阅读说明：本技术 在自适应路由中使用来自相邻节点的负载信息的算法 (Algorithm for using load information from neighboring nodes in adaptive routing ) 是由 D·罗威斯 E·L·弗洛伊斯 J·G·柯普尼克 A·S·科普瑟 R·阿尔弗森 于 2020-03-23 设计创作，主要内容包括：提供了用于在多个交换机之间传递数据的系统和方法,多个交换机具有附接在多个交换机之间的多个链路。在交换机处,从多个相邻交换机接收多个负载信号。多个负载信号中的每一个由指示在多个相邻交换机中的提供负载信号的每个相邻交换机处的负载的一组值构成。该组值内的每个值针对多个链路中与其附接的每个链路提供关于该链路是忙还是安静的指示。基于多个负载信号,选择用于路由接收的包的输出链路并且经由所选的输出链路来路由接收的包。(Systems and methods are provided for communicating data between a plurality of switches having a plurality of links attached between the plurality of switches. At a switch, a plurality of load signals are received from a plurality of neighboring switches. Each of the plurality of load signals is comprised of a set of values indicative of a load at each of the plurality of neighboring switches that provides the load signal. Each value within the set of values provides an indication for each link of the plurality of links to which it is attached as to whether the link is busy or quiet. Based on the plurality of load signals, an output link for routing the received packet is selected and the received packet is routed via the selected output link.)

1. A method of communicating data between a plurality of switches having a plurality of links attached between the plurality of switches, the method comprising:

receiving, at a switch, a plurality of load signals from a plurality of neighboring switches, each of the plurality of load signals comprising a set of values indicative of a load at each of the plurality of neighboring switches providing the load signal, each value within the set of values providing an indication for each link of the plurality of links to which the neighboring switch is attached as to whether the link is busy or quiet; and

based on the plurality of load signals, an output link for routing a received packet is selected and the received packet is routed via the selected output link.

2. The method of claim 1, wherein the switch comprises a source switch.

3. The method of claim 1, wherein the indication of whether the link is busy comprises remote switch busy port masks characterizing ports of the neighboring switch, one or more of the remote switch busy port masks being translated to form entries of a remote switch busy global port table and one or more of the remote switch busy port masks being translated to form entries of a remote switch busy local port table, the remote switch busy port masks being received from the neighboring switch.

4. The method of claim 3, wherein the remote switch busy global port table is indexed according to a destination packet identifier corresponding to a destination packet identifier of a destination packet of the received packet.

5. The method of claim 4, wherein the fabric routing function instance of the switch, after filtering out the candidate ports based on determining candidate ports to filter out are busy from indexed entries in the remote switch busy global port table, determines a recommended port to which the received packet is to be routed according to the routing algorithm executed.

6. The method of claim 3, wherein the remote switch busy local port table is indexed according to a switch identifier corresponding to a switch identifier of a destination switch of the received packet.

7. The method of claim 6, wherein the fabric routing function instance of the switch, after filtering out the candidate ports based on determining candidate ports to filter out are busy based on indexed entries in the remote switch busy local port table, determines a recommended port to which to route the received packet according to the routing algorithm executed.

8. A switch, comprising:

an Application Specific Integrated Circuit (ASIC) to:

receiving a plurality of load signals from a plurality of neighboring switches, each of the plurality of load signals comprising a set of values indicative of a load at each of the plurality of neighboring switches that provides the load signal, each value within the set of values providing an indication of whether the link is busy or quiet for each link of a plurality of links to which the neighboring switch is attached;

selecting an output link for routing the received packet based on the received plurality of load signals; and

the packet is routed via the selected output link.

9. The switch of claim 8, wherein the indication of whether the link is busy comprises remote switch busy port masks characterizing ports of the neighboring switch, one or more of the remote switch busy port masks being translated to form entries of a remote switch busy global port table and one or more of the remote switch busy port masks being translated to form entries of a remote switch busy local port table, the remote switch busy port masks being received from the neighboring switch.

10. The switch of claim 9, wherein the remote switch busy global port table is indexed according to a destination packet identifier corresponding to a destination packet identifier of a destination packet of the received packet.

11. The switch of claim 10, wherein the ASIC includes a fabric routing function instance that determines a recommended port to which the received packet is to be routed according to the executed routing algorithm after filtering out candidate ports to be filtered based on determining the candidate ports to be filtered out from the indexed entries in the remote switch busy global port table.

12. The switch of claim 9, wherein the remote switch busy local port table is indexed according to a switch identifier corresponding to a switch identifier of a destination switch of the received packet.

13. The switch of claim 12, wherein the ASIC includes a fabric routing function instance that determines a recommended port to which the received packet is to be routed according to the executed routing algorithm after filtering out candidate ports to be filtered based on determining the candidate ports to be filtered out from indexed entries in the remote switch busy local port table.

14. A system, comprising:

a plurality of switches interconnected to form a switch fabric;

wherein a first switch of the plurality of switches receives a load signal from a neighboring switch of the plurality of switches, each of the plurality of load signals comprising a set of values indicative of a load at each neighboring switch of the plurality of neighboring switches that provides the load signal, each value within the set of values providing an indication for each link of the plurality of links to which the neighboring switch is attached as to whether the link is busy or quiet; and

wherein, based on the plurality of load signals, the first switch selects an output link for routing a received packet and routes the received packet via the selected output link.

15. The system of claim 14, wherein the indication of whether the link is busy comprises remote switch busy port masks characterizing ports of the neighboring switch, one or more of the remote switch busy port masks being translated to form entries of a remote switch busy global port table and one or more of the remote switch busy port masks being translated to form entries of a remote switch busy local port table, the remote switch busy port masks being received from the neighboring switch.

16. The system of claim 15, wherein the remote switch busy global port table is indexed according to a destination packet identifier corresponding to a destination packet identifier of a destination packet of the received packet.

17. The system of claim 16, wherein the subset of neighboring switches comprises a global minimum path to the destination packet.

18. The system of claim 16, wherein the fabric routing function instance of the first switch, after filtering out the candidate ports based on determining candidate ports to filter out are busy based on indexed entries in the remote switch busy global port table, determines a recommended port to which the received packet is to be routed according to the executed routing algorithm.

19. The system of claim 15, wherein the remote switch busy local port table is indexed according to a switch identifier corresponding to a switch identifier of a destination switch of the received packet.

20. The system of claim 19, wherein the subset of neighboring switches comprises a local non-minimum path and a local minimum path to the destination switch.

21. The system of claim 19, wherein the fabric routing function instance of the first switch, after filtering out the candidate ports based on determining candidate ports to filter out are busy based on indexed entries in the remote switch busy local port table, determines a recommended port to which the received packet is to be routed according to the executed routing algorithm.

Background

As network-enabled devices and applications become more prevalent, various types of traffic and increasing network loads continue to demand higher performance from the underlying network architecture. For example, applications such as High Performance Computing (HPC), streaming media, and internet of things (IOT) may generate different types of traffic that are well characterized. Thus, in addition to conventional network performance metrics such as bandwidth and latency, network architects continue to face challenges such as scalability, versatility, and efficiency.

Drawings

The present disclosure is described in detail in terms of one or more various embodiments with reference to the following figures. The drawings are provided for illustrative purposes only and depict only typical or example embodiments.

Fig. 1 illustrates an example network in which various embodiments may be implemented.

FIG. 2 illustrates an example exchange system that facilitates flow channels in accordance with various embodiments.

Fig. 3A illustrates a crossbar switch implemented within an example crossbar switch, in accordance with various embodiments.

Fig. 3B illustrates an example tile (tile) matrix corresponding to a port of the example edge switching system of fig. 2, in accordance with various embodiments.

Fig. 3C illustrates example tiles comprising the tile matrix of fig. 3B, in accordance with various embodiments.

Fig. 3D illustrates an example age queue implementation in accordance with various embodiments.

Fig. 4A and 4B are block diagrams of example FRF components implemented at each port of the example edge switching system of fig. 2.

Fig. 5 illustrates an example of routing in accordance with various embodiments.

Fig. 6 illustrates an example of local and global load value updates and usage in accordance with various embodiments.

Fig. 7A illustrates an example of routing based on average load.

Fig. 7B illustrates an example of adaptive routing based on neighboring switch load, in accordance with various embodiments.

Fig. 8 illustrates an example computing component capable of executing instructions for load-based adaptive routing, according to one embodiment.

FIG. 9 is an example computing component in which various embodiments described herein may be implemented.

The drawings are not intended to be exhaustive or to limit the disclosure to the precise forms disclosed.

Detailed Description

The present disclosure describes systems and methods that can accommodate billions of sub-computing, for example, performing data intensive tasks such as simulation, data analysis, artificial intelligence workloads, etc., at billions of sub-speeds. In particular, an HPC network or interconnect fabric is provided that may be Ethernet compatible, capable of connecting to third party data storage, and may be constructed using switch components having very high bandwidth (e.g., approximately 12.8Tb/s/dir switches each having, for example, 64 ports of 200Gbps that support creation of large networks at very low diameters (e.g., only three network hops)). Furthermore, low latency can be achieved through the use of novel congestion control mechanisms, adaptive routing, and traffic classification that allows flexibility in bandwidth shaping, priority, and routing policies.

When passing information between switches, an immediate decision must be made on the next hop on the path. The decision is made using information derived from the local state and information communicated from the neighboring switches. Using information from many different sources allows more accurate/efficient decisions to be made. The information includes information from neighbors.

Previous systems carried information about the average load between switches. That is, more detailed information from the relevant or neighboring switch is more helpful. Thus, in some embodiments, a set of values may include information indicating the status of the output ports of the relevant/neighboring switches. By passing this set of values, better routing decisions can be made. In one example, a flag is transmitted back from the neighboring switch, where the flag has one bit for each output port. For example, for a switch with 64 outputs, a 64-bit flag will be transmitted. This is much more accurate than simply passing the global average of adjacent ports.

Fig. 1 shows an example network 100 that includes a plurality of switches, which may also be referred to as a "switch fabric. As illustrated in fig. 1, network 100 may include switches 102, 104, 106, 108, and 110. Each switch may have a unique address or ID within the switch fabric 100. Various types of devices and networks may be coupled to the switch fabric. For example, storage array 112 may be coupled to switch fabric 100 via switch 110; infiniband (IB) -based HPC network 114 may be coupled to switch fabric 100 via switch 108; a plurality of end hosts (e.g., host 116) may be coupled to switch fabric 100 via switch 104; and the IP/ethernet network 118 may be coupled to the switch fabric 100 via the switch 102. For example, a switch, such as switch 102, may receive an 802.3 frame (including an encapsulated IP payload) via an ethernet device, such as a Network Interface Card (NIC), switch, router, or gateway. IPv4 or IPv6 packets, frames, etc. formatted specifically for network 100 may also be received and transmitted through switch fabric 100 to another switch (e.g., switch 110). Thus, the network 100 is capable of handling multiple types of traffic simultaneously. In general, a switch may have edge ports and fabric ports. The edge port may be coupled to a device external to the structure. The fabric port may be coupled to another switch within the fabric via a fabric link.

In general, traffic may be injected into the switch fabric 100 via an ingress port of an edge switch and exit the switch fabric 100 via an egress port of another (or the same) edge switch. The ingress edge switch may divide the injected packets into flows, which may be identified by flow IDs. The concept of a flow is not limited to a particular protocol or layer (e.g., layer 2 or layer 3 in the Open Systems Interface (OSI) reference model). For example, a flow may map to traffic having a particular source ethernet address, traffic between a source IP address and a destination IP address, traffic corresponding to a TCP or UDP port/IP 5-tuple (source and destination IP addresses, source and destination TCP or UDP port numbers, and IP protocol numbers), or traffic generated by a process or thread running on an end host. In other words, a flow may be configured to map to data between any physical or logical entities. The configuration of the mapping may be done remotely or locally on the ingress edge switch.

After receiving the injected packet, the ingress edge switch may assign a flow ID for the flow. The flow ID may be included in a special header that the ingress edge switch may use to encapsulate the injected packet. In addition, the ingress edge switch may also examine the original header field of the injected packet to determine the appropriate egress edge switch address and include that address as the destination address in the encapsulation header. It should be noted that the flow ID may be a link-specific locally valid value, and that this value may be unique only to a specific input port on the switch. When the packet is forwarded to the next hop switch, the packet enters another link and the flow ID may be updated accordingly. Since packets of a flow go through multiple links and switches, the flow ID corresponding to the flow may form a unique chain. That is, at each switch, the packet's flow ID may be updated to the flow ID used by the outgoing link before the packet leaves the switch. This upstream-to-downstream one-to-one mapping between flow IDs may begin at the ingress edge switch and end at the egress edge switch. Because the flow ID only needs to be unique within the incoming link, the switch can accommodate a large number of flows. For example, if the stream ID is 11 bits long, one input port can support a maximum of 2048 streams. Further, the matching pattern (one or more header fields of the packet) for mapping to the stream may include more bits. For example, a 32-bit long matching pattern (which may include multiple fields in the packet header) may map up to 2^32 different header field patterns. If the fabric has N ingress edge ports, a total of N x 2^32 identifiable flows can be supported.

