Patent Grant
11818037
This is generally related to the technical field of networking. More specifically, this disclosure is related to switch devices for facilitating switching in scalable, data-driven intelligent networks.
As network-enabled devices and applications become progressively more ubiquitous, various types of traffic as well as the ever-increasing network load continue to demand more performance from the underlying network architecture. For example, applications such as high-performance computing (HPC), media streaming, and Internet of Things (IOT) can generate different types of traffic with distinctive characteristics. As a result, in addition to conventional network performance metrics such as bandwidth and delay, network architects continue to face challenges such as scalability, versatility, and efficiency.
A switch architecture for a data-driven intelligent networking system is provided. The system can accommodate dynamic traffic with fast, effective congestion control. The system can maintain state information of individual packet flows, which can be set up or released dynamically based on injected data. Each flow can be provided with a flow-specific input queue upon arriving at a switch. Packets of a respective flow are acknowledged after reaching the egress point of the network, and the acknowledgement packets are sent back to the ingress point of the flow along the same data path. As a result, each switch can obtain state information of each flow and perform flow control on a per-flow basis.
In the figures, like reference numerals refer to the same figure elements.
Various modifications to the disclosed embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other embodiments and applications without departing from the spirit and scope of the present disclosure. Thus, the present invention is not limited to the embodiments shown.
Overview
The present disclosure describes systems and methods that accommodate dynamic data traffic with fast, effective congestion control by maintaining state information of individual packet streams. More specifically, packets injected into a network of switches can be categorized into streams, which can be mapped to their layer-2, layer-3, or other protocol-specific header information. Each stream can be marked by a distinctive identifier that is local to an input port of a switch, and provided with a stream-specific input buffer, so that each stream can be individually flow-controlled. In addition, packets of a respective stream can be acknowledged upon reaching the egress point of the network, and the acknowledgement packets can be sent back to the ingress point of the stream along the same data path in the reverse direction. As a result, each switch can obtain state information of active packet streams it is forwarding, and can perform highly responsive, stream-specific flow control. Such flow control can allow the network to operate at higher capacity while providing versatile traffic-engineering capabilities.
In this disclosure, packet streams can also be referred to as “packet flows,” or simply “flows.” The data path traversed by a flow, together with its configuration information maintained by switches, can be referred to as a “flow channel.” Furthermore, the terms “buffer” and “queue” are used interchangeably in this disclosure.
Typically, traffic can be injected into switch fabric 100 via an ingress port of an edge switch, and leave switch fabric 100 via an egress port of another (or the same) edge switch. An ingress edge switch can group injected data packets into flows, which can be identified by flow ID's. The concept of a flow is not limited to a particular protocol or layer (such as layer-2 or layer-3 in the Open System Interface (OSI) reference model). For example, a flow can be mapped to traffic with a particular source Ethernet address, traffic between a source IP address and 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 number), or traffic produced by a process or thread running on an end host. In other words, a flow can be configured to map to data between any physical or logical entities. The configuration of this mapping can be done remotely or locally at the ingress edge switch.
Upon receiving injected data packets, the ingress edge switch can assign a flow ID to the flow. This flow ID can be included in a special header, which the ingress edge switch can use to encapsulate the injected packets. Furthermore, the ingress edge switch can also inspect the original header fields of an injected packet to determine the appropriate egress edge switch's address, and include this address as a destination address in the encapsulation header. Note that the flow ID can be a locally significant value specific to a link, and this value can be unique only to a particular input port on a switch. When the packet is forwarded to the next-hop switch, the packet enters another link, and the flow-ID can be updated accordingly. As the packets of a flow traverses multiple links and switches, the flow IDs corresponding to this flow can form a unique chain. That is, at every switch, before a packet leaves the switch, the packet's flow ID can be updated to a flow ID used by the outgoing link. This up-stream-to-down-stream one-to-one mapping between flow ID's can begin at the ingress edge switch and end at the egress edge switch. Because the flow ID's only need to be unique within an incoming link, a switch can accommodate a large number of flows. For example, if a flow ID is 11 bits long, an input port can support up to 2048 flows. Furthermore, the match pattern (one or more header fields of a packet) used to map to a flow can include a greater number of bits. For instance, a 32-bit long match pattern, which can include multiple fields in a packet header, can map up 2{circumflex over ( )}32 different header field patterns. If a fabric has N ingress edge ports, a total number of N*2{circumflex over ( )}32 identifiable flows can be supported.
A switch can assign every flow a separate, dedicated input queue. This configuration allows the switch to monitor and manage the level of congestion of individual flows, and prevent head-of-queue blocking which could occur if shared buffer were used for multiple flows. When a packet is delivered to the destination egress switch, the egress switch can generate and send back an acknowledgement (ACK) in the upstream direction along the same data path to the ingress edge switch. As this ACK packet traverses the same data path, the switches along the path can obtain the state information associated with the delivery 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 the health of the entire network and fair treatment of the flows. As explained in more detail below, this per-flow queuing, combined with flow-specific delivery acknowledgements, can allow the switch fabric to implement effective, fast, and accurate congestion control. In turn, the switch fabric can deliver traffic with significantly improved network utilization without suffering from congestion.
Flows can be set up and released dynamically, or “on the fly,” based on demand. Specifically, a flow can be set up (e.g., the flow-ID to packet header mapping is established) by an ingress edge switch when a data packet arrives at the switch and no flow ID has been previously assigned to this packet. As this packet travels through the network, flow IDs can be assigned along every switch the packet traverses, and a chain of flow IDs can be established from ingress to egress. Subsequent packets belonging to the same flow can use the same flow IDs along the data path. When packets are delivered to the destination egress switch and ACK packets are received by the switches along the data path, each switch can update its state information with respect to the amount of outstanding, unacknowledged data for this flow. When a switch's input queue for this flow is empty and there is no more unacknowledged data, the switch can release the flow ID (i.e., release this flow channel) and re-use the flow-ID for other flows. This data-driven dynamic flow setup and teardown mechanism can obviate the need for centralized flow management, and allows the network to respond quickly to traffic pattern changes.
Note that the network architecture described herein is different from software-defined networks (SDN's), which typically uses the OpenFlow protocol. In SDN, switches are configured by a central network controller, and packets are forwarded based one or more fields in the 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) headers. In SDN such header-field lookup is performed at every switch in the network, and there is no fast flow ID-based forwarding as is done in the networks described herein. Furthermore, because the OpenFlow header-field lookup is done using ternary content-addressable memory (TCAM), the cost of such lookups can be high. Also, because the header-field mapping configuration is done by the central controller, the setup and tear-down of each mapping relationship is relatively slow and could require a fair amount of control traffic. As a result, an SDN network's response to various network situations, such as congestion, can be slow. In contrast, in the network described herein, the flows can be set up and torn down dynamically based on traffic demand; and packets can be forwarded by a fixed-length flow ID. In other words, flow channels can be data driven and managed (i.e., set up, monitored, and torn down) in a distributed manner, without the intervention of a central controller. Furthermore, the flow ID-based forwarding can reduce the amount of TCAM space used and as a result a much greater number of flows can be accommodated.
Referring to the example shown in
Note that the operations described above can be performed substantially at line speed with little buffering and delay when the first packet is received. After the first packet is processed and scheduled for transmission, subsequent packets from the same flow can be processed by switch 110 even faster because the same flow ID is used. In addition, the design of the flow channels can be such that the allocation, matching, and deallocation of flow channels can have substantially the same cost. For example, a conditional allocation of a flow channel based on a lookup match and a separate, independent deallocation of another flow channel can be performed concurrently in nearly every clock cycle. This means that generating and controlling the flow channels can add nearly no additional overhead to the regular forwarding of packets. The congestion control mechanism, on the other hand, can improve the performance of some applications by more than three orders of magnitude.
At each switch along the data path (which includes switches 110, 106, and 104), a dedicated input buffer can be provided for this flow, and the amount of transmitted but unacknowledged data can be tracked. When the first packet reaches switch 104, switch 104 can determine that the destination fabric address in the packet's fabric header matches its own address. In response, switch 104 can decapsulate the packet from the fabric header, and forward the decapsulated packet to host 116. Furthermore, switch 104 can generate an ACK packet and send this ACK packet back to switch 110. As this ACK packet traverses the same data path, switches 106 and 110 can each update their own state information for the unacknowledged data for this flow.
In general, congestion within a network can cause the network buffers to fill. When a network buffer is full, the traffic trying to pass through the buffer ideally should be slowed down or stopped. Otherwise, the buffer could overflow and packets could be dropped. In conventional networks, congestion control is typically done end-to-end at the edge. The core of the network is assumed to function only as “dumb pipes,” the main purpose of which is to forward traffic. Such network design often suffers from slow responses to congestion, because congestion information often cannot be sent to the edge devices quickly, and the resulting action taken by the edge devices cannot always be effective in removing the congestion. This slow response in turn limits the utilization of the network, because to keep the network free of congestion the network operator often needs to limit the total amount of traffic injected into the network. Furthermore, end-to-end congestion control usually is only effective provided that the network is not already congested. Once the network is heavily congested, end-to-end congestion control would not work, because the congestion notification messages can be congested themselves (unless a separate control-plane network that is different from the data-plane network is used for sending congestion control messages).
In contrast, the flow channels can prevent such congestion from growing within the switch fabric. The flow channel mechanism can recognize when a flow is experiencing some degree of congestion, and in response can slow down or stop new packets of the same flow from entering the fabric. In turn, these new packets can be buffered in a flow channel queue on the edge port and are only allowed into the fabric when packets for the same flow leave the fabric at the destination edge port. This process can limit the total buffering requirements of this flow within the fabric to an amount that would not cause the fabric buffers to become too full.
With flow channels, the switches have reasonably accurate state information on the amount of outstanding in-transit data within the fabric. This state information can be aggregated for all the flows on an ingress edge port. This means that the total amount of data injected by an ingress edge port can be known. Consequently, the flow channel mechanism can set a limit on the total amount of data in the fabric. When all edge ports apply this limit action, the total amount of packet data in the entire fabric can be well controlled, which in turn can prevent the entire fabric from being saturated. The flow channels can also slow the progress of an individual congested flow within the fabric without slowing down other flows. This feature can keep packets away from a congestion hot spot while preventing buffers from becoming full and ensuring free buffer space for unrelated traffic.
Operation of Flow Channel
In general, flow channels can define a path for each communication session across the switch fabric. The path and amount of data belonging to each flow can be described in a set of dynamically connecting flow tables associated with each link of the switch fabric. On every ingress port, edge and fabric, a set of flow channel queues can be defined. There can be one queue for each flow channel. As packets arrive, they either can be assigned to a flow channel on an edge port, or have been assigned to a flow channel by the link partner's egress fabric port on a fabric ingress port. The flow channel information can be used to direct the packets into the appropriate flow channel queue.
By providing flow-specific input queues, the switch can allow each flow to move independently of all other flows. The switch can avoid the head-of-queue blocking problem, which is common with shared input buffers. The flow-specific input queue also allows the packets within a single flow to be kept in order. When a flow passes through the switches, a flow-specific input queue on each input port can be allocated for this flow and these input queues become linked, effectively forming one long queue that reaches across the entire fabric for this flow, and the packets of this flow can be kept in order.
The progress of successful delivery of packets belonging to a flow can be reported by a sequence of ACKs generated by the edge port of an egress switch. The ACK packets can travel in the reverse direction along the data path traversed by the data packets and can be forwarded by the switches according to the forwarding information maintained in flow tables. As ACK packets travel upstream, they can be processed by each switch's input queue manager, which can update the corresponding flow's state information based on information carried by the ACK packets. The ACK packets can have a type field to provide advanced information about the downstream data path, such as congestion. A switch's input queue manager can use this information to make decisions, such as throttling the transmission rate or changing the forwarding path, about the pending data packets currently buffered in its input queues. In addition, the input queue manager can update the information carried in an ACK packet based on state information of a buffered flow, so that the upstream switches can make proper decisions. For example, if an input queue for a given flow is experiencing congestion (e.g., the amount of data in the queue is above a predetermined threshold), the input queue manager can update an ACK packet that is being forwarded to the next upstream switch to include this congestion information.