The switch may assign a separate dedicated input queue for each flow. This configuration allows the switch to monitor and manage the degree of congestion of individual flows and prevent head-of-line congestion that can occur when a shared buffer is used for multiple flows. When a packet is transmitted to a destination egress switch, the egress switch may generate an Acknowledgement (ACK) and send the acknowledgement back to the ingress edge switch along the same data path in the upstream direction. Since the ACK packet travels the same data path, the switches on that path can obtain status information associated with the transmission of the corresponding flow by monitoring the amount of outstanding, unacknowledged data. This state information can then be used to perform flow-specific traffic management to ensure overall network health and fair handling of flows. As explained in more detail below, this flow-by-flow queuing in combination with flow-specific delivery acknowledgements can enable the switch fabric to implement efficient, fast and accurate congestion control. Further, the switch fabric can deliver traffic with significantly improved network utilization without congestion.

The flows may be established and released dynamically or "on the fly" based on demand. In particular, a flow may be established by an ingress edge switch (e.g., establishing a mapping of flow IDs to packet headers) when a data packet arrives at the switch and the packet has not previously been assigned a flow ID. As the packet traverses the network, a flow ID may be assigned along each switch through which the packet passes, and a chain of flow IDs may be established from ingress to egress. Subsequent packets belonging to the same stream may use the same stream ID along the data path. When a packet is transmitted to a destination egress switch and a switch on the data path receives an ACK packet, each switch may update its state information regarding the amount of outstanding, unacknowledged data for the flow. For this flow, when the switch's input queue is empty and there is no more unacknowledged data, the switch may release the flow ID (i.e., release the flow channel) and reuse the flow ID for other flows. Such a data-driven dynamic flow setup and teardown mechanism may eliminate the need for centralized flow management and allow the network to respond quickly to traffic pattern changes.

It should be noted that the network architecture described herein is different from Software Defined Networks (SDNs) that typically use the OpenFlow protocol. In SDN, switches are configured by a central network controller and packets are forwarded based on one or more fields in a layer 2 (data link layer, such as ethernet), layer 3 (network layer, such as IP) or layer 4 (transport layer, such as TCP or UDP) header. In SDN, such header field lookups are performed at each switch in the network and there is no flow ID based fast forwarding as done in the networks described herein. Furthermore, since OpenFlow header field lookups are done using Ternary Content Addressable Memory (TCAM), the cost of such lookups can be high. Moreover, since the header field mapping configuration is done by a central controller, the setup and teardown of each mapping is relatively slow, potentially requiring a large amount of control traffic. As a result, SDN networks may be slow to respond to various network conditions (e.g., congestion). In contrast, in the networks described herein, flows may be dynamically established and torn down based on traffic demand; and packets may be forwarded with a fixed length flow ID. In other words, flow channels may be data driven and managed (i.e., set up, monitored, and torn down) in a distributed manner without intervention by a central controller. Furthermore, flow ID based forwarding can reduce the amount of TCAM space used and thus can accommodate more flows.

Referring to the example shown in FIG. 1, assume that storage array 112 is to send data to host 116 using TCP/IP. During operation, storage array 112 may send a first packet having the IP address of host 116 as the destination address and a predetermined TCP port specified in the TCP header. When the packet arrives at switch 110, a packet processor at the input port of switch 110 may identify the TCP/IP 5 tuple of the packet. The packet processor of switch 110 may also determine that the 5-tuple is not currently mapped to any flow ID and may assign a new flow ID to the 5-tuple. Further, switch 110 may determine the egress switch of the packet, i.e., switch 104, based on the destination (i.e., host 116's) IP address (assuming switch 110 knows that host 116 is coupled to switch 104). Switch 110 may then encapsulate the received packet with a fabric header indicating the newly assigned flow ID and the fabric address of switch 104. Switch 110 may then schedule the encapsulated packet to be forwarded to switch 104 based on a fabric forwarding table, which may be computed by all switches in fabric 100 using a routing algorithm such as a link state or distance vector.

It should be noted that the above operation can be performed substantially at line speed, with little buffering and delay when the first packet is received. Subsequent packets from the same flow may be processed by switch 110 more quickly after the first packet is processed and scheduled for transmission because the same flow ID is used. In addition, the design of the flow channels may be such that the distribution, matching and de-distribution of the flow channels may have substantially the same cost. For example, a lookup match-based conditional assignment of a flow lane and a separate, independent deallocation of another flow lane may be performed simultaneously in nearly every clock cycle. This means that the generation and control of the flow paths adds little additional overhead to the conventional forwarding of packets. On the other hand, congestion control mechanisms may improve the performance of some applications by more than three orders of magnitude.

At each switch on the data path (including switches 110, 106, and 104), a dedicated input buffer may be provided for the flow and the amount of transmitted but unacknowledged data may be tracked. When the first packet arrives at switch 104, switch 104 may determine that the destination fabric address in the fabric header of the packet matches its own address. In response, switch 104 may decapsulate the packet according to the fabric header and forward the decapsulated packet to host 116. In addition, switch 104 may generate an ACK packet and send the ACK packet back to switch 110. When the ACK packet traverses the same data path, switches 106 and 110 may each update their own state information for the flow's unacknowledged data.

Typically, congestion within the network will cause network buffers to be filled. When the network buffer is full, ideally, traffic attempting to pass through the buffer should be slowed or stopped. Otherwise, the buffer may overflow and the packet may be discarded. In conventional networks, congestion control is typically done end-to-end at the edge. The core of the network is considered to act merely as a "bulk water pipe," the primary purpose of which is to forward traffic. Such network designs are typically slow to respond to congestion because congestion information is typically not sent to the edge devices quickly, and the resulting actions taken by the edge devices are not always effective at removing congestion. This slow response in turn limits the utilization of the network, since network operators typically need to limit the total amount of traffic injected into the network in order to keep the network clear. Furthermore, end-to-end congestion control is typically only effective if the network is not congested. Once the network is heavily congested, end-to-end congestion control will not work because the congestion notification message itself may be congested (unless the congestion control message is sent using a separate control plane network than the data plane network).

In contrast, flow paths may prevent such congestion from growing within the switch fabric. When a flow is experiencing some degree of congestion, the flow path mechanism may recognize this and, in response, may slow down or stop new packets of the same flow entering the fabric. Further, these new packets may be buffered in a stream path queue on the edge port and only allowed to enter the fabric when a packet of the same stream leaves the fabric at the destination edge port. This process may limit the total buffering requirements of the stream within the fabric to an amount that does not cause the fabric buffer to become too full.

Through the stream path, the switch is able to obtain fairly accurate status information about the amount of data outstanding en route within the fabric. This status information may be aggregated for all flows on the ingress edge port. This means that the total amount of data injected through the ingress edge port can be known. Thus, the stream lane mechanism may place a limit on the amount of data in the structure. When all edge ports apply this limiting action, the amount of packet data in the entire fabric can be well controlled, which in turn can prevent the entire fabric from saturating. The flow path may also slow down the progress of a single congested flow within the fabric without slowing down other flows. This feature can keep packets away from congestion hot spots while preventing buffers from becoming full and ensuring that free buffer space is provided for irrelevant traffic.

Fig. 2 illustrates an example switch 202 (which may be an embodiment of any one or more of switches 102, 104, 106, 108, and 110) that may be used to create a switch fabric (e.g., switch fabric 100 of fig. 1). In this example, the switch 202 may include a plurality of communication ports, such as port 220. Each port may include a transmitter and a receiver. The switch 202 may also include a processor 204, a storage device 206, and stream path switching logic 208. The stream lane switch module 208 may be coupled to all communication ports and may further include a crossbar switch 210, an EFCT logic block 212, an IFCT logic block 214, and an OFCT logic block 216.

Crossbar switch 210 comprises a crossbar switch that may be configured to forward data packets and control packets (e.g., ACK packets) between communication ports. EFCT logic 212 may process packets received from the edge link and map the received packets to corresponding flows based on one or more header fields in the packets. Additionally, the EFCT logic 212 may assemble FGFC ethernet frames, which may be transmitted to the end host to control the amount of data injected by the respective process or thread. The IFCT logic block 214 may include IFCTs and perform various flow control methods in response to control packets, such as endpoint congestion notification ACKs and structural link credit based flow control ACKs. The OFCT logic 216 may include a memory unit that stores OFCTs and communicates with the IFCT logic of another switch to update the flow ID of the packet as it is forwarded to the next hop switch.

In one embodiment, switch 202 is an Application Specific Integrated Circuit (ASIC) that can provide 64 network ports that can operate at 100Gbps or 200Gbps, thereby achieving an aggregate throughput of 12.8 Tbps. Each network edge port may be capable of supporting IEEE 802.3 ethernet and optimized IP based protocols as well as Portal (an enhanced frame format that supports higher rate mini-messages). Ethernet frames may be bridged based on their L2 address or routed based on their L3(IPv4/IPv6) address. The optimized IP frame may have only an L3(IPv4/IPv6) header and be routed. The dedicated NIC supports a fabric format that may be used for portal enhanced frame formats and may be mapped directly to network 100, such as a fabric format that provides certain control and status fields to support a multi-chip architecture when switch/switch chips (e.g., switches 102, 104, 106, 108, and 110) are connected and communicate with each other. As described above, flow-channel based congestion control mechanisms may be used by such switches, and may also enable high transmission rates of smaller packets (e.g., more than 12 hundred million packets per second per port) to accommodate the needs of HPC applications.

The switch 202 may provide system-wide quality of service (QoS) classification, and the ability to control how network bandwidth is allocated to different classes of traffic and different classes of applications, where a single privileged application may access more than one class of traffic. In the presence of network bandwidth contention, the arbiter will select a packet to forward based on the packet's traffic classification and the credits available for that classification. The network 100 may support a minimum bandwidth and a maximum bandwidth for each traffic class. If a class does not use its minimum bandwidth, other classes may use unused bandwidth, but no class may result in bandwidth that exceeds its maximum allocated bandwidth. The ability to manage bandwidth provides an opportunity to dedicate network resources and CPU and memory bandwidth to a particular application.

In addition to supporting QoS classification, the switch 202 implements flow channel-based congestion control and may reduce the number of network hops in a network having, for example, a dragonfly topology from five network hops to three. The design of the switch 202, described in more detail below, may reduce network cost and power consumption, and may further facilitate the use of innovative adaptive routing algorithms that improve application performance. The fabric created by multiple switches (e.g., multiple switches 202) may also be used to build a fat-tree network, for example, when building a storage subsystem for integration with third party networks and software. Still further, the use of the switch 202 enables fine-grained adaptive routing while maintaining in-order packet delivery. In some embodiments, the switch 202 may be configured to send the header of the packet from the input port to the output port before the complete data payload arrives, allowing the output port load metric to reflect future load, thereby improving adaptive routing decisions made by the switch 202.

Crossbar switch 210 may include individual, distributed crossbars that route data/data elements between input ports and output ports. In some embodiments, as illustrated in fig. 3A, there are five distributed crossbars between input port 220b and output port 220c, including request crossbar 210a, grant crossbar 210b, credit crossbar 210c, Ack crossbar 210d, and data crossbar 210 e.

Request crossbar 210a is used to send requests from the input to the target output age queue. The grant crossbar 210b is used to return grants to the input that satisfy the request. In particular, grant crossbar 210b returns a pointer indicating the location of the packet within the input buffer. It should be noted that when there is room in the output for the corresponding packet, a grant will be returned. Authorization crossbar 201b may also optionally return credits for the requested space in the output. It should be noted that grants are returned when there is a landing site for a packet at the output (e.g., output port 220c), so packets cannot be blocked (although these packets may be subject to transient contention for resources).

It should be appreciated that, according to various embodiments, a credit protocol may be used to ensure that there is landing space for requests at the output. Accordingly, credit crossbar 210c may be used to return credits for the requested space in the output.

Data crossbar 210d is used to move authorized packets from the input buffer to the target output buffer. Ack crossbar 210e is used to propagate Ack packets from output port 220c to input port 220 b. The Ack is manipulated according to the state maintained in the output stream channel table.

It should be understood that data crossbar 210d moves multi-clock packets with both headers and data, while the other four crossbars (request crossbar 210a, grant crossbar 210b, credit crossbar 210c, and Ack crossbar 210e) move only single-clock packet headers. All five crossbars use the same architecture, with the row and column buses within an 8x 4 matrix of 32 dual-port tiles (tiles) (as described below).

Referring back to fig. 2, the switch 202 may have multiple transmit/receive ports, e.g., port 220. The plurality of ports may be configured as a tile matrix. Fig. 3B illustrates an example of such a tile matrix 300. In one embodiment, tile matrix 300 includes 32 tiles, each tile including two ports for implementing crossbar switching between the ports, and for providing the following: a serializer/deserializer (SERDES) interface between the core of the switch 202 and external high-speed serial signals to drive signals out of the switch 202; a Media Access Control (MAC) sublayer interface to a Physical Coding Sublayer (PCS); PCS interface between SERDES and Ethernet MAC function; a Link Level Retry (LLR) function that operates on a packet-by-packet basis and uses an ordered set to convey an initialization sequence, an Ack, and a Nack; and an ingress transform block for converting between different frame structure formats. Each tile contains a crossbar switch, such as crossbar switch 210 for each of the crossbars (210 a-201 e). As will be discussed in more detail below, routing in the switch fabric may be controlled by a Fabric Routing Function (FRF) implemented in the switch 202, where separate instances of FRF components (fig. 4A, 4B) may be implemented within the input logic of each port of the switch 202. As described above, each port includes an instance of an FRF component, but for ease of reference/clarity of illustration, only two example FRF instances (400a, 400b) are labeled.