If an ACK corresponds to the last packet of a flow, a switch can determine that there is no more unacknowledged data for that flow. Correspondingly, the switch can free the flow channel by removing the corresponding entry in the flow table.
As mentioned above, the input queue manager at each switch can maintain information about transmitted but unacknowledged data of a given flow.
In the example in
Correspondingly, because the input queue for the given flow at switch 226 has six queued data units, and two additional data units are in transit between switches 224 and 226, the flow_extent value at switch 224 is 9. Similarly, the flow_extent value at switch 222 is 13, because there are three data units stored in the input queue at switch 224 and one data unit in transit between switches 222 and 224.
In general, a flow channel can remain allocated to a single flow until all the ACKs for all the packets sent on the flow channel have been returned. This means that flow channel table entries can remain active for longer near the fabric ingress edge port than near the egress edge port. If a single packet is injected into the network, a flow channel can be allocated for the ingress edge port and then another flow channel can be allocated for the next fabric link the packet traverses and so on, until the last flow channel is allocated when the packet reaches the last fabric link. Each allocation can generate a flow ID, denoted as variable “flow_id,” to identify the entries of the flow tables of the fabric link. (More details on flow channel tables are provided in the description below in conjunction with
At the input queue of each switch, the flow channel table entries can indicate each flow's state information, including the flow_extent value, from this point downstream to the flow's egress destination edge port. Packets received on the local input port can increase this flow_extent value by the amount of incoming data, and ACKs can reduce the flow_extent by the amount of acknowledged, delivered data.
When a packet reaches the final destination egress port, an ACK packet can be generated and returned for that packet. This ACK can be routed using the data path information stored in the corresponding entry of the flow channel tables at every switch along the data path. Optionally, the ACK packet itself does not need to carry path information and therefore can be small and light weight. If no other data packet is sent on the flow, the ACK can release each flow channel in the reverse order. Once released, the flow channel at each switch can be allocated to a different flow.
If another packet follows the first packet on the same flow, the ACK corresponding to the second packet would need to be received before the flow channel can be released at a given switch. In one embodiment, the flow channel can only be released when ACKs for all the transmitted packets of the same flow have been returned.
Typically, various protocols may require in-order packet delivery. The flow channels can be used to guarantee this delivery order, even when the fabric uses adaptive routing for load balancing across multiple data paths. If packets between an ingress edge port and an egress edge port, perhaps in a different switch on the far side of the fabric, are injected at a very low rate, then each packet injected could reach its destination and return an ACK back to the source before the next packet is injected. In this case, each packet can be a lead packet and free to take any path across the fabric, using the best available dynamic adaptive routing choice. This is possible because the first packet can define the flow's path through the fabric.
Now assume that the packet injection rate is increased slightly to the point where the next packet of the same flow is injected before the current packet's ACK has returned to the source. The second packet can pass the ACK of the first packet somewhere along the flow's data path. Beyond this passing point, the ACK will have released the flow channels allocated to the first packet, because the flow_extent value associated with the first packet is returned to zero when the ACK is processed by the flow channel's logic. Meanwhile, the second packet can now define a new flow, because it is again causing flow channels to be allocated on each of the subsequent fabric links. This second packet, while it is causing flow channels to be allocated beyond the passing point, can be forwarded to a different path based on dynamic adaptive routing. On the other hand, before the passing point, the second packet can extend the outstanding flow created by the first packet to include the second packet. This means the first packet's ACK may not reduce the flow_extent value to zero and the flow channels may remain active before the passing point. It also means that the second packet may follow the exact path taken by the first packet up to the passing point. Note that while it is following the previous packet, the second packet cannot arrive at the egress edge port before the first packet does, and therefore correct packet order can be maintained.
If the injection rate for this flow is increased further, the second packet will pass the first packet's ACK at a location closer to the destination edge port. It is also possible that a third, fourth, fifth, or additional packet may enter the fabric before the first packet's ACK is returned to the source edge port, depending on the data packet injection rate of this flow and the data packet-ACK round trip delay. The maximum packet rate can depend on the size of the packets and the bandwidth of the links. The round trip delay of the data packet and ACK can be an important parameter for a fabric implementation and can be used along with the maximum packet rate to calculate the maximum required number of flow channels for each link. Ideally, a design can provide a reasonable number of unallocated flow channels regardless of the traffic pattern. The demand for the number of flow channels can be high when a large number of packets arriving at an ingress edge port have different destinations and these packets have small sizes and high injection rates. In the most extreme case, each packet could be allocated a different flow channel. These flow channels are freed when the packets' ACKs are returned. Correspondingly, the number of flow channels needed can be calculated as ((Packet rate)*(Average packet to ACK round trip latency)).
Note that packet rate on a single flow channel is not to be confused with packet rate on a link. If the traffic pattern is such that many small packets are being sent to different destinations, then successive packets sent onto the link can have different destinations. This means that each packet could belong to a different flow and could be the only packet to use the corresponding flow channel. In this example, the link can experience a high packet rate, but the packet rate of individual flows can be low. Optionally, a number of ACKs (e.g., 48 ACKs) can be aggregated together into a single ACK frame for transmission over a link and protected by a Frame Check Sequence (e.g., a 32-bit FCS). For example, the ACKs can occupy 25 bits each, and there can be a 9-byte overhead to the frame. That is, the overhead per ACK on a full size frame is approximately 9/(25/8*48)*100%=6%. The logic can optimize the number of ACKs per frame so an ACK does not need to wait too long to be aggregated when the ACKs are arriving slowly. For example, the ACK aggregation logic block can use three timers to manage ACK transmission based on the activity of an outgoing link. These timers can be started when a new ACK arrives at the ACK aggregation logic block. If the outgoing link is idle, a first timer, which can for example be set at 30 ns, can be used to hold the ACK while waiting for additional ACKs to arrive. When this timer expires, all the ACK received within the corresponding time window can be aggregated into one frame and transmitted onto the outgoing link. If the outgoing link is busy, a second timer, which can for example be set at 60 ns, can be used to wait for additional ACKs. Using this second timer can allow more ACKs to be aggregated into a single frame, and this frame can be transmitted only if a predetermined number of ACKs are collected. Note that due to the Ethernet framing constraints, some numbers of ACKs in a single frame can use less wire bandwidth per ACKs than other numbers of ACKs. If no efficient number of ACKs are collected, and the outgoing link remains busy sending normal data packets, then a third timer, which can for example be set at 90 ns, can be used. Once this third timer expires, all the ACKs that have been collected can be aggregated in a frame and transmitted onto the link. By using these three timers, the system can significantly reduce the overhead of sending ACKs on the outgoing link.
In some examples, the ingress edge port of a switch can encapsulate a received data packet with a fabric header, which allows the packet to be forwarded using flow channels.
When a data packet is delivered to its destination, at least one ACK can be generated.
The ACK type field can indicate different types of ACKs. As mentioned above, during normal operation, when a data packet is delivered to the destination edge port, a regular ACK packet can be generated and sent back to the source. Correspondingly, the ACK type field in the ACK packet can indicate a normal ACK. When congestion occurs, the ACK type field can be used to indicate various types and severity of congestion, such as a new congestion on a flow, persistent congestion on a flow, severe congestion at the egress edge port, or mid-fabric localized congestion that calls for rerouting of the flow to rebalance the load across the entire fabric. In addition, under special circumstances such as the presence of a severely congested fabric link, dropped packets, or link error, an ACK can also be generated by an intermediate switch that is not the final destination, and the ACK type field can be used to notify upstream switches of different types of network condition. Other additional fields can also be included in an ACK packet.
In some examples, flow channel tables can be used to facilitate flow channels throughout a fabric. Flow channel tables are data structures that store the forwarding and state information for a given flow at the port of a switch.
When a packet arrives on an ingress edge link 403 of switch 406, the packet's header can be analyzed by an address translate logic block 410. Address translate logic block 410 can determine the destination fabric address of the egress switch (which in this case is switch 432) based on the packet's Ethernet, IP, or HPC header information. Note that header information associated with other protocols or a combination of different protocols can also be used by address translate logic block 410. The fabric destination address determined by address translate logic block 410 can then be used to perform a lookup in an edge flow channel table (EFCT) 412. EFCT 412 can perform a lookup operation for the packet using the packet's fabric destination address and optionally additional values extracted from the packet's header, which can be referred to as a match pattern. EFCT 412 can compare the packet's match pattern against stored match patterns of all existing allocated flows. If a match is found, then this packet is part of an existing flow and the previously allocated flow ID can be returned for this packet. If no match is found, a new flow ID can be allocated for this packet, and a match pattern can be added to EFCT 412. In other words, EFCT 412 can be used to determine whether a flow channel already exists for the incoming packet, or whether a new flow channel needs to be allocated. In addition to the destination fabric address, other packet header information such as traffic class, TCP or UDP port number, and process or thread ID can be used to map or allocate flow IDs.
The flow ID obtained by EFCT 412 can then be used as an index to map to an entry in an input flow channel table (IFCT) 414. Each entry in IFCT 414 can be indexed by a flow ID and store state information for the corresponding flow. An entry in IFCT 414 can store the values of next_data_flow, data_flow, and ack_flow (see
The flow ID can also be used to identify or allocate a flow-specific input queue in which the incoming packet can be temporarily stored. The state information for a particular queue, as well as parameters for monitoring and controlling the queue (such as threshold for detecting congestion) can be stored in the corresponding entry in IFCT 414. An input queue management logic block can determine when a packet can be dequeued from the input queue and sent to a data crossbar switch 413 based on flow-control parameters stored in the entry of IFCT 414.
When a packet is dequeued from the input queue and sent through crossbar switch 413 to an output port, the packet is sent with the input port number on which it has arrived at switch 406. When the packet reaches an output port's transmission buffer, the packet's header can be updated, based on the packet's flow ID and input port number, with a new flow ID to be used by the next-hop switch (i.e., switch 408) for the same flow. This is because each link, in each direction, can have its own set of flow channels identified by their respective flow IDs. The mapping from the incoming flow ID to the outgoing flow ID used on the next link can be done by looking up an output flow channel table (OFCT) 416. OFCT 416 can perform a lookup using a match pattern that is a combination of the local input port number corresponding to link 403 and the packet's flow ID which is produced by EFCT 412. If a match is found, then the flow has already been defined, and the packet's flow ID is updated with the value corresponding to the match pattern (this new outgoing flow ID is to be used by the downstream next-hop switch 408). If a match is not found, then a new flow channel can be allocated with a new, outgoing flow ID, which can be mapped to the input port number and the previous, incoming flow ID. An entry including the outgoing flow ID, input port number, and incoming flow ID can be stored in OFCT 416.
In the case where the packet is the first packet in the flow, a lookup in OFCT 416 would not produce any mapping. In turn, OFCT 416 can allocate for the packet a flow channel with a flow ID to be used by the input port and IFCT 418 on switch 408. This new flow channel, identified by its flow ID, can be added to the packet header for transmission onto link 417, and can be used by the link partner's (which is switch 408) IFCT 418 to access the flow channel's congestion information. As before, OFCT 424 can further generate a new flow channel if no match is found, using the match pattern of its immediate upstream input port number and flow ID associated with link 417. OFCT 424 can then allocate a new flow channel identified by a new flow ID. Note that OFCT 416 can also function as a forwarding table for ACKs of this flow in the upstream direction. After being forwarded upstream from switch 408 to switch 406, the ACK packet can be updated with the flow ID associated with edge link 403 and forwarded to the appropriate input port on switch 406 as indicated by the corresponding entry in OFCT 416. The ACK packets can be forwarded to the input port by an ACK crossbar switch 415 in the upstream direction.
Subsequently, when the packet arrives at switch 408, its flow ID can be used to identify an input queue to use and to determine an entry in IFCT 418. If the packet's flow ID has not been previously allocated by switch 408, a new input queue can be provided and a new entry in IFCT 418 can be created. From this point onward, a similar process can be performed to forward the packet across switches 408 and 430 until the packet reaches egress switch 432.