Each crossbar switch 210 has sixteen inputs 220b (one for each port in its row) and eight outputs 220c (one for each port in its column). The row bus can be driven from each source in a row to all eight crossbars (one-to-all) in that row. Arbitration may be performed at a crossbar from sixteen row buses for the row to eight column buses in a given column. Buffering may be provided at each 16x 8 crossbar of each row bus to absorb packets during contention for the column bus. In some embodiments, non-jumbo packets are kept out of the row bus unless there is room in the input buffer of the target crossbar for the entire packet. Due to area limitations, even if there is not enough space (the input buffer of the crossbar is sized to receive only non-jumbo packets), jumbo packets are allowed to pass through, with the row bus blocked until the packet wins arbitration and frees up space when it moves onto the column bus.

The column bus is driven from a given crossbar to each destination port (all to all) within a column. Each destination may have another level of arbitration between column buses from four rows. Sixteen row buses drive eight crossbars, each feeding eight column buses, with 4 times acceleration between rows and columns. Each row has the same connections, i.e., a pair of all row bus connections for a single row as shown in the row bus. Each tile will have one (request, grant, credit) or two (data, ack) clock delay per tile depending on the crossbar. This provides a maximum of seven or fourteen clock delays between the leftmost column and the rightmost column. Credit returns routed through credit crossbar 210c may have a clock delay of one per tile and, therefore, may take up to a maximum of seven clocks to complete a transmission.

It should be noted that each column may have the same connections, i.e., all pairs of a single column are bus connected to all columns, and there may be two clock delays per tile, resulting in six clock delays from the top row to the bottom row. It should also be understood that both the row and column buses use the credit-based protocol described above to determine when they are able to transmit. In the case of a row bus, the source port maintains a credit count for the input buffer of the in-row crossbar. For a data crossbar, care needs to be taken to determine when to admit a packet to the row bus. If the grants for the input buffers of a particular crossbar all pass through a single queue, space is needed for the packet at the head of the queue before starting the packet transmission. If grants are distributed among multiple queues, packet transfer will not begin unless there is room in the buffer for the entire maximum size packet in order to prevent small packets from locking out large packets. In this way, once a packet transfer on the row bus begins, it does not stop until the entire packet transfer is completed. Thus, the input buffer of the crossbar is configured to be large enough to handle the maximum packet size and have additional space to cover the worst-case round trip (packet send to credit return). This will not be the case for jumbo bags. To save buffer area, the input buffer of the crossbar is only deep enough to handle a non-jumbo sized MTU (1500 bytes), where jumbo packets are allowed to block the row bus while waiting to gain access to the target column bus.

For the column bus, each crossbar maintains a credit count for the input buffer at each destination port in the column. Unlike the row bus, there is no requirement that a credit be available for the largest size packet before the packet begins to be transmitted on the column bus. When credit becomes available, the words of the packet will be moved. Thus, the input buffer at the destination of each column bus need only be large enough to cover the worst-case round-trip (packet-to-credit).

Fig. 3C illustrates in more detail an example embodiment of two ports (e.g., ports 0 and 1) handled by tile 1, and a crossbar 220a in which the crossbar per tile includes a set of row buses and column lanes. In this way, each port has its own row bus across which to communicate, and each tile has the aforementioned 16x 8 crossbar (for making corner turns) and a set of eight column lanes (up to eight ports contained in that column can be provided). In other words, each crossbar switch 210 has sixteen row bus input buffers and eight possible destinations. For example, for data traveling from, for example, input port 17 to output port 52, the data is routed from input port 17 along the row bus, through the local crossbar that performs 16-to-8 arbitration, and then up through the column lanes to output port 52. The internal bandwidth is four times the external bandwidth in terms of the aggregate route through all the set of distributed crossbars, resulting in the ability to keep up with the ingress when routing almost any arbitrary traffic permutation through the switch 202.

Fair round-robin arbitration may be used among sixteen sources per destination. For data crossbar 210d, once the source wins arbitration, the source retains control of the destination column bus until the entire packet has been sent. Each output is granted a limited number of packet payloads and so it is expected that contention for a given column bus when larger packets are involved should be quite limited. Thus, even though there may be large differences in packet size between requesters, polling arbitration is expected to be sufficient.

The portion of the switch 202 associated with the egress function typically operates on frames within the switch fabric format and has a fabric header (e.g., even for frames arriving on ethernet ports alongside within a single switch 202).

The age queue output control is responsible for accepting requests from all input ports (e.g., input port 220b) via request crossbar 210a, buffering the requests, arbitrating between these requests by traffic classification using a traffic shaper, and passing the requests to OFCTs 216 to be granted via grant crossbar 210 b. Age queue buffering is managed to allow each input sufficient space to flow while also allowing inputs with multiple flows for a given output to take up more space. In particular, age queue space is managed by output control. Age queue/output control may also be responsible for managing access to the link, using either credit-based flow control for the input buffer of the connection or pause-based flow control for the unstructured link. When a packet is released by the age queue, it is committed to be placed on the link. In addition, the age queue has a path that allows packets (such as maintenance or reduction packets) that originate on a given port (e.g., one of input ports 220b) to be arbitrated for resources on the given port.

Requests enter the output control block from each row of the matrix 30 via the column bus. Each column bus feeds a separate FIFO (e.g., a first-in-first-out shift register or buffer), where space in the FIFO is managed by credits. The FIFO size (24 depth) may be set to cover the round trip plus additional space to allow requests to be moved out of the crossbars 210a to 210e and prevent head of line blocking. The requested valid Error Correction Code (ECC) may be checked before writing to the FIFO. If the destination field of the ECC check has a Multiple Bit Error (MBE) or a Single Bit Error (SBE) therein (i.e., the request has been routed to the wrong port), the request is considered an invalid request and is discarded, with the error flagged.

A Least Recently Used (LRU) arbitration may be performed between the column bus FIFOs to select which FIFO to forward to age queue management. When a request is removed from each FIFO, credits will be returned to the corresponding crossbar. The row to which the incoming column bus corresponds may depend on the location of the tile in the matrix and on which half of the tile the block is located.

The Output Buffers (OBUF) make requests to the output control block to send the reduction and maintenance packets on the link. These requests may be given the highest priority. A FIFO with 8 positions may be used to buffer these reduce/maintain packet requests as they wait for resources. The reduction package does not need to use a flow path and the maintenance package can create the flow using looping back so that there is no need to check flow path availability or flow through OFCT to create authorization. The reduction and maintenance packets also do not need to use any space in the output buffer, so that no examination space is needed. Instead, a check of the link partner input buffer may be performed. If allowed, the Shaping Queue (SQ) or Virtual Channel (VC) may be granted, preventing any grant from the age queue path during that period.

The size of the next request to be processed from the output buffer is checked against max _ frame _ size. If the request exceeds this setting, the request is not processed and an error flag is set. This will cause the output buffer request path to be blocked until the warm reset position is performed. The error flag will remain set until the reset is complete. The condition can also be relieved by increasing the max _ frame _ size setting to a value above the size of the stuck output buffer request. The size used in the comparison may be the size indicated in the output buffer request (which may include a 4 byte Frame Checksum (FCS) used on the line).

Each input may be assigned the same fixed allocation of age queue space. This age queue space is large enough to reserve a location for each SQ/VC and there is enough extra space to cover request/credit round trips. Managing the space it is given on its SQ/VC depends on the input. The allocation (fixed _ al/oc) is programmable via a Control and Status Register (CSR) in each input queue (INQ) and may range, for example, from 64 to 96 locations. The remaining age queue space (8K-64 fixed _ al/oc) may be a shared space available for all inputs. The shared space may be managed by the output, if there is room in the shared space, it will move incoming requests from the static space to the shared space as they arrive, but limited by each input. When a request is moved to a shared space, credits are returned (e.g., immediately) via credit crossbar 210c and the request is marked as being in the shared space in the age queue.

When a request is authorized, the shared space is credited if the request is marked as using the shared space. If the request is not marked as using shared space, the request is deemed to have used static space and a credit is returned to the input with the authorization.

Credits may not be sent every clock cycle due to conflicts in credit crossbar 210 c. Thus, the FIFO provides a buffer for these transient interruptions. Space in this FIFO is needed before a request is fetched from the request crossbar. A FIFO depth of 32 positions may be used to limit the chance that it will be backed up into the request crossbar 210 a. The shared space may have a limit on how much space any input (from input port 220b) may occupy. These limits may be set as a percentage of the available space. For example, if the limit is set to 50%, if an input port is active, the input port may access 50% of the buffer space; if there are two active input ports, each input port gets 37.5% ((space _ used _ by _ l + space _ left.5)/2 ═ 50% + 50%. 5)/2); in the case of three active input ports, each input port gets 29.2% ((space _ used _ by _2+ space _ left.5)/3 ═ 75% + 25%. 5)/3), and so on. In addition, the total space used by the active input ports may be limited to a given total (50%, 75%, 87.5%). Thus, the space allocated to each input port 220b may vary dynamically depending on the number of currently active input ports. Adding an active input port may cause other active input ports to relinquish their space, which is then taken up by new inputs.

Considering that the partitioning is not easily done in hardware, the age queue credit management function described above may be implemented as a look-up table 310 with 64 entries 312. The number of currently active entries in the age queue 320 indexes 315 the lookup table 310. The value 314 in the lookup table 310 reflects the number of shared space locations that any input may occupy and the limit of the total space they may consume in their entirety. Thus, the values 314 in the look-up table 310 are programmed by software according to how much shared space is in common and the percentage of each input port allowed to occupy. As more input ports 220b become active, less and less space is allowed for each input port 220b, and the total space available increases. Incoming requests from input port 220b above this limit or in amounts exceeding the total space limit are not allowed to occupy more shared space. To track the number of active input ports 220b in the age queue, a set of 64 counters 316 (one for each input port) is used. These counters count up when a request is placed in the age queue 320 and count down when it is fetched (i.e., authorized). The count 319 of the number of non-zero counts is used as an index to the lookup table 310. Additionally, to manage the shared space, an additional set of 64 counters 318 may be used to track the current usage of the shared space by each input. There may also be a single counter 334 that may be used to track overall shared space usage. These counters are compared to the current quota to determine whether the request is allowed to use the shared space. The counters 316, 318 may be, for example, 13 bits wide to provide adequate coverage for a maximum value of an object that may be slightly less than 8K.

The age queue 320 may use a single storage RAM 321 having 8K locations therein. These locations may be dynamically assigned to 32 separate queues (one queue per SQ/VC), where each location consists of a linked list of locations within storage RAM 321. This enables each SQ/VC to take up more space as required.

The age queue 320 may be created using a front pointer 322 to the front of the queue and a next pointer 324 to each location of the next entry in the queue. The last position in the queue may be indicated by a backward pointer 326. Entries are fetched from the front of the queue and inserted into the back of the queue. In addition to the data structures described above, each queue has a FIFO 328 of entries at its head. These FIFOs 328 can ensure that the queue can maintain requests at each clock with multi-clock read access time from the request RAM 321. When a new request arrives, if the head FIFO 328328 of the queue is not full, the request may write directly to the head FIFO 328 bypassing the request RAM 321. Once a request for a given age queue is written to request RAM 321, subsequent requests are also written to request RAM 321 to maintain order. The bypass path may be used again once there are no more requests for the age queue in the request RAM 321 and there is room in the head FIFO 328. Dequeuing is initiated when a request is read from head FIFO 328 and there is a corresponding request queued in request RAM 321. One head FIFO 328 may be read at a time so that a single dequeue operation may be initiated every clock cycle. Logic may be included to handle various race conditions between an ongoing or imminent enqueue operation and the head FIFO 328 being read.

The ECC protection described above for age queue RAM 321 may be extended to FIFO 328 to protect the datapath flip-flops. The resulting structure may include 8K flip-flops (32 queues x deep 5 x SQ bits wide). The age queue number may be included in the calculation (but not stored) as an additional check of free list management when generating the ECC. When checking the ECC, a request may be considered erroneous if there is an MBE or SBE in the queue number bits.

The free list RAM can be a simple FIFO that is initialized with pointers to all 8K entries whenever a reset is performed. A count may be maintained to track how many entries are valid within the free list. When an entry is fetched, the entry is popped from the front of the FIFO and used. When an entry is returned, the entry is pushed to the back of the FIFO. A certain number of entries (e.g., three entries) at the head of the free list may be kept in the trigger so that these entries can be accessed quickly. As with the head FIFO for the age queue, the ECC is carried over the flip-flops to provide protection. The resulting structure may have the fewest flip-flops (57 ═ depth 3x19 bits wide).

To support the full performance of a packet, the age queue supports enqueuing and dequeuing every clock cycle. Operations on the data structure for enqueue operations are discussed below and may differ depending on whether the queue being written is empty.

In some cases, simultaneous enqueuing and dequeuing for a particular queue may be handled easily due to the use and updating of different fields. Some special scenarios may occur, for example, when a dequeue operation empties the age queue. To handle such scenarios, logically dequeue first occurs, followed by an enqueue operation. Thus, the empty flag is considered to be set when the queue is emptied by the dequeue operation, and then cleared due to the enqueue operation.

The arbitration may be performed between requests that are granted subject to input buffer management, output buffer management, and stream channel quotas. Arbitration may also be suspended if the OFCT input FIFO has no credits. In some embodiments, arbitration may be performed in two stages. First, traffic shaping arbitration may be used to arbitrate between SQs. Deficit polling arbitration may be used to arbitrate between VCs within a given SQ. Traffic shaping arbitration may use a series of token buckets to control the bandwidth of each SQ as follows: eight leaf buckets, one leaf bucket for each SQ; four branch barrels; and a single head barrel.

Arbitration may be divided into three packets, the first having the highest priority, followed by a second packet, followed by a third packet. Arbitration may be handled in the same manner between eligible SQs for the first packet and the second packet. For each of the eight priorities, x8 round-robin arbitration (eight parallel round-robin arbitration) may be performed between SQs. A fixed arbitration between the priorities may be performed. For example, packet 3 arbitration has no priority, and therefore is only a single x8 round robin arbitration.

For the arbitration in the first packet, the priority of each arbitration comes from the setting in the leaf bucket. For arbitration in the second packet, the priority comes from the setting in the leaf bucket branch. In all cases, if the request wins arbitration, the bucket that is checked as eligible for the packet is also the bucket from which the packet size token is obtained.

With regard to age queue 320 selection, packets may be classified to select the SQ to which their request is to be forwarded. This allows traffic associated with an application to be shaped differently than traffic originating from a different application or a different traffic classification. This may be useful on the edge port connected to the NIC, as the application will be configured to use resource sharing on the node, and similarly will be granted a certain proportion of network bandwidth. According to one embodiment, the classification is performed by classifying the packet as a traffic classification identifier (FTAG) (e.g., a 4-bit code that is part of the fabric frame header) and a VLAN ID (VNI) as the packet enters the fabric. The FTAG and VNI may then be used to select a shaping queue when a packet leaves the fabric.

Configuration registers may be used to map the FTAG to the SQ. This configuration matches the corresponding configuration in the in-queue (in queue). When the output buffer requests or returns link partner credits, it will convert the given FTAG to SQ. For packet injection, FTAG is located in R _ TF _ OBUF _ CFG _ PFG _ TX _ CTRL. For test generation, the FTAG is located in a test control register. When the Reduce Engine (RED) requests credit return, the FTAG is located in ret _ cdtJtag. The FTAG is located in the frame header when the reduction frame is removed from the outgoing stream and return link partner credits are required.

With respect to SQs discussed herein, each age queue 320 may have 32 SQs addressed by { SQ, VC }. The 3-bit SQ 330 may be considered a shaping function, and the VC selects one of four queues within the shaping function. For ethernet egress (edge) ports, no VC is needed to avoid deadlock. Thus, all 32 SQ 330 are available. In this scenario, SQ 330 may be selected by adding the SQ base from R _ TF _ OBUF _ CFG _ FTAG _ SQ _ MAP to the lower bits of the VNI. The 5 bit sum defines { SQ, VC } to be sent to the age queue. It should be noted that when injecting frames at the egress port, the VNI is not available and therefore the SQ base can be used directly. For fabric links, SQ 330 is taken from the upper three bits of the SQ base. The VC may be obtained from the frame header when credit to the reduction frame is returned, or from the appropriate control CSR (R _ TF _ OBUF _ CFG _ TEST _ CTRL or R _ TF _ OBUF _ CFG _ PFG _ TX _ CTRL) when the frame is injected.

Link partner input buffer management may depend on the type of device to which the link is attached. Devices such as switch 202 may use credit-based flow control, where each credit represents a memory location in an input buffer. Other devices may use standard-based ethernet suspension or priority-based suspension flow control. Requests marked to terminate locally (lac term group) do not need to take into account link partner input buffer flow control and do not need to update any associated counters. When the link is in a drained state, there is no need to consider the link partner space.

For credit-based flow control, the link partner input buffer may be classified into eight buffers. Each SQ 330 may be assigned to one of the 8 buffer classifications. Credits are maintained for each buffer classification, where each credit represents 32 bytes of storage in a link partner input buffer. To allow credit-based flow control to work with various devices (switches, enhanced NICs), each device may have a different cell size, which is a programmable value in units of 32 bytes.

There may be two sets of VCs, with each SQ 330 assigned to one of the sets. A useful space of maximum frame size may be reserved for each VC, and each set of VCs may have a different maximum frame size. The remainder of the link partner input buffer is a shared dynamic space that any SQ/VC can use, but is limited by each VC and buffer classification.

The size of the accompanying request indicates the size of the packet on the line, including 4 bytes of FCS. This is converted to an internal 2 byte FCS at the link partner before writing the packet into the link partner input buffer, so the credit needs to account for this difference, which may be a factor involved at the cell size boundary. For example, for a 96 byte unit, a size of 97 or 98 would occupy a single unit. To know when this occurs, the request includes a correction term that is calculated as follows: len _ correct ═ (byte _ len% 16) ═ 1 or 2.

This entry needs to be further validated to convert it to any possible cell size boundary. This term is valid when the length is just beyond the cell size. In this case, the verified fen _ correct entry may be determined by: len _ correct ═ (((16 byte size)% (2 × 32 byte unit size)) ═ 1) & req

The following table illustrates examples of how these values apply to a few units and packet sizes:

length correction calculation

The size of the accompanying request uses 8 bytes as a unit, and the link partner input buffer unit size is a multiple of 32 bytes (32 x y, where y is the unit size from the CSR). First, the 8-byte size is converted into a 16-byte size (ROUNDUP ((8-byte size)/2)). Also, the cell size is converted to a unit of 16 bytes (2 × y). Mathematically, the number of units requested to be used can be calculated by: round ddn (((16 bytes size) +2 × y-1-len _ correct)/(2 × y)) # units

Although the partitioning operation can be performed in hardware, it cannot be done in the critical path of arbitration for timing reasons. Instead, alternative credit management is used. That is, the credit is maintained in units of 32 bytes. When the request wins arbitration, the number of credits occupied is adjusted according to the maximum error term (2 x y-1) using the following calculation: round ddn (((16 bytes size) +2 x y-1)/2) ═ credit requiring a maximum of 32 bytes. Since this calculation overestimates the credit required for the packet, on the next clock, a modulo operation (X (16 byte size) MOD 2 y, y 32 byte unit size from CSR) can be performed to determine the actual remainder. This value is used with the len _ correct entry to adjust the credit counter. The formula for creating the adjustment value (adf _ val) for X is: if (X ═ 0), adj _ val ═ y-1, otherwise if (X ═ 1 and fen _ correct), adj _ val ═ y, otherwise, adj _ val ═ ROUNDDN ((X-1)/2).

The following table illustrates a request credit example for a 96 byte unit, which shows the values used by the 96 byte unit (y 3) of the switch input buffer over several packet lengths.

Request credit example for 96 byte unit

If the request is filtered before forwarding to the link partner input buffer, the output buffer logic returns the SQ and VC so that the SQ and VC can be used to return credits to the appropriate credit counter. No size is needed because the packet size is always the same, i.e. the length of the frame is reduced (69 bytes or 16 bytes size 5).

The local (primary) side of the link maintains a count of the number of packets (8 total) sent from each VC on both sets, a count of the number of packets (4) sent to each VC (in 32 bytes), and a count of the number of packets (8) sent for each buffer class (in 32 bytes). The link partner (slave) side of the link maintains the same set of counts and sends these counts on the link periodically. The difference between the master-slave counts is a count of the number of packets in the link partner input buffer from each VC on both sets and a count of the amount of space currently occupied (in units of 32 bytes) by each VC and each buffer class. A count of the total amount of space used by all packets is also maintained. The counters are summarized as follows: master _ vcx _ cnt [4]/slave _ vcx _ cnt [4] -a master-slave count of the number of packets sent to each VC in set X; master _ vcy _ cnt [4]/slave _ vcy _ cnt [4] -a master-slave count of the number of packets sent to each VC in set Y; master _ bc _ cnt [8]/slave _ bc _ cnt [8] -master-slave count of the amount of space occupied by each buffer classification, in units of 32 bytes; master _ VC _ cnt [4]/slave _ VC _ cnt [4] -a master-slave count of the amount of space occupied by each VC, in units of 32 bytes; master-tot-cnt/slave-tot-cnt-master-slave count of the total amount of space occupied, in units of 32 bytes.

All counters are set to zero at the warm reset. The counter is also forced to zero when the link is in a drained state or a state where the DBG _ RESET CSR bit is set to clear the counter. The output buffer filter directs the reduction packets to a location outside of the path to the link partner input buffer. In this case, the signal may be returned along with the SQ and VC of the packet. Also, since the size of these packets is fixed, no length is required. This information is used to adjust the appropriate master credit count.

A request is allowed to participate in arbitration if its VC count is 0 (indicating that one of its statically assigned slots is available) or there is space in the dynamic space for the largest sized frame (subject to the target buffer classification and VC limit). The maximum frame size may have a single programmable value for all VCs and SQ. The request checking of the input buffer space can be solved using credit-based flow control.

Credit-based flow control can be used to partition the dynamic space in two ways, each independent of the other: first, based on the limit on how much dynamic space each of the four VCs can occupy; and secondly, a limit based on how much dynamic space each of the eight buffer classifications can occupy. In both cases, the limit is set as a percentage of the available space. For a given packet, available space should be provided in both its target VC and buffer classification. For example, if the limit for each space is set to 50%, if one space is active, it may access 50% of the buffer space, if there are two activities, each space gets 37.5% ((50+50 x.5)/2), if there are three activities, each space gets 29.2% ((75+25 x.5)/3), and so on. Moreover, the total space usage of these activities may be limited to a given total (50%, 75%, 87.5%). Thus, the space allocated to each dynamically changes according to the number of current activities. When another space becomes active, it causes the other active spaces to give up some of their spaces and then occupy them by the new one.

As with the partitioning function discussed above, this function is implemented as a lookup table. For VC space in this example, there are 16 entries, where each entry specifies the space available for each VC and the total space available for all VCs. For buffer classification, there may be 256 entries, where each entry specifies the space available for each buffer classification and the total space available for all buffer classifications. The space for each is represented in 2048 bytes. Each table is deep enough to cover all combinations of active members (VC or buffer classification), where each combination can set the percentage of these combinations independently. In this case, the values in the table are programmed by the software according to how much dynamic space is in common and the percentage each space is allowed to occupy in all possible combinations. As more members become active, less and less space is allowed for each member, and the total space available increases. Requests for space above this limit or a total amount above the total limit do not allow more dynamic space to be taken up.

A VC or buffer classification is considered active if it has a request in the age queue or if it has outstanding credits for the link partner to enter the buffer space. As an example, consider a table with only 4 SPACEs (16 entries), the percentages set to SPACE0 (50%), SPACE1 (40%), SPACE2 (30%), SPACE3 (10%), and the total dynamic SPACE is 16 KB. This will result in a value in units of 16 bytes as given in the following buffer space example table.

Buffer space example

Index	SPACE3	SPACE2	SPACE1	SPACE0	Total of
						0	N/A	N/A	N/A	N/A	N/A
1	N/A	N/A	N/A	512	512
						2	N/A	N/A	410	N/A	410
3	N/A	N/A	319	398	717
						4	N/A	307	N/A	N/A	307
5	N/A	250	N/A	416	666
						6	N/A	255	339	N/A	594
7	N/A	202	270	337	809
						8	102	N/A	N/A	N/A	102
9	94	N/A	N/A	469	563
						10	94	N/A	377	N/A	471
11	75	N/A	299	374	748
						12	95	284	N/A	N/A	379
13	78	234	N/A	389	701
						14	80	239	319	N/A	638
15	79	236	315	394	1024

As an example, the values in the row of index 7 are calculated as follows: total% (% 0.5+ (1-0.5) × 0.4+ (1-0.5- (1-0.5) × 0.4) × 0.3 ═ 0.79; SPACE0 ═ (0.5/(0.5+0.4+0.3)) × 0.79 ═ 1024 ═ 337; SPACE1 ═ (0.4/(0.5+0.4+0.3)) × 0.79 ═ 1024 ═ 270; SPACE2 ═ (0.3/(0.5+0.4+0.3)) × 0.79 ═ 1024 ═ 202; the total of 337+270+202 809.

As described above, and referring back to fig. 2, a switch, such as switch 202, can be used to create a switch fabric in which switch ports 220 can be configured to operate as edge ports or fabric ports. As also described above, the switch 202 may support various network topologies including, but not limited to, for example, dragonfly and fat-tree topologies. A network may be considered to include one or more slices, each having the same overall topology, although the slices may differ in the manner in which each slice is populated. The nodes are connected to one or more ports on each slice. When a network has multiple slices and a node is connected to more than one slice, it is assumed that the node is connected at the same location in each slice.

Routing in the switch fabric may be controlled by Fabric Routing Functions (FRFs) implemented in the switch 202. An example FRF component 400 is illustrated in fig. 4A and 4B. It should be understood that separate instances of FRF component 400 may be implemented within the input logic of each port of switch 202. The routing decisions made by the FR component 400 can be applied to frames that are not already part of an established flow. It should be noted that FRF component 400 does not necessarily know whether a particular frame is associated with a flow, but rather makes an independent forwarding decision for each frame presented at an input port. FRF component 400 may include filters, tables, circuits, and/or logic (such as selection circuitry/logic) to enable routing of data throughout the switch fabric as described herein. As illustrated, FRF component 400 includes at least: a minimum port selection component 402 (which includes a minimum table component 402A), various port filters (allowed port filter, operational port filter, busy port filter); a preferred port identifying part 402B; pseudo-random down-select component/logic 402C; an exception table 404 (including an exception list table 404A); an operation port component 406 comprising a global fault table 406A; and a routing algorithm table 408. As illustrated in fig. 4B, the FRF part 400 may further include: non-minimum port selection component 410, which includes local non-minimum selection component 410A and global non-minimum selection component 410B; and an output logic component 412 (which is part of the output control block of the switch) that includes an adaptive selection component or logic 412A. FRF component 400 includes other components and is described herein.