When the packet reaches switch 432, after the packet is forwarded by a data crossbar switch 423, an ACK generator logic block 420 can generate an ACK packet based on the packet's flow ID and input port number. This ACK packet can then be forwarded in the upstream direction by an ACK crossbar switch 422. At the same time, based on the ACK packet, an IFCT 421 can update the state information for the flow in the corresponding table entry. When the ACK packet reaches switch 430, an OFCT 419 can be looked up to determine the upstream flow ID and upstream input port to which the ACK packet is to be forwarded. The ACK packet can then have its flow ID updated and be forwarded to the appropriate input port in the upstream direction. As the ACK packet traverses the data path upstream in a similar way, the IFCT at each switch can update its table entry for the flow based on the ACK.
Note that the flow_extent variable can be an important parameter, because it represents the total amount of downstream packet data for a flow. A flow channel is considered free to be reallocated to another flow when the flow_extent of an entry is zero. In general, on receipt of a new packet, the input logic can make a request to send data to an output port. The selected output port can be a function of the flow_extent stored in the IFCT. If flow_extent is zero, there are no packets downstream in the flow to the destination egress edge port. As a result, the switch can use a load based adaptive route selection to choose any valid path that leads to the destination. In a multi-path network, dynamic adaptive routing can be done without the packet being reordered. If flow_extent is not zero, and if in-order delivery is required, the packet can use the same route taken by previous packets. The IFCT can have a field that stores a previous output port number, which is loaded when a packet request is made to an output port and can be used to ensure a connection to the previously used output port.
As mentioned before, the flow channels can use a match function to recognize packets belonging to an existing flow. Received Ethernet frames or other types of packets can be parsed in real time when the frame or packet is received on an ingress edge port and some fields of the packet header can be used for a lookup in a CAM or Ternary Content Addressable Memory (TCAM). If there is a match, the match address can become the flow ID used to select a flow channel. When no match occurs, the switch hardware can load the pattern that fails to match directly onto a free line of the CAM, which can be done without additional delay. As a result, any following packet can be matched to this new entry without significant amount of buffering. The free entry chosen becomes the new flow ID for the new flow channel entry. Note that no external software intervention is required for the loading of the new entry. The process can be completed autonomously by the switch hardware.
The de-allocation of flow IDs and corresponding CAM match lines can also be automatically performed by the hardware when the last ACK is returned for the flow. The de-allocation can occur in hardware with respect to potentially matching new packets, without external software intervention.
In some examples, ingress edge switch 406 can include a fine-grain flow control logic block 434, which can communicate with a network interface controller (NIC) 401 on host 402 to apply flow control on a per-flow basis. More details on find-grain flow control are provided below in conjunction with the description on congestion management.
Exemplary Switch Architecture
In one embodiment, a switch chip supporting the aforementioned features can provide 64 network ports, each of which can operate at 100 Gbps or 200 Gbps, with an aggregate throughput of 12.8 Tbps. Other numbers of ports and data rates are also possible. Each network edge port can support various types of protocols, such as IEEE 802.3 Ethernet, Optimized-IP based protocols, and HPC Portals protocol. Ethernet frames can be bridged based on their layer-2 addresses or be routed based on their layer-3 (IPv4/IPv6) addresses. Optimized-IP frames only have a layer-3 (IPv4/IPv6) header, therefore they are typically routed based on layer-3 addresses. The enhanced Portals format frames typically use specialized NIC and can map directly onto the switch's enhanced fabric format.
When a switch chip is connected to another switch chip, they can communicate using the enhanced fabric frame format, which provides additional control and status fields to support a multi-chip fabric. One of the differentiating features of the present switch architecture when compared with Ethernet switch or alternative technologies such as InfiniBand is that the present switch can provide flow channel based congestion control. The enhanced fabric frame format that operates between switch chips can provide forward and reverse path signaling of the state for flows.
In one embodiment, the switch chip can be implemented based on a crossbar architecture with combined virtual output queuing and crossbar queuing. Buffering and forwarding of data packets can be done with a credit-based request and grant mechanism.
In one embodiment, IEEE 802.3 and Optimized-IP format packets can be passed through ELU block 506. ELU block 506 can extract appropriate headers for use in the lookup process, and return a lookup result to an Ethernet ingress queues (EIQ) block 508, which queues headers for flow channel allocation in EFCT block 510. EIQ block 508 can also associate the addresses of packets stored in the input buffer (IBUF) block 512 with their translated header. For IEEE 802.3 and optimized-IP packets, ELU block 506 can perform a lookup to create fields for forwarding the packets within the fabric.
For an ingress edge port, EIQ block 508 can queue the packet headers waiting for EFCT block 510 to allocate a flow channel. If EFCT block 510 runs out of flow channels, the FIFO queue in EIQ block 508 can fill up, and as configurable thresholds are exceeded, Pause packets can be generated. For packets received from a fabric port, these packets do not need flow channel allocation, and hence their headers are not queued in EIQ block 508.
An input header (IHDR) block 514 which is coupled to IBUF block 512 can perform modifications to a received packet, and update a packet's fabric header fields. IHDR block 514 can receive packet data, ingress time stamps, and grant headers (which can contain modification data and instructions) from IBUF block 512. Such modifications can include removing various Ethernet layer-2 header fields and adding a fabric header. IHDR block 514 can modify packets “on the fly” as packets are read out of IBUF bock 512 and sent to data crossbar 516.
IBUF block 512 can store unmodified packets when they are received by the switch chip, and can support different formats. The stored packet address, which is a pointer denoted as sop_ptr, and the index of the packet can be sent by IBUF block 512 to EIQ block 508, which can match the packet to the ELU block 506's header lookup result.
At some point in time, each packet stored in IBUF block 512 is either sent to a destination port via data crossbar 516 based on a grant sent via a grant crossbar 518 and an input queues (INQ) block 520 (explained below), or is discarded. Both operations can be performed based on a reference to sop_ptr. A grant can also contain other fields from ELU block 506 and EFCT 510, which can be sent with the packet to IHDR block 514. IHDR block 514 in turn can use the control information from the grant header to make appropriate packet modifications before forwarding the packet to the destination port via data crossbar 516. When the buffer in IBUF block 512 fills up, configurable thresholds may be exceeded, which can trigger various flow control and congestion management mechanisms.
EFCT 510 can allocate flow channels to packets, depending on the FTAG, the destination address, and the VNI. The match pattern can provide the separation between flows with separate ordering and priority constraints, between the same source and destination fabric ports. Typically, different cores on a node may be running with different VNIs and this separation of the flows allows the different cores to be decoupled.
If the match value is currently unique, a new flow channel can be allocated. If the match value is identical to the match value of an existing allocated flow channel, then the packet is assigned to the corresponding existing flow. The size of the packet can be used to increment the flow's data_flow value. In one embodiment, for an edge port, an OFCT 522 can be used as the EFCT. The acknowledgments, which are returned from downstream flow channel tables, are used to advance a flow's ack_flow value. When this value catches up with the data_flow value the flow channel can be automatically deallocated and its match pattern can be invalidated.
INQ block 520 can receive the header requests from EIQ block 508 and, for an ingress edge port, from EFCT 510 as well. INQ block 520 can store the lookup result header in its header RAMs. The pointer to each header can be stored in one of a set of queues, based on the header's corresponding flow channel. At edge ports, packet headers can be arbitrated for routing in a manner that is fair by application groups (APPGs), which can be used to group applications into different traffic classes. At fabric ports, headers can be arbitrated based on their flow channels. When a header is considered for routing, it can be forwarded to a fabric routing function (FRF) block 524 and subsequently to an IFCT 526 as well.
FRF block 524 can perform the routing function based on the network topology, and select the output port (or ports for multicast) to which a packet is to be forwarded. This routing result can be passed to IFCT 526, where it is combined with the rest of the header, and IFCT 526 can either use the result from FRF block 524 or choose to use the previous route for a given flow if maintaining packet order is important. IFCT 526 can then pass the forwarding result (i.e., the output port information for a given packet) back to INQ block 520 as a new request. This request can then be used to schedule the packet to traverse data crossbar 516 toward the desired output port.
The request can then be placed in a request queue (or queues) in INQ block 520 based on a shaping function corresponding to the flow channel, a virtual channel (VC) identifier, and the output port. (Note that VCs can be used to separate a physical link into groups of virtual links for the purpose of avoiding deadlocks.) After arbitration, the request can be sent over a request crossbar 528 to an age queues (AGEQ) block 530. Later, a corresponding grant can be returned via grant crossbar 518. When the grant is returned, INQ block 520 can fetch the corresponding header and return it to IBUF block 512, where the header is rejoined with its payload before being forwarded to IHDR block 514 and subsequently data crossbar 516.
As described earlier, IFCT 526 can measure the amount of flow data buffered in the local flow queues. It can also measure the amount of unacknowledged data downstream in the flow. IFCT 526 can also store returned acknowledgment code values in its tables and use this flow-specific state information with configuration information indexed by a packet's FTAG value to determine whether the header of the packets it receives is to be forwarded, discarded, or made to wait longer. The “made to wait” case can be realized by not dequeuing the header from the flow queue. The header can eventually be dequeued, and the decision to forward, discard, or made to wait can be made again. In one embodiment, IFCT 516 can have a “discard” interface to IBUF block 512, which allows the sop_ptr value to be passed to IBUF block 512 when a packet is to be discarded. In response, the packet's header can be dropped before being added to a request queue. IFCT 516 can further increment the corresponding statistics for dropped packets.
FRF block 524 can receive routing requests from INQ block 520 for each received packet, and for each routing request can return a routing response to IFCT 526. The routing response can indicate which port or ports the packet should be forwarded to and the VC on which it should be forwarded. For non-multicast requests, the response can indicate both a preferred port and a set of acceptable ports to which the packet can be forwarded, thereby allowing IFCT 526 to use the preferred port for a new flow or a rerouted flow, or, for an existing flow, to maintain the current path via a port that may not be FRF block 524's current preferred choice. In the presence of errors, the FRF may also indicate to the IFCT that there is no legal port the packet can be forwarded to. When this occurs, the packet is discarded.
FRF block 524's routing decisions can be based on a combination of software-configurable table-based rules, dynamic load information, and pseudo-random selection. Rules can take into account factors including the packet's destination, where it is along its path (such as source group, intermediate group, destination group, destination switch), the VC it is received on, and the type of port (edge, local, or global) at which it is received. AGEQ block 530 can provide FRF block 524 with the current load present at the output side of the port associated with a given FRF instance. Each FRF instance can communicate with every other FRF instance within the switch chip to learn the current load present at each output port, and the link up/down status of each port. FRF instances can also communicate with FRF instances in neighboring switch chips to obtain load-related status of the neighboring devices. In one embodiment, FRF block 524 can be configured to support multiple network topologies.
AGEQ block 530 can accept requests from all of the input ports via request crossbar 528, buffer them, arbitrate between them by traffic class using a traffic shaper, and pass them to OFCT block 522 to be granted grant crossbar 518. The buffering of request within AGEQ block 530 can be managed to allow each input to have sufficient space to send requests while also allowing an input with multiple flows targeting a given output to take more space. AGEQ block 530 can also be responsible for managing access to the link either using credit based flow control for the IBUF block of a neighbor switch chip or pause-based flow control for non-fabric links. When a packet is released by AGEQ block 530 (that is, a corresponding grant is issued for the packet that is waiting in IBUF block 512), the packet is to be placed on the outgoing link. Additionally, AGEQ block 530 can have a path allowing for packets initiated on a given port (such maintenance or reduction packets) to arbitrate for resources on the port.
OFCT 522 can be programmed to operate either as an EFCT for an egress edge port or an OFCT for a fabric port. For egress edge port operation, when the block is programmed as EFCT, the headers received from AGEQ block 530 can be passed through the EFCT to grant crossbar 518, largely unchanged. The EFCT can also receive new ACKs from an output buffer (OBUF) block 532 to acknowledge packets leaving the fabric. These ACKs can be returned to ACK crossbar 534 and will be the ACKs that close flows in the upstream flow tables. The EFCT can also generate congestion notification messages when AGEQ block 530 is reporting congestion. This congestion, on an egress edge port, usually represents an incast forming and is used to slow down the flow back at the ingress edge port.
For fabric port operation, OFCT 522 can manage the allocation of the flow channels for the next hop switch over an outgoing link. It can work in tandem with the fabric link partner's IFCT and create extension to the flows that the link partner's IFCT can use to manage the forward progress of the packets.