In particular, FRF component 400 uses preferred port discriminator 402B to determine a preferred port to forward each frame presented at an input port based on: a destination structure address (DFA) of the received frame; the current routing state of the frame (the position of the frame on its path, and the path(s) taken to reach its current routing state); switch fabric routing algorithms and configurations; and a load metric associated with the output port that uses the busy port filter (the above-mentioned preferred port to which the frame is to be forwarded).

FRF component 400 may include a routing algorithm table 408, which may be embodied as a software configurable table that determines valid selections based on the current routing state of the frame. An efficient choice is a decision such as whether to allow a local minimum, global minimum, local non-minimum, or global non-minimum path to be selected for the next hop of the frame. The routing state includes information such as the VC on which the frame was received and whether the frame is in the source, destination, or intermediate packet. Routing algorithm table 408, along with adaptive selection function or logic 412A (described below), also determines the VCs to be used for the next hop of the frame.

Frame routing using unicast DFA is described as an example. It should be noted, however, that the DFA of the routing request may be in unicast or multicast format. The unicast format may include a 9-bit global ID field (global _ ID), a 5-bit switch ID field (switch _ ID), and a 6-bit endpoint ID field (endpoint _ ID). The global ID may uniquely identify a packet within the network. In particular, the global ID identifies the final packet to which the frame must be transferred. The switch ID uniquely identifies the switch within the globally ID identified packet. The endpoint ID field, along with the global ID and switch ID, identifies the edge connected to the network fabric and the endpoint to which the frame is to be transmitted. This field maps to a port or a group of ports on the switch identified by the global ID and the switch ID.

The multicast format includes a multicast ID field (multicast _ ID) of 13 bits. This field is mapped by FRF component 400 to a set of ports on the current switch to which the frame is to be forwarded.

From this information, FRF component 400 determines the updated routing state of the frame and then carries that state within the frame. For example, to implement routing in a dragonfly topology, the current state of a frame (discussed above) may be collected from the VC of the frame. FRF component 400 determines the particular VC to be used for the next hop of the frame to avoid any deadlock based on the algorithmic switch fabric routing rules specified for the switch fabric (the selection of which are described below). Additional routing state information may be provided depending on the location of the frame on its path (e.g., whether the frame is in its source packet, in an intermediate packet, or in its destination packet). It should be noted that FRF component 400 performs port filtering (described in more detail below) using an allow port filter, an operate port filter, a busy port filter, etc. to determine whether the preferred port to which the frame is to be forwarded is currently faulty, busy, absent, etc.

The switches 202 distribute load information among the switches. FRF component 400 receives its load measurement from its associated output port. FRF component 400 receives aggregated load information for neighboring switches from their associated input ports. Each FRF component 400 exchanges load information with all other FRF instances within the same switch. FRF component 400 provides the aggregated load information to its associated output port for transmission to the neighboring switch. Through the load distribution mechanism, each FRF component 400 learns the load measured at each output port of its switch. In addition, each FRF learns the aggregated load information of all neighboring switches.

It should be noted that FRF component 400 may support frame multicasting. When a multicast DFA is received, FRF component 400 determines a set of ports to which frames associated with the multicast DFA should be forwarded. The set of ports may be determined by accessing a lookup table that maps software configured multicast fabric addresses to output ports. This helps to avoid problems associated with duplicate copies of multicast frames.

Fig. 5 illustrates an example routing process involving down-selection of candidate ports and load-based adaptive routing. FRF component 400 considers three types of candidate ports to which a frame may be forwarded: a preferred minimum path candidate port 502; non-preferred minimum path candidate ports 504; and non-minimum path candidate ports 506. The non-minimum path candidate port is either a global non-minimum path candidate port or a local non-minimum path candidate port depending on the position of the frame on its path.

Filtering may be applied to these three types of candidate ports, such as operational port filtering, available port filtering, and busy port filtering. Port filtering as applied herein may be used to reduce a set of valid ports that are considered path candidate ports by identifying and excluding absent and/or failed and/or busy ports from consideration.

Operational port filtering (or non-operational port filtering) may refer to removing non-operational ports from groups of ports considered as route candidates (e.g., preferred minimum path candidate port 502, non-preferred minimum path candidate port 504, and non-minimum path candidate port 506). That is, the switch 202 may identify some ports as non-operational. These non-operational ports may be reported in a non-operational port mask. It should be noted that in some embodiments, software may use the non-operational port CSR to force additional ports of the switch 202 to be considered non-operational, e.g., when the port(s) are disconnected due to scheduled maintenance.

Available (or unavailable) port filtering may involve filtering out the following candidate ports for consideration: the candidate port is generally acceptable, but due to a failure within network 100, for example, the candidate port has become unacceptable/unavailable to reach one or more destination switches, destination (switch) packets, etc., but is still acceptable or available to reach one or more other destination switches. In some embodiments, global failure table 406A may be used to block global minimum path port candidates and global non-minimum path port candidates depending on the destination packet to which the frame is routed. For example, when routing a frame to a particular destination (switch) packet, candidate ports leading to intermediate (switch) packets without connecting to the destination packet may be excluded from consideration, but for other destination packets, the same candidate ports may not necessarily be blocked. The global fault table 406A may be determined or indexed by the global _ id field of the DFA of the frame.

In some embodiments, the exception list maintained by exception list table 404A may be used to conditionally exclude port candidates depending on the destination packet or switch to which the frame is routed. It should be noted that the exception list table 440A may be used to identify a preferred global minimum path port. Thus, when the exception list table is not used to identify the preferred global minimum path port, the exception list table 404A is used to exclude candidate ports.

It should be noted that knowledge of which ports in a neighboring switch are busy can be used to determine whether a port connected to the neighboring switch is a poor candidate for receiving a frame based on whether the neighboring switch subsequently needs to forward the frame to an already busy port. For example, when considering a candidate port for global minimum routing, a port connected to a neighboring switch is a poor candidate if the global ports of the neighboring switches connected to the destination packet of the frame are all busy. Similarly, when considering candidate ports for local non-minimum routes in a destination packet, a port connected to a neighboring switch of the destination switch of a frame is a poor candidate if the local ports of the neighboring switch are all busy.

Thus, busy port filtering may be performed by FRF component 400 by using a busy port mask to remove heavily loaded ports rather than being considered candidate ports. It should be noted that in some embodiments, heavily loaded ports are excluded from consideration when there are other candidate ports that are not heavily loaded. Otherwise, busy port filtering does not exclude heavily loaded ports from consideration when there are no less heavily loaded ports. FRF component 400 maintains four busy port (i.e., port whose load exceeds a software defined threshold) masks: a local switch busy port mask; a global non-minimum busy global port mask; a global non-minimum busy local port mask; the remote switch busy port mask. The remote busy port mask and other load information is communicated between the switches to populate the remote switches.

The local switch busy port mask may be applied to the smallest path candidate port as well as to the local non-smallest path candidate port. FRF generates 64 bits ls busy port mask by comparing the local _ load of each port to a software defined threshold. Ports with loads above the threshold are marked as busy in the mask.

A global non-minimum busy port global mask may be applied to global ports of the global non-minimum path candidate ports. FRF generates a 64-bit gnmbgp _ mask by comparing the gnmgp _ load of each port to a software-defined threshold. Ports with loads above the threshold are marked as busy in the mask.

A global non-minimum busy local port mask may be applied to local ports of the global non-minimum path candidate ports. The FRF generates a 64-bit gnmblp _ mask by comparing the gnmblp _ load of each port to a software-defined threshold. Ports with loads above the threshold are marked as busy in the mask.

The destination packet dependent busy port mask obtained from the remote switch busy global port table may be applied to the global minimum path candidate port. Accordingly, the destination switch associated busy port mask obtained from the remote switch busy local port table may be applied to the local non-minimum path candidate port when routing the frame in its destination packet.

A surviving set of path candidate ports 508 may be generated when the filtering or down-selection phase described above is applied. That is, after removing non-operational and unavailable, heavily loaded port candidates, a reduced number of candidate ports may be determined, leaving a set of path candidate ports. In some embodiments, the pseudo-random selection process is used to further reduce the number of surviving path candidate ports 508 to the determined port threshold number associated with each candidate port class (preferred minimum path candidate port, non-preferred minimum path candidate port, and non-minimum path candidate port). In some embodiments, the threshold number of candidate ports may be four candidate ports per category. If the minimum threshold number of candidate ports is not met, then any candidate ports in the category are not excluded from consideration.

In some embodiments, this pseudo-random selection (or down-selection) of candidate ports may be weighted. As will be described below, this weight may be applied to offset potential bias in distributing global non-minimum traffic among the global links of the packet. That is, each port may be assigned a weight, e.g., between 0 and 15, per CSR configuration. This weighting may be used to influence the probability of selecting each candidate port so that higher weighted ports have a greater chance of being selected. For example, a weight of 15 would result in a port being selected 15 times more likely in the pseudo-random selection process. In some embodiments, candidate ports may be filtered into four packets (GW1, GW2, GW4, GW8) based on their assigned weights, wherein the candidate ports may belong to multiple packets depending on the assigned weights, e.g., a candidate port with weight 1 belongs to only one weight packet while a candidate port with weight 5 belongs to two packets (GW1 and GW4, i.e., 1+ 4-5), and a candidate port with weight 15 belongs to all four packets (GW1, GW2, GW4, GW8, i.e., 1+2+4+ 8-15). The number of candidate ports in each packet may be determined (nW1, nW2, nW4, nW8) and a pseudo-random selection applied to each packet to select one candidate port from each packet (cW1, cW2, cW4, cW 8). The weight of each packet and their total weight can be calculated: wW1 ═ nW 1; wW2 ═ 2 × nW 2; wW4 ═ 4 × nW 4; wW8 ═ 8 × nws; total is wW1+ wW2+ wW4+ wW 8. A fifth pseudorandom selection may be performed to select the number j in the range 0 … Wtotal-1. Based on the value of j, one of the candidates cW1, cW2, cW4, cW8 is selected as a down-selected candidate port as follows: if j < wW1, then cW1 is selected; otherwise if j < wW1+ wW2, then cW2 is selected; otherwise if j < wW1+ wW2+ wW4, then cW4 is selected; otherwise, cW8 is selected.

It should be noted that the same candidate port may be selected by more than one instance/iteration of instances/iterations of weighted pseudo-random down-selection logic 410C, effectively reducing the number of candidate ports that are selected. The probability that the same candidate port is selected by more than one of the 4 global non-minimally weighted pseudo-random selectors in mgnm decreases as the number of candidate ports to be selected increases. In the context of a network with, for example, a dragonfly topology, and with full global bandwidth, there may be approximately 48 possible global non-minimum candidate ports at an edge port in a source packet: 16 global ports and 32 local ports. If the local hop is already occupied, the next hop is a global hop, thereby reducing the number of candidate ports to about 16. However, if the network is tapered such that it only supports one quarter of the full global bandwidth, there may be only 4 global candidates to choose from after the local hop has been occupied.

FRF component 400 may generate a remote switch busy global port table of busy port masks described above using the received remote switch busy port masks that identify ports connected to neighboring switches that should be avoided as global minimum path candidates based on the destination packet to which the frame is routed. Similarly, a remote switch busy local port table of the above busy port masks may also be generated using the received remote switch busy port masks that identify ports connected to neighboring switches that should be avoided as local non-minimum path candidates when routing in a destination packet based on the destination switch to which the frame is routed.

rs busy port mask is used to evaluate the suitability of a neighboring switch (whether its port is busy or quiet) to reach a particular destination packet via a global minimum path and to reach a particular destination switch via a local non-minimum path or via a local minimum path. Each FRF instance corresponding to a local port or a global port may be configured to generate a 64-bit rs _ busy _ port _ mask. The generated mask is transmitted to a partner switch connected to the port. Similarly, the partner switch may also generate and return rs _ busy _ port _ mask.

Each FRF instance transmits rs _ busy _ port _ mask it received from its partner switch to the FRF instances in all other ports of the switch using the port state ring. The port state ring carries load and state information between all FRF instances in the switch. Each FRF instance captures all rs busy port masks so that all FRF instances learn the remote busy port status provided by all neighboring switches. Each FRF instance uses the rs busy _ port _ mask it receives to generate a Remote Switch Busy Global Port (RSBGP) table and a Remote Switch Busy Local Port (RSBLP) table.

The generation process of rs _ busy _ port _ mask is divided into two steps. The first step is to compare the local _ load of each port to a software configurable threshold, generating an intermediate mask of all the ports that are each busy. Ports classified as non-operational are also recorded as busy in the intermediate mask. The second step considers link bundles such that a port is marked as busy in rs busy _ port _ mask only if that port and all other ports that are part of the same bundle are marked as busy in the intermediate mask. All ports that are members of the same bundle are marked as busy or not marked in rs busy _ port _ mask. Global ports that are part of the same bundle are all connected to the same remote packet. Local ports that are part of the same bundle are all connected to the same remote switch within the current packet.

When rs busy port mask is used to determine whether the switch generating the mask is a good candidate for routing a frame to another packet or another switch in the current packet, bundling is used to provide a consistent view of the applicability of the generating switch when the busy states on its links connected to the destination switch in the destination packet or current packet are inconsistent. The basic principle of the bundling process described herein is that the switch that generates rs busy port mask is still a candidate for reaching the destination switch in the destination packet or current packet, as long as the switch has at least one non-busy link to the destination packet or switch; adaptive routing at the switch that generated rs busy port mask should direct the frame to a non-busy link.