OFCT 522 can also manage the ACKs received from the fabric link and return these ACKs upstream through ACK crossbar 534. Having created or extended an existing flow, OFCT 522 can generate the flow_ID and data_flow values which can be added to the next hop fabric header by IHDR block 514 and add these values to the grant that is returned with other header values back to grant crossbar 518.
The output buffer (OBUF) block 532 can capture packets that have been sent over data crossbar 516 to the corresponding output port. The packets can arrive on, for example, four different column buses and are enqueued on, for example, four separate FIFO queues (explained in more detail in conjunction with
OBUF block 532 can also generate ACK values to indicate mid-fabric congestion if AGEQ block 530 starts to fill or discard ACKs if AGEQ block 530 has discarded a packet.
In one embodiment, a control packet receiver (CFRX) block 538 can process all the control-related packets which can be extracted from IBUF block 512. These control-related packets can include congestion signaling packets, flow control credit issuing packets, and flow channel ACKs, among others. The congestion signaling information can be sent to FRF block 524 and used for making routing decisions. The credit-based flow control information can be sent to AGEQ block 530 for scheduling packet forwarding to the downstream switch. ACKs can be sent to OFCT 522, which in turn can identify the input port the ACK is to be forwarded to, and subsequently sent to ACK crossbar 534.
Correspondingly CFTX block 536 can send the ACKs (based on IFCT 526), credit-based flow control packets (based on the state of IBUF block 512), and congestion signaling packets to the corresponding output port.
OBUF block 532 can also generate a credit feedback, which indicates the landing space available for outgoing data packets, to AGEQ block 530 (note that this credit is used for crossbar scheduling between inputs and outputs of the crossbar, and is different from credits used for inter-switch flow control). This credit information is passed by AGEQ block 530, optionally via a credit crossbar 540, to INQ block 520, which uses this credit information to schedule packet extraction from IBUF block 512.
As mentioned above, there can be five crossbars in a switch chip: request crossbar 528, grant crossbar 518, credit crossbar 540, ACK crossbar 534, and data crossbar 516.
Request crossbar 528 can send requests from an input to the targeted output AGEQ block. A credit protocol can be used to guarantee that there is a landing space for a request at the output. Each request can contain a pointer (sop_ptr) to where the packet is stored in IBUF block 512.
Grant crossbar 518 can return a grant back to the input which satisfies a request. The grant can return the pointer (sop_ptr). A grant is only returned when there is space in OBUF block 532 for the corresponding packet. Grants can also optionally return credit for request space in OBUF block 532.
Credit crossbar 540 can return credit for request space in OBUF block 532. ACK crossbar 534 can propagate ACK packets from output ports to input ports, based on OFCT 522. Data crossbar 516 can move granted packets from IBUF block 512 to the targeted OBUF block 532. Grants are only returned when there is a guaranteed landing space for the packet at the output so packets cannot be blocked.
Now referring back to
In one embodiment, row buses and column buses can both use a credit based protocol to determine when they are able to send (see arbiters 576 and 578 in
For column buses, each crossbar tile can maintain credit counts for the input buffers at each destination port in that column. Unlike row buses, there is no requirement that credits be available for a largest-sized packet before starting transfer of that packet on a column bus. Individual words of the packet can be moved as credits become available. Therefore, the input buffer at the destination for each column bus needs to be only large enough to cover the worst-case round trip (e.g., from packet-send to credit-return).
As shown in
In one embodiment, an output control block can be responsible for accepting requests from all of the input ports via the request crossbar, buffering them, and passing them to the OFCT to be granted via the grant crossbar. The AGEQ space can be managed by the output control block to allow a single input with multiple flows targeting a given output to move its requests to the AGEQ. The output control block can also be responsible for managing the use of space in the input buffer on a downstream neighboring switch (i.e., the link partner corresponding to an output port) and allocation of flow channels. Additionally, the output control block can have a path allowing for packets initiated on a given port, such as maintenance or reduction packets, to be arbitrated for resources on the port.
Requests can come into the output control block via a column bus from each row of the matrix. Each column bus can feed an independent FIFO queue with space in the FIFO queue managed via credits. These FIFOs can be sized sufficiently deep to cover the maximum round-trip delay plus some extra space to allow requests to be moved out of the crossbars and to prevent head-of-line blocking. Prior to being written into a FIFO queue the request can be checked for valid error check code (ECC). If an error is detected, the packet can be discarded with an error flagged.
In one embodiment, least recently used (LRU) arbitration can be used between the column bus FIFO queues to choose which FIFO queue is selected with the corresponding request forwarded to the AGEQ block. As requests are removed from each FIFO queue, credits can be returned to the corresponding crossbar.
The output buffer can make requests to the output control block for sending reduction and maintenance packets across the corresponding outgoing link. These requests can be given a higher priority. In one embodiment, reduction packets do not use flow channels, and maintenance packets can use loopback to create a flow, so there is no need to either check for flow channel availability or to use the OFCT to create a grant. They also do not use space in the output buffer so check of space is not required.
The size of the next request to be processed from the output buffer can be checked against the maximum packet size. If it exceeds this value, the request is not processed and an error flag can be set. This can result in the output buffer request path being blocked until a warm reset is performed.
In one embodiment, each input port can be allocated a fixed amount, denoted as fixed_alloc, of AGEQ space. This space can be sufficiently large to accommodate each traffic class associated with a respective input port, with sufficient additional space to cover the request-credit round trip. The allocation of this fixed space among different traffic classes within the same input port can be configurable. A traffic class can be identified by a combination of the shaping queue (SQ) identifier and virtual channel (VC) identifier. In one embodiment, the AGEQ can have 8 k locations, each location corresponding to a unit of traffic. The total amount of fixed allocated space can be (64*fixed_alloc), and the remaining space can be 8 k-64*fixed_alloc. This remaining space can be shared among all inputs.
The shared space can be managed by the output. Incoming requests can be moved from static to shared space as they arrive if there is room in the shared space, subject to per-input limits. When moving a request to the shared space a credit can be returned immediately via the credit crossbar with the request marked in the AGEQ as being in the shared space. When the request is granted, if it is marked as using the shared space, the shared space is credited. If it is not marked as using shared space, it is considered to have used the static space and a credit is returned to the input with the grant.
Due to conflicts in the credit crossbar, it is possible that credits may not be able to be sent every clock period. A FIFO queue can be used to provide buffering for these transient disruptions. In one embodiment, a request can be accepted from the request crossbar only if there is space in this FIFO queue. A FIFO queue with a depth of 32 locations, for example, can be used to limit the possibility of it backing up into the request crossbar.
The shared space in AGEQ can impose limits for how much space any single input can occupy. These limits can be set as a percentage of the available space. For instance, if the limit is set to 50%, and if one input is active, the input can have access to 50% of the shared space. With two active inputs, each input can have access to 37.5% of the share space, which is calculated as (space_used_by_1+space_left*0.5)/2=(50%+50%*0.5)/2=37.5%. With three active inputs, each input can have access to 29.2% of the shared space, which is calculated as (space_used_by_2+space_left*0.5)/3=(75%+25%*0.5)/3=29.2%, and so on. The total shared space that can be used by the all the active inputs is limited to the total, which are 50%, 75%, and 87.5% in these three examples, respectively. With this configuration, the shared space allocated to each input can vary dynamically based on how many inputs are currently active. The addition of an active input can cause other active inputs to give up their shared space, which is then allocated to the new input.
Given that division can be costly to implement in hardware, this dynamic allocation function of the shared AGEQ space can be implemented as a lookup table with, for example, 64 entries, where each entry corresponds to a number of active input ports. The number of active input ports can be used as an index to the table. The values in the table can be the limit of the shared space any input can access, along with the total space they can consume as a whole. A software-based function can be used to program the values in the table according to how much total shared space there is and what percentage each input is allowed to use. As more inputs become active, each input is allowed less space and the total space available increases. Incoming requests from inputs that are above this limit, or in total above the total space limit, may not be allowed to take more shared space.
In order to track the number of active inputs in the AGEQ, a set of 64 counters (one for each input) can be used. These counters can count up when a request is placed in AGEQ and count down as they are taken out (granted). A second counter that counts of the number of non-zero counts can be used to index into the lookup table for shared space allocation. In addition, in order to manage the shared space, an additional set of 64 counters can be used to track the current usage of the shared space by each input. There can also be a single counter which tracks overall shared space usage. These counters can be compared against the current quotas to determine whether a request is allowed to use the shared space. In one embodiment, all counters can be 13-bits wide which is sufficient to cover, for example, the 8K locations in AGEQ.
As shown in
In addition to the linked-list data structure, each queue can also have a FIFO queue such as FIFO queue 584 of requests at its head. These FIFO queues can be used to ensure that a queue can sustain a request every clock with a mutli-clock read access time from the request RAM. When a new request arrives, if the head FIFO queue for that queue is not full, the request can bypass the request RAM and be written directly into the head FIFO queue. Once requests for a given queue are being written to the request RAM, subsequent requests are also written to the request RAM to maintain order. The bypass path can be used again once there are no more requests for that queue in the request RAM and there is room in the corresponding head FIFO.
When a request is read from a head FIFO queue, and there are corresponding requests queued in the request RAM, a dequeue operation can be initiated. Since only one head FIFO queue is read at a time, only a single dequeue operation can be initiated each clock period. Logic can be included to handle the various race conditions between an ongoing or imminent enqueue operation and a head FIFO queue being read.
The Free List RAM can be a simple FIFO queue which is initialized with pointers to all entries (e.g., 8 k entries) whenever a reset is done. A count can be maintained to keep track of how many entries are valid within the Free List RAM. As entries are taken, they are popped off the front of the FI FO and used. As entries are returned, they are pushed onto the back of a FIFO queue 585. A number of entries (for example, 3) at the head of the Free List RAM can be kept in flops so they are available for quick access.
In order to support full performance for small packets the age queues need to support both an enqueue operation and a dequeue operation every clock period. The operations on the data structures for an enqueue operation are given below. They differ depending on whether the queue being written is empty or not. In most cases a simultaneous enqueue and dequeue to a specific queue can be handled as they are using and updating separate fields. A special case would be the one where the dequeue operation empties the queue. In order to handle this case, the dequeue operation can take place first logically followed by the enqueue operation. This can be made possible by using an empty flag for the queue, which can be set when the queue is emptied by the dequeue operation and then cleared based on the enqueue operation.
Arbitration can be done among requests that are permitted to be granted subject to input buffer management, output buffer management, and flow channel quotas. Arbitration can also be stopped if there are no credits for the OFCT input FIFO queue. In one embodiment, arbitration can be done in two levels, one for among the SQs and one for among the VCs. A traffic shaping arbitration can be used to arbitrate between the SQs. A Deficit Round-robin arbitration can be used to arbitrate between VCs within a given SQ.
In one embodiment, the traffic shaping arbitration can use a series of token buckets to control the bandwidth of each SQ. For example, if there are 8 SQs, there can be 8 leaf buckets (one for each SQ), 4 branch buckets, and one head bucket. Arbitration can be divided into three groups with group 1 having the highest priority, followed by group 2 and then group 3. For group 1 and 2, the arbitration can be done in the same way among eligible SQs. A ×8 Round-robin arbitration can be done between the SQs for each of the 8 priority levels (8 parallel Round-robin arbitration operations). A fixed arbitration can be done between the priority levels. The group 3 arbitration has no priorities and can be a single ×8 Round-robin arbitration. For group 1 arbitration, the priority for each comes from the setting in the leaf buckets. For group 2 arbitration, the priority comes from the setting in the branch buckets. In all cases, the buckets which are checked to be eligible for that group are also the buckets from which packet size tokens are taken if that request wins arbitration.
Packets can be classified in order to select the SQ to which their request is forwarded. This allows traffic associated with an application to be shaped differently to traffic from a different application or a different traffic class. This feature is especially important on the edge ports which connect to a NIC, because the applications are typically configured to use a share of the resources on the node, and similarly they are granted a proportion of the network bandwidth. In one embodiment, this classification can be performed by labeling the packets with an FTAG and a VNI as the packet enters the fabric. The FTAG and VNI are then used as the packet leaves the fabric to select the shaping queue. A configuration register can be used to map FTAGs to SQs.