Ports must be included in either the bundled port mask CSR or the unbundled port mask CSR (both CSRs are part of the static description of the wire) in order to mark these ports as busy in rs busy port mask. The bundle port masks are scanned to identify the bundles and ports they contain. In addition, the unbundled port mask is queried to identify any other ports that are not members of the bundle, but whose busy state should also be included in the generated rs _ busy _ port _ mask.

A different software defined threshold is used in calculating rs busy port mask because of the greater latency involved in transmitting and processing rs busy port mask and transmitting frames bound by the mask to the remote switch that generated the mask. Due to the greater latency, it may be useful to require the port to be more loaded before the port is considered so busy that it is not a good candidate to receive additional frames from the remote switch. The busy remote port should be loaded enough so that the port remains loaded for the entire time it takes to receive a frame that has been mask bound.

The RSBGP table stores a busy port mask indexed by a destination packet (global _ id). The RSBGP table is dynamically populated using information contained in the received rs busy port mask and topology information related to the system stored in the configuration state. Also, the RSBGP table is used to evaluate global minimum paths consisting of hops to neighboring switches having global ports connected to the destination packet to filter out the ports of the current switch that are used when arriving at the destination packet are a poor choice because the corresponding global port or ports of the neighboring switches that the filtered ports of the current switch arrive at are heavily loaded.

The RSBLP table stores a busy port mask indexed by the destination switch (switch _ id) and may also be used to evaluate local non-minimum paths consisting of a local hop to a neighboring switch followed by another local hop to the destination switch. The RSBLP table is dynamically populated using the information contained in the received rs busy port mask and topology information related to the system stored in the configuration state. For topologies such as fat-tree, where a local minimum path may consist of a local hop to a neighboring switch followed by another local hop to the destination switch, the RSBLP table may also be used to evaluate the local minimum path. The RSBLP table is used to filter out ports on the current switch, which is a poor choice when indirectly reaching the destination switch because one or more ports of neighboring switches connected to the destination switch are heavily loaded.

It should be noted that the RSBGP and RSBLP tables are never accessed simultaneously for the same route request. The RSBGP table is accessed when frames are not routed in the destination packet and the RSBLP table is accessed only when frames are routed in the destination packet. Thus, both may be implemented within the same memory (referred to as the remote switch busy port table).

The busy port filter is used to remove the candidate smallest path port that is a poor choice because it causes the path to be heavily loaded. Separate copies of the busy port filter are applied independently to the preferred minimum path port and the non-preferred minimum path port. If global minimum routing is enabled, the first stage of the filter removes ports marked as busy in the index entry of the remote switch busy global port table. If local minimum routing is enabled in the destination packet, an optional first stage of the filter will remove ports marked as busy in the index entries of the remote switch busy local port table or remote switch busy global port table, where the table used is based on the CSR configuration. When in the destination group, the first phase of the filter is not applicable and not used with the dragonfly topology; however, this first stage may be applicable to other topologies, including fat trees. The second stage of the filter removes the ports marked as busy in the ls busy port mask. If no candidates remain after the filter second stage, the filter second stage is disabled so that the only remaining candidate ports (if any) will be considered to have high load during the adaptive selection stage.

A local non-minimum busy port filter is used in the destination packet to remove local non-minimum candidate ports that are poor choices because of heavy resulting path load. The first stage of the filter removes ports marked as busy in the index entries of the remote switch busy local port table or remote switch busy global port table, where the table used is based on the CSR configuration. The second stage of the filter removes the ports marked as busy in the ls busy port mask. If no candidates remain after the filter second stage, the filter second stage is disabled so that the only remaining candidate ports (if any) will be considered to have high load during the adaptive selection stage.

A global non-minimum busy port filter is used when global non-minimum routing is enabled to remove candidate global non-minimum path ports that are poor choices because of causing the path load to be severe. Specifically, the global port marked as busy in the gnmbgp _ mask is removed, and the local port marked as busy in the gnmblp _ mask is removed. If no candidates remain after applying these masks, both masks are disabled so that the only remaining candidate ports (if any) will be considered to have high load during the adaptive selection phase.

The various busy port filters (busy port filter, local non-minimum (LN) busy port filter, global non-minimum (GN) busy port filter) may not be allowed to block all candidate ports in common provided there is at least one valid candidate port. If there are viable candidate port choices, they are allowed if there are no better choices, even if they are "poor" choices. Otherwise, an empty route response will be generated for the route request when the route is actually available.

To prevent the generation of incorrect null route responses, the first stages of the preferred and non-preferred minimum path busy port filters (fig. 4A) and the first stage of the local non-minimum path busy port filter (fig. 4B) are all disabled if the following conditions are all true: no candidate survives the first phase of the preferred minimum path busy port filter (busy port filter); no candidates survive the first phase of the non-preferred minimum path busy port filter (busy port filter); no candidates survive the first phase of the local non-minimum busy port filter (local non-minimum (LN) busy port filter); and no candidate survives the global non-minimum busy port filter (global non-minimum (GN) busy port filter).

It should be noted that if the minimum route is disabled, there will be no minimum path candidate port (port filter allowed). Only if local non-minimum routing is enabled, there will be local non-minimum path candidate ports (candidate local non-minimum path ports). Only if global non-minimum routing is enabled will there be global non-minimum path candidates (candidate global non-minimum path ports). Local and global non-minimal routing are not typically enabled at the same time. When the first stages of the preferred and non-preferred minimum path busy port filters and the local non-minimum busy port filters are disabled due to the above conditions, the only candidate ports that will be seen in the adaptive selection stage (described below) will be poor candidates because these candidate ports will all be ports to other switches whose ports (which are connected to the destination packet or destination switch) are heavily loaded. This is because these ports are the only candidate ports blocked by disabled filters, and there are no other candidates without disabling these filters.

The adaptive selection phase will select between these remaining/surviving candidate ports based on their biased local loads, but their local loads do not necessarily reflect the reason for their poor performance. Their poor characteristics may be due to high downstream loads on some ports of other switches that these candidate ports reach. This is because the adaptive selection phase may not see how bad the candidates are, so coordination between different busy port filters as described herein may be used. If each busy port filter independently decides whether to disable its RSBGP based table and RSBLP based filter, the following may occur. A non-preferred minimum path busy port filter may produce one or more bad candidates without disabling any of its filter stages. A preferred minimum path busy port filter may only be able to generate one or more candidate ports by disabling both of its filter stages. Therefore, the preferred minimum path busy port filter can produce all the bad candidates. In the adaptive selection phase, the down-selected, bad, non-preferred minimum path candidate port is compared with the down-selected, bad, preferred minimum path candidate port. However, the adaptive selection stage lacks poor visibility into the preferred minimum path candidate, and therefore it may select a poor preferred minimum path candidate instead of a poor non-preferred minimum path candidate.

An alternative mechanism to the busy port filter coordination mechanism described herein is that all busy port filters act independently, but the minimum path and local non-minimum path busy port filters each forward a signal to the adaptive selection stage to indicate whether their respective candidate ports are poorly selected because the ports of the downstream switch are busy. If so, the adaptive selection function may lower the priority of its candidate ports to support other ports. The results will be the same as those achieved through coordination between the various busy port filters described herein.

As illustrated in fig. 5, load-based adaptive selection may be performed on surviving path candidate ports 510 remaining after FRF component 400 performs the pseudo-random selection process. The adaptive selection stage will produce a single, least loaded candidate port 512 to which frames can be routed, wherein the current load present on the candidate port surviving the pseudo-random down-selection (surviving path candidate port 508) is compared to determine the least loaded candidate port in the set of remaining candidate ports.

In some embodiments, the preferred minimum path candidate port is selected over the non-preferred minimum path candidate ports, and the minimum path candidate port is selected over the non-minimum path candidate ports. To accomplish this preference, an offset value may be added to the load of each candidate port before performing the adaptive selection comparison. The bias value used may be configured using CSR and may vary depending on the type of path to which the bias value applies (i.e., non-preferred minimum, and non-minimum), the traffic classification of the frame being routed, and the location of the frame on its path. For example, frames belonging to a low latency traffic classification may be more biased toward the smallest path than frames in other traffic classifications to be more likely to achieve/meet low latency requirements or requirements. The more a frame may be biased toward minimum path routing, the closer the frame is to its destination.

In particular, the load value represents a busy condition for a port of switch 202 and is used to evaluate a load-based port mask and compare candidate ports during an adaptive routing process. A load-based port mask is used in the busy port filter to remove ports from the set of candidates under consideration that are poor candidates based on the current load. A plurality of different types of load values are used within the switch and some of the load values are communicated to adjacent switch devices. These load values are described in the following section, note that fig. 6 illustrates how load values are measured, transmitted, and used.

When there is more than one port to which a frame can be routed, several load metrics are calculated and used to determine which port to route the frame. The load metrics are also used to generate busy port masks, which, as described above, are used to exclude heavily loaded ports from consideration.

Five load metrics are described herein: a local load; a packet load; global non-minimum global port loading; average global load; and global non-minimum local port loading.

With respect to local loads, the load of each output port of the switch (e.g., output port 220c of switch 202) is constantly evaluated and provided as an 8-bit value local _ load to the corresponding FRF instance. A larger value indicates a higher load. The current load present at each output port is measured by the output control age queue block. The output port load is provided by each age queue instance to the FRF instance (of FRF component 400) associated with the input side of the same port. The load value provided to the FRF is an 8-bit value, referred to as local _ load. The age queue determines the local load based on a combination of the amount of traffic queued, waiting to leave the port, and the amount of traffic queued at the other side of the link in the input buffer of the link partner switch device. Each FRF instance distributes the local _ load value that FRF instance receives from its associated age queue instance to all other FRF instances. In this way, each FRF instance learns the current local _ load for each output port.

When comparing port loads of candidate ports to determine the best port to route a frame to, the local _ load values of the ports are used for the port under consideration for minimum path routing and for the port under consideration for local non-minimum path routing.

Packet load is a metric that measures how appropriate a dragonfly packet is to be used as an intermediate packet in a global non-minimum path. The 8-bit packet load value is not calculated by a switch, such as switch 202, but is software configurable. The software may use a measure of the network injection load present across the input side of the edge ports of the packet to derive a group _ load value, or may determine the group _ load value based on a policy that discourages (perhaps based on jobs or services running in the packet) the use of certain packets as non-minimal intermediate packets. That is, the group _ load value is intended to represent the amount of local traffic within a packet.

The network management stack sets the group _ load value by periodically writing to the CSR. The software configured packet load value is transmitted over the global link. FRF instances associated with the global link forward these FRF instances to all other FRF instances in the switch from the group _ load values they receive from their link partners in packets at the other side of the link. In this way, each FRF instance learns the packet load values of packets at opposite ends of each global link where the switch terminates.

In terms of global non-minimum global port load (gnmgp _ load), this load is a metric used to evaluate the suitability of a global port for directing frames to intermediate packets arriving on a global link connected to the global port. gnmgp _ load is nominally equal to the maximum of the global port's local _ load and the group _ load value received for packets arriving from the global link. However, with CSR configuration, the group _ load component may be excluded.

When comparing port loads of candidate ports to determine the best selected port to route a frame to, the value of gnmgp _ load for a port is used for the global port under consideration for global non-minimum path routing.

The average global load (mean _ global _ load) is intended to be used to evaluate the suitability of a switch for reaching any intermediate packets directly connected to the switch. The mean _ global _ load value is an 8-bit value equal to the arithmetic mean of the gnmgp _ load values for all switch global ports. Ports classified as non-operational by hardware or software are excluded from the computation. The port to be included in the mean _ global _ load calculation is determined by the CSR configuration.

For any port whose load is included in the mean _ global _ load calculation, if the port's group _ load value is not received (either because the port's link partner data is not received or because the link partner data being received is not global link data), then the port's contribution to the mean _ global _ load is based only on the port's local load. The calculated mean _ global _ load value is transmitted over the local link. FRF instances associated with the local links forward mean _ global _ load values that these FRF instances receive from their link partners in the local switch at the other side of the link to all other FRF instances in the switch. In this way, each FRF instance learns the mean _ global _ load values of local switches at opposite ends of each local link terminated by a switch and can use these values for global non-minimum path selection.

A global non-minimum local port load metric is calculated by each FRF instance. Global non-minimum local port load (gnmlp _ load) is an indicator for evaluating the suitability of a local port for intermediate packets directing frames to a global non-minimum path. Global non-minimum local port load takes into account the load on the local port and the suitability of the local packet switch to which the port is connected for reaching intermediate packets. The value of gnmlp _ load for a port is equal to the maximum of the local _ load for the port and the mean _ global _ load reported by the partner switch for the port. Through software configuration, the mean _ global _ load component can be removed so that gnmlp _ load of a port becomes simply equal to local _ load of the port.

The gnmlp _ load value is an 8-bit value. Each FRF instance calculates the gnmlp _ load values for all switch ports of the FRF instance at which link partner data for the local link is received. If no load status information is received from the port's partner switch, then the port's gnmlp _ load value is set equal to the port's local _ load.

Fig. 6 illustrates example load measurements and how the load measurements may be exchanged between switches in a packet. Fig. 6 illustrates a set of switches, e.g., packet 1 including switches 602, 604, and 606, each of which may be an embodiment of switch 202 (fig. 2). As described above, the packet load values may be switched over the global link and, as shown, the group _ load value is transmitted/received from/by each of the switches 602, 604, and 606 of the other packets/switches in the switch fabric. Within packet 1, switches 602, 604, and 606 exchange a mean _ global _ load value. Switch 602 is also shown calculating a gnmlp _ load value at the output side of a local port based on local _ load at the port and mean _ global _ load reported by the link partner. Still further, the local _ load value at the output of all ports is calculated by switch 602. It should be noted that the mean _ global _ load and group _ load values are sent symmetrically from each switch to the other over the link between the connected switches.