In one embodiment, the AGEQ can have a number of shaping queues that are addressed by {SQ, VC}. For example, if there are 8 SQs and 4 VCs, there can be a total of 32 individual shaping queues. The corresponding 3-bit SQ index can be a shaping function, and the VC value maps to one of four queues (corresponding to the 4 VCs) within that shaping function. For Ethernet egress (edge) ports, the VC is not needed for deadlock avoidance, therefore all 32 shaping queues are available.
Congestion Management
As described above, each flow at a given switch can have its own private queue of packets. This configuration facilitates separate flow control for each flow. As a result, the network can remain mostly lossless, and one flow using a link can be blocked without blocking any of the other flows using the same link. Unlike a traditional packet switched network, congestion in one part of the network can only affect the flows that are contributing to the congestion. For example, in a conventional network, the buffers before a congested link can quickly fill up with the packets causing the congestion. This in turn can force the switch to issue a pause command or use some other flow control method to prevent neighboring switches from sending packets toward the congested link. Consequently, the packets causing congestion can be stopped or slowed down, and all other packets, which may not be heading to the congested link, can also be stopped or slowed down. As a result, the congestion could spread sideways and increase the size of the saturation tree from a topological perspective.
In contrast, with flow channels, the load corresponding to flows contributing to congestion can be reduced on the links leading up to the congestion. This reduction of load can allow other flows that are sharing these links to use more link bandwidth and deliver their payload more quickly, while only the packets contributing to the congested link are slowed down.
Typically, conventional networks can operate normally provided the network load is not at or near full capacity. This can be the case for small or medium sized networks most of the time. With large or very large networks operating with multiple bandwidth-hungry applications, however, at any point in time part of the network can be saturated with traffic load. Under these circumstances, unfair packet delivery could occur even if individual switches implement locally fair policies.
The next nearest switch, which is switch 606, can do the same and so on. In this example, with only four stages of switches and only three or four inputs on each stage, and only with a total of 11 inputs trying to send to the destination node L, three input sources (nodes A, B, and C) only take 1/48 the bandwidth taken by two other input sources (nodes J and K) on the outgoing link toward destination node L. Hence, even with locally fair arbitration policies, nodes that are far away from the destination can suffer from very unfair treatment. A more realistic network topology can involve more switching stages, greater numbers of switch inputs, and more sources trying to send to a single destination. A moderate-sized incast could result in six orders of magnitude difference between the delivered bandwidths among different sources.
The unfairness problem described above is often caused by the fact that the arbitration policies implemented by a switch are based on the input ports. That is, the bandwidth throttling is done with a per-port granularity. In contrast, by facilitating flow channels and implementing flow-specific throttling, a network can significantly reduce the amount of unfairness among different flows. For example, in the scenario shown in
In addition to fairness, another challenge faced by network architects is congestion. In general, two types of congestions can occur in a network. The first type is endpoint congestion, where an egress edge link coupled to a destination device is congested. The second type is fabric link congestion, where an intermediate fabric link is congested.
A switch can detect and mitigate endpoint congestion by monitoring the output buffer on an egress edge link and by sending ACKs with congestion information to upstream switches and source nodes. More specifically, the output buffer coupled to an egress edge link can monitor the state of the buffer and detect congestion when certain criteria are met. When a packet arrives at or leaves an output buffer, the output buffer can compute three congestion-detection parameters, such as: (1) the amount of data stored in the buffer, (2) the number of packets stored in the buffer, and (3) the rate of change of buffer depth (amount of data stored in the buffer). Three threshold values can be set respectively for these three monitored parameters, although more or less can be set. Congestion is considered to be present when at least one of these parameters exceeds the corresponding threshold.
When congestion is detected, the switch can generate and transmit an endpoint-congestion-notification ACK corresponding to the packet that has just entered the output buffer. The ACK can include a value indicating the severity of the congestion. Note that this endpoint-congestion-notification ACK is not intended to notify upstream switches of the successful delivery of the packet, but to inform them of the presence and degree of congestion at the egress edge link. (In fact when this endpoint-congestion-notification ACK is sent, the packet may still be stored in the output buffer waiting to be transmitted onto the egress edge link.) This fast, explicit congestion notification mechanism allows the switches to act quickly on a specific flow contributing to the congestion.
In addition, the output buffer can update the congestion-detection parameters when a packet is dequeued and transmitted onto the egress edge link. If no congestion is present, a regular ACK is generated and sent, which can clear any previous congestion notifications received by the upstream switches operating on the corresponding flow. If congestion is present, the ACK can be marked with a flag, which allows the ACK to notify the switches of persistent congestion at the egress edge link as well as the successful delivery of the packet.
Note that the endpoint congestion management logic block shown in
As an endpoint-congestion-notification ACK traverses the fabric, the IFCT's of the switches along the path can apply bandwidth limitations to the flow corresponding to the ACK. Effectively, the fabric can slow down the delivery of that flow in a distributed way at each switch along the data path. When an endpoint-congestion-notification ACK passes an IFCT its value can be stored in the flow's table entry as an ep_congestion value, which can be used to select a desired maximum bandwidth for the flow. Each value of ep_congestion can have a corresponding set of high, target, and drop watermark values. For high levels of congestion, when ep_congestion has a high value, the watermark values can have lower values, so that the congestion can be mitigated more aggressively. For low levels of congestion, when ep_congestion has a low value, a different set of greater high, target, and drop watermark values can be used for higher flow bandwidth. For example, a table indexed by the ep_congestion value can be used. For each ep_congestion value, the table can indicate a corresponding set of high, target, and drop watermark values. The entries of this table can be predetermined, so that when an endpoint-congestion-notification ACK is received, the switch can use the ep_congestion value to perform a lookup in this table, and apply the three corresponding watermark values to the identified flow.
In some cases, if the source is injecting data in a greedy manner, only slowing down the forwarding inside the network might not be sufficient to fully remove the congestion. To address this problem, an ingress edge switch can be configured to instruct the source device (which typically resides outside the fabric) to limit data injection on a fine-grain, per-flow basis. This switch-to-host flow control mechanism can be referred to as Fine Gran Flow Control (FGFC).
In particular, especially in an HPC environment, an end host or computing node could have a large number of cores running numerous threads, processes, or virtual machines, each of which could be injecting their own stream of data into the network through a common physical network interface controller (NIC). When congestion is present, a per-port based flow control can only throttle the overall data rate over a single port on the NIC, which can be 40 Gb/s or more. Pushing back on the total data rate on the entire port can cause unfairness to flows that are not contributing to congestion. FGFC can extend the concept of the individual flows or group of associated flows to their ultimate source, which can be a single thread executing on one of the cores.
To slow down data injection from the source, an FGFC logic block on an ingress edge switch (for example, FGFC logic block 434 in edge switch 406 in
FGFC can implement the control communication between an edge switch and an end host NIC using an Ethernet frame with an Organizationally Unique Identifier (OUI) extended Ether_Type field. These frames can indicate one or more of the following: (1) the protocol used by the flow being controlled; (2) an identifier to indicate the source (e.g., application, process, or thread) generating the packets that need to be throttled; (3) a pause time value for which the flow control is to last (which can prevent a lockup if subsequent FGFC frames are lost due to errors), and (4) a credit value, which can be zero, to indicate the number of frames or amount of data that can be sent during the pause period.
Note that the identifier for indicating the source flow subject to flow control can be different based on the protocol associated with the flow. For layer-2 Ethernet virtual local area network (VLAN) traffic, the identifier can include the VLAN number. For IPv4 traffic, the identifier can include a source/destination IP address pair, a UDP or TCP/IP 5-tuple that includes UDP or TCP port numbers, or an optional flow label. For IPv6 traffic, the identifier can include one or more IPv6 addresses or an IPv6 flow label. For proprietary HPC protocol traffic, the identifier can include a process or thread ID. In general, this identifier is also stored in the EFCT of the edge switch, since it is used to map the corresponding traffic to a flow ID.
To trigger FGFC, the IFCT of an ingress edge switch can monitor its flow-specific input queues. For each queue, the corresponding IFCT entry can indicate three watermark values: high, target, and drop, which can be used to measure the queue depth. In some examples, these watermark values can be included as additional fields in the IFCT as shown in
To facilitate FGFC, a NIC can be configured to process the FGFC Ethernet frame, so that the NIC can communicate to the application or process on an end host that is generating the data. Parsing of the FGFC Ethernet frame and communication to the application or process can be done in software, hardware, or a combination of both.
As mentioned above, two types of congestion can occur in a network. A first type is endpoint congestion, and a second type is fabric link congestion.
To mitigate fabric link congestion, a switch can apply dynamic per-flow credit-based flow control. At a switch, if an input queue starts to fill up, and the queue_extent value for this flow reaches a predetermined threshold, the switch can generate a special ACK to notify the upstream switch's IFCT of the congestion. This special per-hop ACK can be referred to as “HeadroomACK.” Upon receiving the HeadroomACK, the upstream switch's IFCT can start a credit based flow control with the downstream switch. In the downstream IFCT entry, a flag Upstream Metering (UM) can be set to indicate that the data transmission from the upstream switch is now metered based on the credits. The HeadroomACK packet can also include a credit value.
When the upstream switch receives a HeadroomACK, a flag called Downstream Metered (DM) can be set in the corresponding entry of the IFCT. The IFCT can also store a signed headroom field in the IFCT entry with the credit value carried by the HeadroomACK (i.e., the headroom value indicates the number of credits). This headroom field can represent the amount of data that can be forwarded to the downstream switch. This establishes a credit based flow control for the corresponding flow. If the upstream IFCT receives a HeadroomACK while the DM flag in the flow's entry is already set, the credit value carried by the HeadroomACK can be added to the existing headroom value.
New packets received by the upstream IFCT can be blocked if the headroom value is not greater than zero (i.e., there is no credit available). These packets can fill this flow's input queue and may in turn cause the IFCT to initiate per-flow credit based flow control with its upstream IFCT, and so on. If the headroom value is greater than zero, a packet stored in the input queue can be dequeued and forwarded to the downstream switch, and the headroom value can be decremented by the size of the forwarded packet, which may cause the headroom value to become zero or negative.
With the flow restricted from sending new packets to the downstream IFCT, the downstream IFCT's input queue can start to drain at some rate depending on its downstream congestion. As described above, each flow's input queue can have three queue-depth watermark values, namely high, target, and drop, which can be used to manage credit-based flow control. The target watermark can be approximately the ideal queue depth for the desired flow bandwidth. It indicates sufficient buffering is available for transmitting data downstream. When there is congestion, the credit-based flow control mechanism can attempt to keep the flow's queue_extent value approximately at this target watermark.
If the queue_extent value is between the high watermark and drop watermark, and is greater than the target watermark, when a packet is forwarded, slightly less than this packet's size of credit can be returned with a HeadroomACK to the upstream switch. If the queue_extent value does not exceed the target watermark, when a packet is forwarded, slightly more than this packet's size of credit can be returned with the HeadroomACK to the upstream switch.
If the queue_extent depth is greater than the high watermark, no credit is returned when packets are forwarded. This mechanism can bring the queue_extent value down more quickly and is usually used when congestion is detected for the first time. If the congestion clears, the flow's input queue can start to empty more quickly. When the queue depth is less than the drop watermark, the credit-based flow control can be switched off. This can done by clearing the UM flag in the IFCT entry and returning a HeadroomACK with the maximum credit value to the upstream switch. When received by the upstream IFCT the HeadroomACK clears the entry's DM flag and flow control against the headroom value is turned off.
Note that in a typical network topology there can be a number of switches and between two endpoints there can be multiple data paths. In a multi-path network, it is possible to use various methods to control fabric link congestion. For example, the injection limits, described later in this document, can control the maximum total amount of data in the entire fabric. This means that if a particular fabric link is overloaded, a flow can use a different data path that does not go through the congested link. It is possible to detect an overloaded link and generate “reroute” ACKs for a set of flows. The reroute ACKs can temporarily block the flow in an upstream switch, and when all the ACKs for that flow have been returned, the flow can be unblocked and become free to use a different path across the fabric. A dynamic load-based adaptive routing mechanism can then direct the lead packet to use a different uncongested fabric link. In turn the load across the entire fabric can become more balanced.
In general, a flow channel switch can use a combination of several congestion detection and control mechanisms. For example, different degrees of endpoint congestion can be reported using the endpoint-congestion-notification ACK that can be returned from the final fabric egress edge port. This ACK type can be used to manage the bandwidth of flows into a significantly congested egress edge port. The system can also use a per-hop credit-based flow control to manage fabric link congestion. This per-hop congestion management mechanism can be effective against low to moderate levels of congestion, because the response time can be much shorter than the network-wise round trip delay.
If the congestion is severe, perhaps caused by a wide incast, the system can also apply a per-flow injection limit. A flow's injection limit can be determined based on the ep_congestion value. The injection limit can be compared with the flow_extent value in all IFCTs the flow passes through. If the flow_extent is greater than this limit the IFCT can block the forwarding of packets from the input queue for this flow. This mechanism can reduce the rate of forwarding of packets over an entire flow to as little as a single packet.
The system can also protect unrelated traffic from extreme congestion caused by incasts with a large number of contributors. In this case, the ep_congestion value can be set to a high value and the average amount of data of a flow can be reduced to a small fraction of a packet. This can be achieved by only releasing the next packet of an individual flow into the fabric from the ingress edge port's IFCT after a programmable delay has elapsed since when the ACK of the previous packet has been received.
In addition to per-flow injection limits, the system can measure the amount of data that has been injected into the fabric on a per-ingress-port basis, and set injection limits to impose a cap on the total amount of data a port can inject into the fabric. Since every ingress port can apply this injection limit, the system can control the maximum amount of data allowed inside the fabric. Limiting the total amount of data into the fabric can ensure that buffer exhaustion does not occur where bandwidth is scarce. As a result, traffic which is not using the paths with reduced bandwidth are not affected.
To facilitate per-port injection limit, an IFCT can maintain a total traffic count. Each time a packet is injected into the fabric from the edge port the total count can be incremented. When a flow's ACK is returned, the total traffic count can be decremented. Once all the ACKs of all the flows of an ingress port have been returned (i.e., when the sum of the flow_extent values for all the flows becomes zero), the total traffic count can be set to zero.
Crossbar switch 1210 can include one or more crossbar switch chips, which can be configured to forward data packets and control packets (such as ACK packets) among the communication ports. EFCT logic block 1212 can process packets received from an edge link and map the received packets to respective flows based on one or more header fields in the packets. In addition, EFCT logic block 1212 can assemble FGFC Ethernet frames, which can be communicated to an end host to control the amount of data injected by individual processes or threads. IFCT logic block 1214 can include the IFCT, and perform various flow control methods in response to control packets, such as endpoint-congestion-notification ACKs and fabric-link credit-based flow control ACKs. OFCT logic block 1216 can include a memory unit that stores the OFCT and communicate with another switch's IFCT logic block to update a packet's flow ID when the packet is forwarded to a next-hop switch.
Crossbar switch 1310 can include one or more crossbar switch chips, which can be configured to forward data packets and control packets (such as ACK packets) among the communication ports. EFCT logic block 1312 can process packets received from an edge link and map the received packets to respective flows based on one or more header fields in the packets. In addition, EFCT logic block 1312 can assemble FGFC Ethernet frames, which can be communicated to an end host to control the amount of data injected by individual processes or threads. IFCT logic block 1314 can include the IFCT, and perform various flow control methods in response to control packets, such as endpoint-congestion-notification ACKs and fabric-link credit-based flow control ACKs. OFCT logic block 1316 can include a memory unit that stores the OFCT and communicate with another switch's IFCT logic block to update a packet's flow ID when the packet is forwarded to a next-hop switch.
In summary, the present disclosure describes a data-driven intelligent networking system that can accommodate dynamic traffic with fast, effective congestion control. The system can maintain state information of individual packet flows, which can be set up or released dynamically based on injected data. A packet flow can be mapped to their layer-2, layer-3, or other protocol-specific header information. Each flow can be provided with a flow-specific input queue upon arriving at a switch. Packets of a respective flow are acknowledged after reaching the egress point of the network, and the acknowledgement packets are sent back to the ingress point of the flow along the same data path in the reverse direction. As a result, each switch can obtain state information of each flow and perform flow control of a per-flow basis. Such flow control allows the network to be better utilized while providing versatile traffic-engineering and congestion control capabilities.
The methods and processes described above can be performed by hardware logic blocks, modules, or apparatus. The hardware logic blocks, modules, or apparatus can include, but are not limited to, application-specific integrated circuit (ASIC) chips, field-programmable gate arrays (FPGAs), dedicated or shared processors that execute a piece of code at a particular time, and other programmable-logic devices now known or later developed. When the hardware logic blocks, modules, or apparatus are activated, they perform the methods and processes included within them.
The methods and processes described herein can also be embodied as code or data, which can be stored in a storage device or computer-readable storage medium. When a processor reads and executes the stored code or data, the processor can perform these methods and processes.
The foregoing descriptions of embodiments of the present invention have been presented for purposes of illustration and description only. They are not intended to be exhaustive or to limit the present invention to the forms disclosed. Accordingly, many modifications and variations will be apparent to practitioners skilled in the art. Additionally, the above disclosure is not intended to limit the present invention. The scope of the present invention is defined by the appended claims.
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/US2020/024267 | 3/23/2020 | WO |
| Publishing Document | Publishing Date | Country | Kind |
|---|---|---|---|
| WO2020/236286 | 11/26/2020 | WO | A |
| Number | Name | Date | Kind |
|---|---|---|---|
| 4807118 | Lin et al. | Feb 1989 | A |
| 5138615 | Lamport et al. | Aug 1992 | A |
| 5457687 | Newman | Oct 1995 | A |
| 5937436 | Watkins | Aug 1999 | A |
| 5960178 | Cochinwala et al. | Sep 1999 | A |
| 5970232 | Passint et al. | Oct 1999 | A |
| 5983332 | Watkins | Nov 1999 | A |
| 6112265 | Harriman et al. | Aug 2000 | A |
| 6230252 | Passint et al. | May 2001 | B1 |
| 6246682 | Roy et al. | Jun 2001 | B1 |
| 6493347 | Sindhu et al. | Dec 2002 | B2 |
| 6545981 | Garcia et al. | Apr 2003 | B1 |
| 6633580 | Toerudbakken et al. | Oct 2003 | B1 |
| 6674720 | Passint et al. | Jan 2004 | B1 |
| 6714553 | Poole et al. | Mar 2004 | B1 |
| 6728211 | Peris et al. | Apr 2004 | B1 |
| 6732212 | Sugahara et al. | May 2004 | B2 |
| 6735173 | Lenoski et al. | May 2004 | B1 |
| 6894974 | Aweva et al. | May 2005 | B1 |
| 7023856 | Washabaugh et al. | Apr 2006 | B1 |
| 7110360 | Hui | Sep 2006 | B1 |
| 7133940 | Blightman et al. | Nov 2006 | B2 |
| 7218637 | Best | May 2007 | B1 |
| 7269180 | Bly et al. | Sep 2007 | B2 |
| 7305487 | Blumrich et al. | Dec 2007 | B2 |
| 7337285 | Tanoue | Feb 2008 | B2 |
| 7397797 | Alfieri et al. | Jul 2008 | B2 |
| 7430559 | Lomet | Sep 2008 | B2 |
| 7441006 | Biran et al. | Oct 2008 | B2 |
| 7464174 | Ngai | Dec 2008 | B1 |
| 7483442 | Torudbakken et al. | Jan 2009 | B1 |
| 7562366 | Pope et al. | Jul 2009 | B2 |
| 7593329 | Kwan et al. | Sep 2009 | B2 |
| 7596628 | Aloni et al. | Sep 2009 | B2 |
| 7620791 | Wentzlaff et al. | Nov 2009 | B1 |
| 7633869 | Morris et al. | Dec 2009 | B1 |
| 7639616 | Manula et al. | Dec 2009 | B1 |
| 7734894 | Wentzlaff et al. | Jun 2010 | B1 |
| 7774461 | Tanaka et al. | Aug 2010 | B2 |
| 7782869 | Chitlur Srinivasa | Aug 2010 | B1 |
| 7796579 | Bruss | Sep 2010 | B2 |
| 7856026 | Finan et al. | Dec 2010 | B1 |
| 7933282 | Gupta et al. | Apr 2011 | B1 |
| 7953002 | Opsasnick | May 2011 | B2 |
| 7975120 | Sabbatini, Jr. et al. | Jul 2011 | B2 |
| 8014278 | Subramanian et al. | Sep 2011 | B1 |
| 8023521 | Woo et al. | Sep 2011 | B2 |