The fabric routing process described in the previous section is performed on each frame received. The switch determines whether to perform packet-by-packet adaptive routing (using the recommended port selection) or flow-based adaptive routing (where the port determined for the first packet in each flow is selected for all packets in that flow) according to the ordering requirements of the traffic.

As already described herein, switches such as switch 202 support minimum and non-minimum path routing. It should be understood that the minimum path is based on the destination. If the destination NIC is local, an output port connected to the destination switch is selected. If the destination is in another packet, the packet is routed to a switch within the local packet connected to the destination packet. In large systems there may be only one such path, but in small systems there may be many such paths, some connected to the incoming switches and others connected to switches elsewhere within their packets. The input switches select between them.

The candidate ports considered for minimum path routing are further divided into preferred and non-preferred port groups, where the ports in the preferred group may result in paths containing fewer hops. The non-minimal path routes packets via an intermediate switch (referred to as a root switch). The root switch is selected on a packet-by-packet or flow-by-flow basis depending on the ordering requirements of the traffic.

Non-minimal traffic is routed "up" to the root switch and then minimally "down" to the destination. In some embodiments, the intermediate root switch is randomly selected to distribute the load evenly. A network (e.g., network 100) provides control over intermediate packet selection, enabling traffic to be routed towards intermediate packets that are known to be lightly loaded or away from packets having a particular function or that are known to be heavily loaded. The root switch may be distributed among all packets, where the non-minimum path may detect the root switch in the source packet, the destination packet, or any intermediate packet. The global non-minimal route takes an indirect path through the root switch in the intermediate packet. These paths require two global hops, one from the source packet to the intermediate packet and one from the intermediate packet to the destination packet. A global non-minimum path requires a maximum of three local hops, one per group. Regardless of system size, the maximum path length is typically five switch-to-switch hops. Although unlikely, a maximum of two additional local hops can be employed to address a failure in the network.

Since the path is shorter, a minimum route is preferred and thus the network load is lower. However, a single minimum route may result in poor performance for certain traffic patterns, such as when all nodes in one packet communicate with nodes in another packet. Achieving good performance in a wide range of traffic patterns requires a mix of minimal and non-minimal routes.

At each hop along the path of the frame, the routing mode that may be used to advance the frame along its next hop is controlled by the configuration of FRF routing algorithm table 408. When a frame is received at a switch input port, determining a path type that can be forwarded with the frame: local minimum, global minimum, local non-minimum, and global non-minimum. The set of output ports to which a frame may be forwarded is determined by the type of path allowed at that point.

The type of path that is allowed to be taken depends on where the frame is in its journey between the ingress and egress ports of the network. The path types are as follows. The local minimum path selects the link that is directly connected to the frame's destination switch and may be used when the frame is in its destination packet. When the frame is not in its destination packet, a global minimum path may be used and a global link directly connected to the frame's destination packet or a local link connected to a switch having a working global link directly connected to the frame's destination packet is selected. A local non-minimum path may be used when a frame is in its destination packet or when a frame is in an intermediate packet. The local non-minimum path selects local links connected to other switches in the packet regardless of the destination of the frame. When in a destination packet, after making a local non-minimum hop, the destination switch of the frame must be able to be reached within another hop. When in an intermediate packet, after making a local non-minimum hop, it must be possible to reach the switch with the working global link connected to the destination packet of the frame within another hop. A local link cannot be selected that is connected to a switch for which such is not possible.

A global non-minimum path may be used when a frame is in its source packet and not in its destination packet. The global non-minimum path selects either a global link connected to other packets or a local link connected to other switches in the source packet, regardless of the destination of the frame. A global link must be selected only when it is connected to a packet that has a working link connecting it to the frame's destination packet. Similarly, local links must be selected only when they are connected to switches with global links (these global links are themselves valid global non-minimum path choices).

Adaptive routing selects between the minimum and non-minimum paths (as described above) based on their current load.

In a minimum routing aspect, a local minimum route is generated by looking up the switch _ id field of the destination fabric address in the local switch minimum table when in the destination packet rather than at the destination switch (fig. 4A). The lookup returns a set of valid links. The local switch minimum table contains 128 entries of 64 bits each, where each bit represents one possible output port. At the destination switch, an egress port or port selection is generated by looking up the endpointid field of the destination fabric address in a local port minimum table. The lookup returns a set of valid links. The local port minimum table contains 64 entries of 64 bits each, where each bit represents one possible output port.

The global minimum route is generated by looking up the global _ id field of the destination fabric address in the global minimum table (fig. 4A). The lookup returns a set of valid links. The global minimum table contains 512 entries, 64 bits each, where each bit represents one possible output port.

The local minimum path takes at most one switch-to-switch hop from the source switch to the destination switch, both within the same packet. There may be several such paths. The local non-minimal path takes two switch-to-switch hops, both from the source switch to an intermediate switch (called the root switch) and from the intermediate switch to the destination switch. There are many such paths.

The global minimum path takes one global hop from the source packet to the destination packet. There is at most one local hop in both the source packet and the destination packet. Regardless of system size, the global minimum path requires a maximum of three switch-to-switch hops.

In some system configurations where there are multiple global links connecting a source packet to a destination packet, an offset may occur such that the proportion of traffic injected at the source packet is distributed to each global link is not equal. As an example, consider the case of switch A, B, and C (both in group X), where switch B has three global links connecting it to packet V, and switch C has two global links connecting it to packet V. If the injected traffic at switch a destined for packet V is distributed equally between switches B and C, each of the two global links of switch C will be more heavily loaded than each of the three global links of switch B.

To be able to offset the bias, the global minimum table may be divided into several blocks, each of which is able to generate an efficient set of global routing options for any destination packet. The blocks for the service request are pseudo-randomly selected by the FRF on a frame-by-frame basis. Within each global minimum table block instance, only a subset of the possible candidate ports that can be used to reach the destination packet are populated. The subset is selected in a manner that cancels the bias. The subset populated may vary between different block instances such that all possible candidate ports can be used.

A global minimum path between an edge port in one packet and an edge port in another packet may require one, two, or three hops across fabric links. A hop is if the switch in the source packet that contains the ingress edge port has a global link that directly connects it to the switch in the destination packet that contains the egress edge port. A frame is two hops if it traverses a global link between two packets that are directly connected to an ingress switch in a source packet or an egress switch in a destination packet. In this case, one hop across the local link in the source or destination packet is also required. Finally, three hops are required if the global link through which the frame passes is not directly connected to either the ingress switch in the source packet or the egress switch in the destination packet. In this case, hops across the local link are also required in both the source and destination packets. Local link hops in the source packet bring the frame from the ingress switch to the source packet switch connected to the global link. Local link hops in the destination packet bring the frame from the destination packet switch connected to the global link to the egress switch.

In identifying the minimum path candidate port, the FRF can classify the candidates as a set of preferred ports and a set of non-preferred ports. Preferred ports are those that allow a global minimum path requiring two or fewer fabric link hops. Non-preferred ports are all the smallest path candidates that are not classified as preferred. Using the preferred path when available and not heavily loaded reduces the average load on the local fabric links of the system because it reduces the average number of local fabric links traversed per frame. The use of the preferred path may also reduce the end-to-end fabric delay experienced by the frame.

Any local link may be a candidate when performing local non-minimal routing. However, if some local links lead to switches that may not be able to reach the destination because of the presence of links within the system or switch failures, it may be desirable to exclude these local links from consideration. The CSR controls which ports are candidates for local non-minimum path routing.

When performing global non-minimum routing, generally any global link may be a candidate. In addition, any local link that would normally reach a switch with an operational global link may also be a candidate. However, if some links lead to switches or packets that may not be able to reach the destination packet because there is a link or switch failure within the system, then these links may need to be disregarded. The CSR controls which ports are candidates for global non-minimum path routing.

In selecting candidate ports for global non-minimum routing, if all candidate ports (global and local) are equally likely to be selected, for many system configurations, global non-minimum traffic will not be evenly distributed in the global link out of the packet. For example, consider the case of three switches A, B, and C within a packet, where each switch is connected to each other by four local links, and switch A and switch B both terminate 14 global links, and switch C terminates 16 global links. Ingress traffic arriving at an edge port of switch a may be globally non-minimally routed to any global or local link terminated by switch a. If the ingress traffic is routed to any of these links with equal probability, each global link terminated by switch B will only receive 4/14(4 local links to 14 global links) of traffic routed to each global link terminated by switch a. Likewise, each global link terminated by switch C will receive 4/16(4 local links to 16 global links).

To counteract this potential bias in global non-minimum traffic distribution among the global links of a packet, when a global non-minimum set of candidate ports is pseudo-randomly selected down to a small number of ports that will participate in the adaptive routing phase, a weight may be applied to each candidate port so that some candidate ports will be more likely to survive the down-selection process than other candidate ports.

Adaptive routing selects between these minimum and non-minimum paths based on their load. Adaptively routed traffic starts with the smallest path and goes to the non-smallest path if the load on the smallest path is high (this is called progressive adaptive routing). It can be said that such a path has diverged.

A non-minimum path may be selected at an injection point or an egress router in the source packet. Local non-minimal routing may be employed within a source packet, an intermediate packet, or a destination packet. The dragonfly routing algorithm allows non-minimal paths in both the intermediate packet and the destination packet (considering the case where all incoming traffic on a global port of a particular switch is destined for a NIC on another switch in the packet). However, in general, non-minimum traffic is distributed well enough to avoid this, but in the presence of errors on the local link, additional hops in the destination packet may still be beneficial. Upon reaching the intermediate packet, the packet may take a minimum route to the destination packet or a local non-minimum route to a switch having a path to the destination packet.

Again, this decision is made based on load. After a hop is made within the intermediate packet, the packet must detect the root and take the smallest path to the destination. The adaptive decision is made based on the load and the bias towards the preferred minimum, non-preferred minimum or non-minimum path. As described above, the more the routing algorithm increases the bias towards the minimum path, the closer the packet is to its destination. The algorithm tends to cross the direct path of the intermediate packet, provided the load is low.

Restricted routing is used at points other than the injection and root detection points to prevent packets from flowing back in the direction from which they came. In the switch, the condition that a packet makes one hop from an injection point and a root detection point is detected, and a global port is ensured to be adopted. For local minimum routing, after a certain hop from the root, the packet will reach the switch connected to the destination NIC. For global minimum routing, after a certain hop from the injection point, the packet will reach the switch with the global link that must be taken. In the intermediate packets, packets are allowed to make local hops at the injection points. From this point on, the packet must follow the minimum path to its destination.

When passing information between switches, an immediate decision must be made on the next hop on the path. The decision is made using information derived from the local state and information communicated from the neighboring switches. Using information from many different sources allows more accurate/efficient decisions to be made. The information includes information from neighbors.

Previous systems carried information about the average load between switches. That is, more detailed information from the relevant or neighboring switch is more helpful. In current switch ASICs, a set of values may include information indicating the state of the output ports of the relevant/adjacent switches. By passing this set of values, better routing decisions can be made. In one example, a flag is transmitted back from the neighboring switch, where the flag has one bit for each output port. For example, for a switch with 64 outputs, a 64-bit flag will be transmitted. This is much more accurate than simply passing the global average of adjacent ports.

Fig. 7A illustrates an example scenario in which the average load is used as a basis for routing packets in a switch fabric that includes a source switch 702, a destination switch 708, and two possible intermediate switches 704 and 706. As illustrated in fig. 7A, based on adaptive routing as described herein, source switch 702 may determine two candidate intermediate switches (704 and 706) through which to route packets from source switch 702 to destination switch 708. When making adaptive routing decisions based on the average load of the intermediate switches 704 and 706, in this example, the intermediate switch 704 may have an average load value of two, while the intermediate switch 706 may have an average load value of three. Making an adaptive routing decision based on the average load value will then result in the source switch 702 selecting to reach the destination switch 708 through the intermediate switch 704 (i.e., route or path 710). However, the link between the intermediate switch 704 and the destination switch 708 may be busy, which may have negative effects, such as latency, as described above.

According to one embodiment, and as described above, a port busy condition may be considered. That is, and referring back to fig. 5, the down-selection and adaptive routing of candidate ports may be based on load. Likewise, FRF component 400 considers three types of candidate ports to which frames may be forwarded: a preferred minimum path candidate port 502; non-preferred minimum path candidate ports 504; and non-minimum path candidate ports 506. The non-minimum path candidate port is either a global non-minimum path candidate port or a local non-minimum path candidate port depending on the position of the frame on its path.

Filtering may be applied to these three types of candidate ports, such as operational port filtering, available port filtering, and busy port filtering. Port filtering as applied herein may be used to reduce a set of valid ports that are considered path candidate ports by identifying and excluding absent and/or failed and/or busy ports from consideration.

In the example of figure 7B, busy port filtering may result in a different path being selected to route a frame from source switch 702 to destination switch 708. Also, knowledge of which ports in the neighboring switches (in this example, intermediate switches 704 and 706) are busy can be used to determine whether a port connected to the neighboring switch is a poor candidate for receiving a frame based on whether the neighboring switch subsequently needs to forward the frame to an already busy port. Therefore, busy port filtering may be performed by FRF component 400 using a busy port mask to remove heavily loaded ports from being considered candidate ports. FRF component 400 maintains four busy port (i.e., port whose load exceeds a software defined threshold) masks: a local switch busy port mask; a global non-minimum busy global port mask; a global non-minimum busy local port mask; the remote switch busy port mask. Information from these masks is transferred between switches to populate remote switches.

Here, the local switch busy port mask may be applied to the minimum path candidate port as well as to local non-minimum path candidate ports, e.g., output ports of intermediate switches 704 and 706. FRF component 400 generates a 64-bit ls _ busy _ port _ mask by comparing the local _ load of each port to a software-defined threshold. Ports with loads above the threshold are marked as busy in the mask. In this example, the output port from the intermediate switch 704 that is directed to the destination switch 708 meets or exceeds the software-defined threshold specified for the switch fabric. Thus, ls busy port mask indicates that the link from intermediate switch 704 to destination switch 708 is busy. On the other hand, because the busy threshold is not met/exceeded, the output port of intermediate switch 706 to destination switch 708 is not marked or marked with ls busy port mask, which reflects that the link between intermediate switch 706 and destination switch 708 is quiet and not busy.

The output port of the intermediate switch 704 that is linked to the input port of the destination switch 708 is considered a poor candidate and is excluded from consideration based on the busy port mask indication that the link between the intermediate switches 704 and 708 is busy. Thus, in contrast to the scenario of fig. 7A in which route 710 is selected based on average load, route or path 712 is selected to route the frame(s) from source switch 702 through intermediate switch 706 to its destination, destination switch 708. It should be appreciated that the quiet link condition between the intermediate switch 706 and the destination switch 708 is a better choice for routing frames from the source 702 despite the higher average load.

Similarly, as described above, FRF component 400 may use the received remote switch busy port mask to generate a remote switch busy global port table and a remote switch busy local port table of busy port masks described above that identify ports connected to neighboring switches that should be avoided as global minimum path candidates based on the destination packet to which the frame is routed and that should be avoided as local non-minimum path candidates based on the destination switch to which the frame is routed. Thus, according to some embodiments, rs busy port mask may also be used to identify ports connected to neighboring switches, which may be busy or quiet. In the example of fig. 7B, the input port of destination switch 708 may be considered busy by the application of rs _ busy _ port _ mask thereto. Such a port may be avoided as a local non-minimum path candidate when routing in a destination switch based on the destination switch to which the frame is to be routed.

Fig. 8 is a flowchart illustrating example operations that may be performed to implement adaptive routing or path selection in accordance with one embodiment of the disclosed technology. As described above, the switch 202 may be an ASIC, and thus provides one or more electronic circuits including electronic components to control or instantiate the FRF component 400 (fig. 4A and 4B) by performing the disclosed example operations.

At operation 800, a plurality of load signals are received from a plurality of neighboring switches, each of the plurality of load signals including a set of values indicative of a load at each of the plurality of neighboring switches providing the load signal, each value within the set of values providing an indication for each of a plurality of links attached thereto as to whether the link is busy or quiet. As described above, a switch, such as switch 202, may have multiple ports (64 ports in some embodiments). Packets may be routed between an ingress and egress port of one switch to one or more other switches (each having its own ingress and egress port). Routing of packets between switches may be performed according to certain characteristics of the switches through which the packets may travel on the path from the source switch to the destination switch.

Between the source switch and the destination switch may be one or more intermediate switches. To select a path, the ports of the switch that packets may traverse are filtered to identify certain path candidate ports having desired characteristics regarding, for example, how busy they are in traffic traversing the ports. Thus, adaptive routing is implemented by determining whether a packet should traverse a path that includes a candidate port for a switch (e.g., an intermediate switch) based on the load (busy indication) associated with the candidate port. In some embodiments, the local load is compared to a threshold that may be specified by the CSR configuration, and ports may be excluded based on whether the local load is below or above the threshold.

At operation 802, an output link for routing the received packet is selected based on the plurality of load signals, and the received packet is routed via the selected output link. That is, a mask may be applied to those ports whose local load exceeds a threshold, i.e., the ports are marked as busy in the mask. It should be understood that other busy port masks may be applied depending on circumstances according to other embodiments, such as a global non-minimum busy global port mask, a global non-minimum busy local port mask, or a destination packet dependent busy port mask. For example, a destination packet dependent busy port mask obtained from a remote switch busy global port table may be applied to the global minimum path candidate port. Accordingly, when routing a frame in its destination packet, the destination switch associated busy port mask obtained from the remote switch busy local port table may be applied to the local non-minimum path candidate port. Thus, the use of load information need not be limited to simply determining the ports of the intermediate switches. When a port is marked busy in the busy port mask, the port may be removed or disregarded as a candidate port. As illustrated in fig. 5, and as described above, the load-based adaptive selection may narrow surviving candidate ports down to the least loaded candidate port. The path along which the packet is routed may be selected to include the least loaded port and the packet is ultimately routed through the selected one of the plurality of switch ports.

It should be noted that the terms "optimize," "optimal," and the like as used herein may be used to mean making or achieving as efficient or as perfect a performance as possible. However, one of ordinary skill in the art reading this document will recognize that perfection is not always possible. Thus, these terms may also encompass achieving or achieving as good or effective or practical a performance as possible in a given situation, or achieving a performance that is better than what may be achieved with other settings or parameters.

Fig. 9 depicts a block diagram of an example computer system 900 in which various embodiments described herein may be implemented. Computer system 900 includes a bus 902 or other communication mechanism for communicating information, one or more hardware processors 904 coupled with bus 902 for processing information. Hardware processor(s) 904 may be, for example, one or more general-purpose microprocessors.

Computer system 900 also includes a main memory 906, such as a Random Access Memory (RAM), cache, and/or other dynamic storage device, coupled to bus 902 for storing information and instructions to be executed by processor 904. Main memory 906 also may be used for storing temporary variables or other intermediate information during execution of instructions to be executed by processor 904. Such instructions, when stored in a storage medium accessible to processor 904, cause computer system 900 to become a special-purpose machine that is customized to perform the operations specified in the instructions.

Computer system 900 further includes a Read Only Memory (ROM)908 or other static storage device coupled to bus 902 for storing static information and instructions for processor 904. A storage device 910, such as a magnetic disk, optical disk or USB thumb drive (flash drive), is provided and coupled to bus 902 for storing information and instructions.

Computer system 900 may be coupled via bus 902 to a display 912, such as a Liquid Crystal Display (LCD) (or touch screen), for displaying information to a computer user. An input device 914, including alphanumeric and other keys, is coupled to bus 902 for communicating information and command selections to processor 904. Another type of user input device is cursor control 916, such as a mouse, a trackball, or cursor direction keys for communicating direction information and command selections to processor 904 and for controlling cursor movement on display 912. In some embodiments, the same directional information and command selections as the cursor control device may be implemented by receiving a touch on the touch screen without a cursor.

Computing system 900 may include a user interface module for implementing a GUI, which may be stored in mass storage as executable software code executed by the computing device(s). By way of example, this and other modules may include components (such as software components, object-oriented software components, class components, and task components), processes, functions, attributes, procedures, subroutines, segments of program code, drivers, firmware, microcode, circuitry, data, databases, data structures, tables, arrays, and variables.

In general, as used herein, the terms "component," "engine," "system," "database," "data store," and the like may refer to logic implemented in hardware or firmware, or to a set of software instructions written in a programming language such as, for example, Java, C, or C + +, possibly with entry and exit points. The software components may be compiled and linked into an executable program, installed in a dynamically linked library, or written in an interpreted programming language such as, for example, BASIC, Perl, or Python. It should be understood that software components may be invoked from other components or from themselves, and/or may be invoked in response to a detected event or interrupt. The software components configured for execution on the computing device may be provided in a computer readable medium, such as a compact disc, digital video disc, flash drive, magnetic disk, or any other tangible medium, or may be provided as a digital download (and may be originally stored in a compressed or installable format that requires installation, decompression, or decryption prior to execution). Such software code may be stored, in part or in whole, on a memory device executing the computing device for execution by the computing device. The software instructions may be embedded in firmware, such as an EPROM. It will be further appreciated that the hardware components may include connected logic units such as gates and flip-flops, and/or may include programmable units such as programmable gate arrays or processors.

Computer system 900 may implement the techniques described herein using custom hard-wired logic, one or more ASICs or FPGAs, firmware, and/or program logic that, in conjunction with the computer system, makes computer system 900 a special-purpose machine or programs it to a special-purpose machine. According to one embodiment, the techniques herein are performed by computer system 900 in response to processor(s) 904 executing one or more sequences of one or more instructions contained in main memory 906. Such instructions may be read into main memory 906 from another storage medium, such as storage device 910. Execution of the sequences of instructions contained in main memory 906 causes processor(s) 904 to perform the process steps described herein. In alternative embodiments, hard-wired circuitry may be used in place of or in combination with software instructions.

The term "non-transitory medium" and similar terms, as used herein, refer to any medium that stores data and/or instructions that cause a machine to operate in a specific manner. Such non-transitory media may include non-volatile media and/or volatile media. Non-volatile media includes, for example, optical or magnetic disks, such as storage device 910. Volatile media includes dynamic memory, such as main memory 906. Common forms of non-transitory media include, for example, a floppy disk, a flexible disk, hard disk, solid state drive, magnetic tape, or any other magnetic data storage medium, a CD-ROM, any other optical data storage medium, any physical medium with patterns of holes, a RAM, a PROM, and EPROM, a flash-EPROM, NVRAM, any other memory chip or cartridge, and networked versions of the medium.

The non-transitory medium is different from but may be used in conjunction with a transmission medium. Transmission media participate in the transfer of information between non-transitory media. For example, transmission media includes coaxial cables, copper wire and fiber optics, including the wires that comprise bus 902. Transmission media can also take the form of acoustic or light waves, such as those generated during radio wave and infrared data communications.

Computer system 900 also includes a communication interface 918 coupled to bus 902. Network interface 918 provides a two-way data communication coupling to one or more network links that connect to one or more local networks. For example, communication interface 918 may be an Integrated Services Digital Network (ISDN) card, cable modem, satellite modem, or a modem to provide a data communication connection to a corresponding type of telephone line. As another example, network interface 918 may be a Local Area Network (LAN) card to provide a data communication connection to a compatible LAN (or WAN component for communicating with a WAN). Wireless links may also be implemented. In any such implementation, network interface 918 sends and receives electrical, electromagnetic or optical signals that carry digital data streams representing various types of information.

The network link typically provides data communication through one or more networks to other data devices. For example, the network link may provide a connection through a local network to a host computer or to data equipment operated by an Internet Service Provider (ISP). ISPs in turn provide data communication services through the global packet data communication network now commonly referred to as the "internet". Both the local network and the internet use electrical, electromagnetic or optical signals that carry digital data streams. The signals through the various networks and the signals on the network links and through communication interface 918, which carry the digital data to and from computer system 900, are example forms of transmission media.

Computer system 900 can send messages and receive data, including program code, through the network(s), network link and communication interface 918. In the Internet example, a server might transmit a requested code for an application program through the Internet, an ISP, local network and communication interface 918.

The received code may be executed by processor 904 as it is received, and/or stored in storage device 910, or other non-volatile storage for later execution.

Each of the processes, methods, and algorithms described in the preceding sections may be implemented in and automated in whole or in part by code means executed by one or more computer systems or computer processors comprising computer hardware. The one or more computer systems or computer processors may also operate to support the performance of related operations in a "cloud computing" environment, or to operate as a "software as a service" (SaaS). The processes and algorithms may be implemented in part or in whole in application specific circuitry. The various features and processes described above may be used independently of one another or may be combined in various ways. Various combinations and subcombinations are intended to be within the scope of the disclosure, and certain method blocks or process blocks may be omitted in some embodiments. The methods and processes described herein are also not limited to any particular order, and the blocks or states associated with the methods and processes may be performed in other orders as appropriate, or may be performed in parallel, or in some other manner. Blocks or states may be added to or removed from the disclosed example embodiments. The performance of certain operations or processes may be distributed among multiple computer systems or computer processors, rather than residing only within a single machine, but being deployed across multiple machines.

As used herein, a circuit may be implemented using any form of hardware, software, or combination thereof. For example, one or more processors, controllers, ASICs, PLAs, PALs, CPLDs, FPGAs, logic components, software routines, or other mechanisms may be implemented to form a circuit. In implementations, the various circuits described herein may be implemented as discrete circuits, or the functions and features described may be shared, in part or in whole, among one or more circuits. Even though various features or functions may be described or claimed separately as separate circuits, these features and functions may be shared between one or more common circuits, and such description should not require or imply that separate circuits are required to implement such features or functions. Where circuitry is implemented, in whole or in part, using software, such software may be implemented to operate with a computing system or processing system (such as computer system 900) capable of performing the functions described with respect to the software.

As used herein, the term "or" may be interpreted in an inclusive or exclusive sense. Furthermore, the description of a resource, operation, or structure in the singular is not to be construed as excluding the plural. Conditional language (such as "can", "might", "migt", or "may", among others) is generally intended to convey that certain embodiments include (while other embodiments do not include) certain features, elements and/or steps unless specifically stated otherwise, or otherwise understood within the context as used.

Terms and phrases used in this document, and variations thereof, unless expressly stated otherwise, should be construed as open ended as opposed to limiting. Adjectives such as "conventional", "traditional", "normal", "standard", "known", and terms of similar meaning should not be construed as limiting the item described to a given time period or item available at a given time, but rather should be read to encompass conventional, traditional, normal, or standard technologies that may be available or known at any time, now or in the future. In some instances, the presence of broad words and phrases such as "one or more," "at least," "but not limited to," or other like phrases, is not to be construed as referring to the narrower case that such broad phrases may not be present.