| 8050180 | Judd | Nov 2011 | B2 |
| 8077606 | Litwack | Dec 2011 | B1 |
| 8103788 | Miranda | Jan 2012 | B1 |
| 8160085 | Voruganti et al. | Apr 2012 | B2 |
| 8175107 | Yalagandula et al. | May 2012 | B1 |
| 8249072 | Sugumar et al. | Aug 2012 | B2 |
| 8281013 | Mundkur et al. | Oct 2012 | B2 |
| 8352727 | Chen et al. | Jan 2013 | B2 |
| 8353003 | Noehring et al. | Jan 2013 | B2 |
| 8443151 | Tang et al. | May 2013 | B2 |
| 8473783 | Andrade et al. | Jun 2013 | B2 |
| 8543534 | Alves et al. | Sep 2013 | B2 |
| 8619793 | Lavian et al. | Dec 2013 | B2 |
| 8626957 | Blumrich et al. | Jan 2014 | B2 |
| 8650582 | Archer et al. | Feb 2014 | B2 |
| 8706832 | Blocksome | Apr 2014 | B2 |
| 8719543 | Kaminski et al. | May 2014 | B2 |
| 8811183 | Anand et al. | Aug 2014 | B1 |
| 8948175 | Bly et al. | Feb 2015 | B2 |
| 8971345 | McCanne et al. | Mar 2015 | B1 |
| 9001663 | Attar et al. | Apr 2015 | B2 |
| 9053012 | Northcott et al. | Jun 2015 | B1 |
| 9088496 | Vaidya et al. | Jul 2015 | B2 |
| 9094327 | Jacobs et al. | Jul 2015 | B2 |
| 9178782 | Matthews et al. | Nov 2015 | B2 |
| 9208071 | Talagala et al. | Dec 2015 | B2 |
| 9218278 | Talagala et al. | Dec 2015 | B2 |
| 9231876 | Mir et al. | Jan 2016 | B2 |
| 9231888 | Bogdanski et al. | Jan 2016 | B2 |
| 9239804 | Kegel et al. | Jan 2016 | B2 |
| 9269438 | Nachimuthu et al. | Feb 2016 | B2 |
| 9276864 | Pradeep | Mar 2016 | B1 |
| 9369397 | Pathakota | Jun 2016 | B1 |
| 9436651 | Underwood et al. | Sep 2016 | B2 |
| 9455915 | Sinha et al. | Sep 2016 | B2 |
| 9460178 | Bashyam et al. | Oct 2016 | B2 |
| 9479426 | Munger et al. | Oct 2016 | B2 |
| 9496991 | Plamondon et al. | Nov 2016 | B2 |
| 9544234 | Markine | Jan 2017 | B1 |
| 9548924 | Pettit et al. | Jan 2017 | B2 |
| 9594521 | Blagodurov et al. | Mar 2017 | B2 |
| 9635121 | Mathew et al. | Apr 2017 | B2 |
| 9658978 | Lakshmanamurthy | May 2017 | B2 |
| 9742855 | Shuler et al. | Aug 2017 | B2 |
| 9762488 | Previdi et al. | Sep 2017 | B2 |
| 9762497 | Kishore et al. | Sep 2017 | B2 |
| 9830273 | BK et al. | Nov 2017 | B2 |
| 9838500 | Ilan et al. | Dec 2017 | B1 |
| 9853900 | Mula et al. | Dec 2017 | B1 |
| 9887923 | Chorafakis et al. | Feb 2018 | B2 |
| 10003544 | Liu et al. | Jun 2018 | B2 |
| 10009270 | Stark et al. | Jun 2018 | B1 |
| 10031857 | Menachem et al. | Jul 2018 | B2 |
| 10050896 | Yang et al. | Aug 2018 | B2 |
| 10061613 | Brooker et al. | Aug 2018 | B1 |
| 10063481 | Jiang et al. | Aug 2018 | B1 |
| 10089220 | McKelvie et al. | Oct 2018 | B1 |
| 10169060 | Vincent et al. | Jan 2019 | B1 |
| 10178035 | Dillon | Jan 2019 | B2 |
| 10200279 | Aljaedi | Feb 2019 | B1 |
| 10218634 | Aldebert et al. | Feb 2019 | B2 |
| 10270700 | Burnette et al. | Apr 2019 | B2 |
| 10305772 | Zur et al. | May 2019 | B2 |
| 10331590 | MacNamara et al. | Jun 2019 | B2 |
| 10353833 | Hagspiel et al. | Jul 2019 | B2 |
| 10454835 | Contavalli et al. | Oct 2019 | B2 |
| 10498672 | Graham et al. | Dec 2019 | B2 |
| 10567307 | Fairhurst et al. | Feb 2020 | B2 |
| 10728173 | Agrawal et al. | Jul 2020 | B1 |
| 10802828 | Volpe et al. | Oct 2020 | B1 |
| 10817502 | Talagala et al. | Oct 2020 | B2 |
| 11128561 | Matthews et al. | Sep 2021 | B1 |
| 11271869 | Agrawal et al. | Mar 2022 | B1 |
| 11416749 | Bshara et al. | Aug 2022 | B2 |
| 11444886 | Stawitzky et al. | Sep 2022 | B1 |
| 20010010692 | Sindhu et al. | Aug 2001 | A1 |
| 20010047438 | Forin | Nov 2001 | A1 |
| 20020174279 | Wynne et al. | Nov 2002 | A1 |
| 20030016808 | Hu et al. | Jan 2003 | A1 |
| 20030041168 | Musoll | Feb 2003 | A1 |
| 20030110455 | Baumgartner et al. | Jun 2003 | A1 |
| 20030174711 | Shankar | Sep 2003 | A1 |
| 20030200363 | Futral | Oct 2003 | A1 |
| 20030223420 | Ferolito | Dec 2003 | A1 |
| 20040008716 | Stiliadis | Jan 2004 | A1 |
| 20040059828 | Hooper et al. | Mar 2004 | A1 |
| 20040095882 | Hamzah et al. | May 2004 | A1 |
| 20040133634 | Luke et al. | Jul 2004 | A1 |
| 20040223452 | Santos et al. | Nov 2004 | A1 |
| 20050021837 | Haselhorst et al. | Jan 2005 | A1 |
| 20050047334 | Paul | Mar 2005 | A1 |
| 20050088969 | Carlsen et al. | Apr 2005 | A1 |
| 20050091396 | Nilakantan et al. | Apr 2005 | A1 |
| 20050108444 | Flauaus et al. | May 2005 | A1 |
| 20050108518 | Pandya | May 2005 | A1 |
| 20050152274 | Simpson | Jul 2005 | A1 |
| 20050182854 | Pinkerton et al. | Aug 2005 | A1 |
| 20050270974 | Mayhew | Dec 2005 | A1 |
| 20050270976 | Yang et al. | Dec 2005 | A1 |
| 20060023705 | Zoranovic et al. | Feb 2006 | A1 |
| 20060067347 | Naik et al. | Mar 2006 | A1 |
| 20060075480 | Noehring et al. | Apr 2006 | A1 |
| 20060174251 | Pope et al. | Aug 2006 | A1 |
| 20060203728 | Kwan et al. | Sep 2006 | A1 |
| 20070061433 | Reynolds et al. | Mar 2007 | A1 |
| 20070070901 | Aloni et al. | Mar 2007 | A1 |
| 20070198804 | Moyer | Aug 2007 | A1 |
| 20070211746 | Oshikiri et al. | Sep 2007 | A1 |
| 20070242611 | Archer et al. | Oct 2007 | A1 |
| 20070268825 | Corwin et al. | Nov 2007 | A1 |
| 20080013453 | Chiang et al. | Jan 2008 | A1 |
| 20080013549 | Okagawa et al. | Jan 2008 | A1 |
| 20080071757 | Ichiriu et al. | Mar 2008 | A1 |
| 20080084864 | Archer et al. | Apr 2008 | A1 |
| 20080091915 | Moertl et al. | Apr 2008 | A1 |
| 20080147881 | Krishnamurthy et al. | Jun 2008 | A1 |
| 20080159138 | Shepherd et al. | Jul 2008 | A1 |
| 20080253289 | Naven et al. | Oct 2008 | A1 |
| 20090003212 | Kwan et al. | Jan 2009 | A1 |
| 20090010157 | Holmes et al. | Jan 2009 | A1 |
| 20090013175 | Elliott | Jan 2009 | A1 |
| 20090055496 | Garg et al. | Feb 2009 | A1 |
| 20090092046 | Naven et al. | Apr 2009 | A1 |
| 20090141621 | Fan et al. | Jun 2009 | A1 |
| 20090198958 | Arimilli et al. | Aug 2009 | A1 |
| 20090259713 | Blumrich et al. | Oct 2009 | A1 |
| 20090285222 | Hoover et al. | Nov 2009 | A1 |
| 20100061241 | Sindhu et al. | Mar 2010 | A1 |
| 20100169608 | Kuo et al. | Jul 2010 | A1 |
| 20100172260 | Kwan et al. | Jul 2010 | A1 |
| 20100183024 | Gupta | Jul 2010 | A1 |
| 20100220595 | Petersen | Sep 2010 | A1 |
| 20100274876 | Kagan et al. | Oct 2010 | A1 |
| 20100278195 | Wagh | Nov 2010 | A1 |
| 20100302942 | Shankar et al. | Dec 2010 | A1 |
| 20100316053 | Miyoshi et al. | Dec 2010 | A1 |
| 20110051724 | Scott et al. | Mar 2011 | A1 |
| 20110066824 | Bestler | Mar 2011 | A1 |
| 20110072179 | Lacroute et al. | Mar 2011 | A1 |
| 20110099326 | Jung et al. | Apr 2011 | A1 |
| 20110110383 | Yang et al. | May 2011 | A1 |
| 20110128959 | Bando et al. | Jun 2011 | A1 |
| 20110158096 | Leung et al. | Jun 2011 | A1 |
| 20110158248 | Vorunganti et al. | Jun 2011 | A1 |
| 20110164496 | Loh et al. | Jul 2011 | A1 |
| 20110173370 | Jacobs et al. | Jul 2011 | A1 |
| 20110264822 | Ferguson et al. | Oct 2011 | A1 |
| 20110276699 | Pedersen | Nov 2011 | A1 |
| 20110280125 | Jayakumar | Nov 2011 | A1 |
| 20110320724 | Mejdrich et al. | Dec 2011 | A1 |
| 20120093505 | Yeap et al. | Apr 2012 | A1 |
| 20120102506 | Hopmann et al. | Apr 2012 | A1 |
| 20120117423 | Andrade et al. | May 2012 | A1 |
| 20120137075 | Vorbach | May 2012 | A1 |
| 20120144064 | Parker et al. | Jun 2012 | A1 |
| 20120144065 | Parker et al. | Jun 2012 | A1 |
| 20120147752 | Ashwood-Smith et al. | Jun 2012 | A1 |
| 20120170462 | Sinha | Jul 2012 | A1 |
| 20120170575 | Mehra | Jul 2012 | A1 |
| 20120213118 | Lindsay et al. | Aug 2012 | A1 |
| 20120250512 | Jagadeeswaran et al. | Oct 2012 | A1 |
| 20120287821 | Godfrey et al. | Nov 2012 | A1 |
| 20120297083 | Ferguson et al. | Nov 2012 | A1 |
| 20120300669 | Zahavi | Nov 2012 | A1 |
| 20120314707 | Epps et al. | Dec 2012 | A1 |
| 20130010636 | Regula | Jan 2013 | A1 |
| 20130039169 | Schlansker et al. | Feb 2013 | A1 |
| 20130051397 | Guo | Feb 2013 | A1 |
| 20130060944 | Archer et al. | Mar 2013 | A1 |
| 20130103777 | Kagan et al. | Apr 2013 | A1 |
| 20130121178 | Mainaud et al. | May 2013 | A1 |
| 20130136090 | Liu et al. | May 2013 | A1 |
| 20130182704 | Jacobs et al. | Jul 2013 | A1 |
| 20130194927 | Yamaguchi et al. | Aug 2013 | A1 |
| 20130203422 | Masputra et al. | Aug 2013 | A1 |
| 20130205002 | Wang et al. | Aug 2013 | A1 |
| 20130208593 | Nandagopal | Aug 2013 | A1 |
| 20130246552 | Underwood et al. | Sep 2013 | A1 |
| 20130290673 | Archer et al. | Oct 2013 | A1 |
| 20130301645 | Bogdanski et al. | Nov 2013 | A1 |
| 20130304988 | Totolos et al. | Nov 2013 | A1 |
| 20130311525 | Neerincx et al. | Nov 2013 | A1 |
| 20130329577 | Suzuki et al. | Dec 2013 | A1 |
| 20130336164 | Yang et al. | Dec 2013 | A1 |
| 20140019661 | Hormuth et al. | Jan 2014 | A1 |
| 20140032695 | Michels et al. | Jan 2014 | A1 |
| 20140036680 | Lih et al. | Feb 2014 | A1 |
| 20140064082 | Yeung et al. | Mar 2014 | A1 |
| 20140095753 | Crupnicoff et al. | Apr 2014 | A1 |
| 20140098675 | Frost et al. | Apr 2014 | A1 |
| 20140119367 | Han et al. | May 2014 | A1 |
| 20140122560 | Ramey et al. | May 2014 | A1 |
| 20140129664 | McDaniel et al. | May 2014 | A1 |
| 20140133292 | Yamatsu et al. | May 2014 | A1 |
| 20140136646 | Tamir et al. | May 2014 | A1 |
| 20140169173 | Naouri et al. | Jun 2014 | A1 |
| 20140185621 | Decusatis et al. | Jul 2014 | A1 |
| 20140189174 | Ajanovic et al. | Jul 2014 | A1 |
| 20140207881 | Nussle et al. | Jul 2014 | A1 |
| 20140211804 | Makikeni et al. | Jul 2014 | A1 |
| 20140226488 | Shamis et al. | Aug 2014 | A1 |
| 20140241164 | Cociglio et al. | Aug 2014 | A1 |
| 20140258438 | Ayoub | Sep 2014 | A1 |
| 20140301390 | Scott et al. | Oct 2014 | A1 |
| 20140307554 | Basso et al. | Oct 2014 | A1 |
| 20140325013 | Tamir et al. | Oct 2014 | A1 |
| 20140328172 | Kumar et al. | Nov 2014 | A1 |
| 20140347997 | Bergamasco et al. | Nov 2014 | A1 |
| 20140362698 | Arad | Dec 2014 | A1 |
| 20140369360 | Carlstrom | Dec 2014 | A1 |
| 20140379847 | Williams | Dec 2014 | A1 |
| 20150003247 | Mejia et al. | Jan 2015 | A1 |
| 20150006849 | Xu et al. | Jan 2015 | A1 |
| 20150009823 | Ganga et al. | Jan 2015 | A1 |
| 20150026361 | Matthews et al. | Jan 2015 | A1 |
| 20150029848 | Jain | Jan 2015 | A1 |
| 20150055476 | Decusatis et al. | Feb 2015 | A1 |
| 20150055661 | Boucher et al. | Feb 2015 | A1 |
| 20150067095 | Gopal et al. | Mar 2015 | A1 |
| 20150089495 | Persson et al. | Mar 2015 | A1 |
| 20150103667 | Elias et al. | Apr 2015 | A1 |
| 20150124826 | Edsall et al. | May 2015 | A1 |
| 20150146527 | Kishore et al. | May 2015 | A1 |
| 20150154004 | Aggarwal | Jun 2015 | A1 |
| 20150161064 | Pope | Jun 2015 | A1 |
| 20150180782 | Rimmer et al. | Jun 2015 | A1 |
| 20150186318 | Kim et al. | Jul 2015 | A1 |
| 20150193262 | Archer et al. | Jul 2015 | A1 |
| 20150195388 | Snyder et al. | Jul 2015 | A1 |
| 20150208145 | Parker et al. | Jul 2015 | A1 |
| 20150220449 | Stark et al. | Aug 2015 | A1 |
| 20150237180 | Swartzentruber et al. | Aug 2015 | A1 |
| 20150244617 | Nakil et al. | Aug 2015 | A1 |
| 20150244804 | Warfield et al. | Aug 2015 | A1 |
| 20150261434 | Kagan et al. | Sep 2015 | A1 |
| 20150263955 | Talaski et al. | Sep 2015 | A1 |
| 20150263994 | Haramaty et al. | Sep 2015 | A1 |
| 20150288626 | Aybay | Oct 2015 | A1 |
| 20150365337 | Pannell | Dec 2015 | A1 |
| 20150370586 | Cooper et al. | Dec 2015 | A1 |
| 20160006664 | Sabato et al. | Jan 2016 | A1 |
| 20160012002 | Arimilli et al. | Jan 2016 | A1 |
| 20160028613 | Haramaty et al. | Jan 2016 | A1 |
| 20160065455 | Wang et al. | Mar 2016 | A1 |
| 20160094450 | Ghanwani et al. | Mar 2016 | A1 |
| 20160134518 | Callon et al. | May 2016 | A1 |
| 20160134535 | Callon | May 2016 | A1 |
| 20160134559 | Abel et al. | May 2016 | A1 |
| 20160134573 | Gagliardi et al. | May 2016 | A1 |
| 20160142318 | Beecroft | May 2016 | A1 |
| 20160154756 | Dodson et al. | Jun 2016 | A1 |
| 20160182383 | Pedersen | Jun 2016 | A1 |
| 20160205023 | Janardhanan | Jul 2016 | A1 |
| 20160226797 | Aravinthan et al. | Aug 2016 | A1 |
| 20160254991 | Eckert et al. | Sep 2016 | A1 |
| 20160259394 | Ragavan | Sep 2016 | A1 |
| 20160283422 | Crupnicoff et al. | Sep 2016 | A1 |
| 20160285545 | Schmidtke et al. | Sep 2016 | A1 |
| 20160285677 | Kashyap et al. | Sep 2016 | A1 |
| 20160294694 | Parker et al. | Oct 2016 | A1 |
| 20160294926 | Zur et al. | Oct 2016 | A1 |
| 20160301610 | Amit et al. | Oct 2016 | A1 |
| 20160344620 | G. Santos et al. | Nov 2016 | A1 |
| 20160381189 | Caulfield et al. | Dec 2016 | A1 |
| 20170024263 | Verplanken | Jan 2017 | A1 |
| 20170039063 | Gopal et al. | Feb 2017 | A1 |
| 20170041239 | Goldenberg et al. | Feb 2017 | A1 |
| 20170048144 | Liu | Feb 2017 | A1 |
| 20170054633 | Underwood et al. | Feb 2017 | A1 |
| 20170091108 | Arellano et al. | Mar 2017 | A1 |
| 20170097840 | Bridgers | Apr 2017 | A1 |
| 20170103108 | Datta et al. | Apr 2017 | A1 |
| 20170118090 | Pettit et al. | Apr 2017 | A1 |
| 20170118098 | Littlejohn et al. | Apr 2017 | A1 |
| 20170153852 | Ma et al. | Jun 2017 | A1 |
| 20170177541 | Berman et al. | Jun 2017 | A1 |
| 20170220500 | Tong | Aug 2017 | A1 |
| 20170237654 | Turner et al. | Aug 2017 | A1 |
| 20170237671 | Rimmer et al. | Aug 2017 | A1 |
| 20170242753 | Sherlock et al. | Aug 2017 | A1 |
| 20170250914 | Caulfield et al. | Aug 2017 | A1 |
| 20170251394 | Johansson et al. | Aug 2017 | A1 |
| 20170270051 | Chen et al. | Sep 2017 | A1 |
| 20170272331 | Lissack | Sep 2017 | A1 |
| 20170272370 | Ganga et al. | Sep 2017 | A1 |
| 20170286316 | Doshi et al. | Oct 2017 | A1 |
| 20170289066 | Haramaty et al. | Oct 2017 | A1 |
| 20170295098 | Watkins et al. | Oct 2017 | A1 |
| 20170324664 | Xu et al. | Nov 2017 | A1 |
| 20170371778 | McKelvie et al. | Dec 2017 | A1 |
| 20180004705 | Menachem et al. | Jan 2018 | A1 |
| 20180019948 | Patwardhan et al. | Jan 2018 | A1 |
| 20180026878 | Zahavi et al. | Jan 2018 | A1 |
| 20180077064 | Wang | Mar 2018 | A1 |
| 20180083868 | Cheng | Mar 2018 | A1 |
| 20180097645 | Rajagopalan et al. | Apr 2018 | A1 |
| 20180097912 | Chumbalkar et al. | Apr 2018 | A1 |
| 20180113618 | Chan et al. | Apr 2018 | A1 |
| 20180115469 | Erickson et al. | Apr 2018 | A1 |
| 20180131602 | Civanlar et al. | May 2018 | A1 |
| 20180131678 | Agarwal et al. | May 2018 | A1 |
| 20180150374 | Ratcliff | May 2018 | A1 |
| 20180152317 | Chang et al. | May 2018 | A1 |
| 20180152357 | Natham et al. | May 2018 | A1 |
| 20180173557 | Nakil et al. | Jun 2018 | A1 |
| 20180183724 | Callard et al. | Jun 2018 | A1 |
| 20180191609 | Caulfield et al. | Jul 2018 | A1 |
| 20180198736 | Labonte et al. | Jul 2018 | A1 |
| 20180212876 | Bacthu et al. | Jul 2018 | A1 |
| 20180212902 | Steinmacher-Burow | Jul 2018 | A1 |
| 20180219804 | Graham et al. | Aug 2018 | A1 |
| 20180225238 | Karguth et al. | Aug 2018 | A1 |
| 20180234343 | Zdornov et al. | Aug 2018 | A1 |
| 20180254945 | Bogdanski et al. | Sep 2018 | A1 |
| 20180260324 | Marathe et al. | Sep 2018 | A1 |
| 20180278540 | Shalev et al. | Sep 2018 | A1 |
| 20180287928 | Levi | Oct 2018 | A1 |
| 20180323898 | Dods | Nov 2018 | A1 |
| 20180335974 | Simionescu et al. | Nov 2018 | A1 |
| 20180341494 | Sood et al. | Nov 2018 | A1 |
| 20190007349 | Wang et al. | Jan 2019 | A1 |
| 20190018808 | Beard et al. | Jan 2019 | A1 |
| 20190036771 | Sharpless et al. | Jan 2019 | A1 |
| 20190042337 | Dinan et al. | Feb 2019 | A1 |
| 20190042518 | Marolia | Feb 2019 | A1 |
| 20190044809 | Willis et al. | Feb 2019 | A1 |
| 20190044827 | Ganapathi et al. | Feb 2019 | A1 |
| 20190044863 | Mula et al. | Feb 2019 | A1 |
| 20190044872 | Ganapathi et al. | Feb 2019 | A1 |
| 20190044875 | Murty et al. | Feb 2019 | A1 |
| 20190052327 | Motozuka et al. | Feb 2019 | A1 |
| 20190058663 | Song | Feb 2019 | A1 |
| 20190068501 | Schneider et al. | Feb 2019 | A1 |
| 20190081903 | Kobayashi et al. | Mar 2019 | A1 |
| 20190095134 | Li | Mar 2019 | A1 |
| 20190104057 | Goel et al. | Apr 2019 | A1 |
| 20190104206 | Goel et al. | Apr 2019 | A1 |
| 20190108106 | Aggarwal et al. | Apr 2019 | A1 |
| 20190108332 | Glew et al. | Apr 2019 | A1 |
| 20190109791 | Mehra et al. | Apr 2019 | A1 |
| 20190121781 | Kasichainula | Apr 2019 | A1 |
| 20190140979 | Levi et al. | May 2019 | A1 |
| 20190146477 | Cella et al. | May 2019 | A1 |
| 20190171612 | Shahar et al. | Jun 2019 | A1 |
| 20190196982 | Rozas et al. | Jun 2019 | A1 |
| 20190199646 | Singh et al. | Jun 2019 | A1 |
| 20190253354 | Caulfield et al. | Aug 2019 | A1 |
| 20190280978 | Schmatz et al. | Sep 2019 | A1 |
| 20190294575 | Dennison et al. | Sep 2019 | A1 |
| 20190306134 | Shanbhogue et al. | Oct 2019 | A1 |
| 20190332314 | Zhang et al. | Oct 2019 | A1 |
| 20190334624 | Bernard | Oct 2019 | A1 |
| 20190356611 | Das Sharma et al. | Nov 2019 | A1 |
| 20190361728 | Kumar et al. | Nov 2019 | A1 |
| 20190379610 | Srinivasan et al. | Dec 2019 | A1 |
| 20200036644 | Belogolovy et al. | Jan 2020 | A1 |
| 20200084150 | Burstein et al. | Mar 2020 | A1 |
| 20200145725 | Eberle et al. | May 2020 | A1 |
| 20200177505 | Li | Jun 2020 | A1 |
| 20200177521 | Blumrich et al. | Jun 2020 | A1 |
| 20200259755 | Wang et al. | Aug 2020 | A1 |
| 20200272579 | Humphrey et al. | Aug 2020 | A1 |
| 20200274832 | Humphrey et al. | Aug 2020 | A1 |
| 20200334195 | Chen et al. | Oct 2020 | A1 |
| 20200349098 | Caulfield et al. | Nov 2020 | A1 |
| 20210081410 | Chavan et al. | Mar 2021 | A1 |
| 20210152494 | Johnsen et al. | May 2021 | A1 |
| 20210263779 | Haghighat et al. | Aug 2021 | A1 |
| 20210334206 | Colgrove et al. | Oct 2021 | A1 |
| 20210377156 | Michael et al. | Dec 2021 | A1 |
| 20210409351 | Das Sharma et al. | Dec 2021 | A1 |
| 20220131768 | Ganapathi et al. | Apr 2022 | A1 |
| 20220166705 | Froese | May 2022 | A1 |
| 20220200900 | Roweth | Jun 2022 | A1 |
| 20220210058 | Bataineh et al. | Jun 2022 | A1 |
| 20220217078 | Ford et al. | Jul 2022 | A1 |
| 20220217101 | Yefet et al. | Jul 2022 | A1 |
| 20220245072 | Roweth et al. | Aug 2022 | A1 |
| 20220278941 | Shalev et al. | Sep 2022 | A1 |
| 20220309025 | Chen et al. | Sep 2022 | A1 |
| 20230035420 | Sankaran et al. | Feb 2023 | A1 |
| 20230046221 | Pismenny et al. | Feb 2023 | A1 |
| Number | Date | Country |
|---|---|---|
| 101729609 | Jun 2010 | CN |
| 102932203 | Feb 2013 | CN |
| 110324249 | Oct 2019 | CN |
| 110601888 | Dec 2019 | CN |
| 0275135 | Jul 1988 | EP |
| 2187576 | May 2010 | EP |
| 2219329 | Aug 2010 | EP |
| 2947832 | Nov 2015 | EP |
| 3445006 | Feb 2019 | EP |
| 2003-244196 | Aug 2003 | JP |
| 3459653 | Oct 2003 | JP |
| 10-2012-0062864 | Jun 2012 | KR |
| 10-2012-0082739 | Jul 2012 | KR |
| 10-2014-0100529 | Aug 2014 | KR |
| 10-2015-0026939 | Mar 2015 | KR |
| 10-2015-0104056 | Sep 2015 | KR |
| 10-2017-0110106 | Oct 2017 | KR |
| 10-1850749 | Apr 2018 | KR |
| 2001069851 | Sep 2001 | WO |
| 0247329 | Jun 2002 | WO |
| 2003019861 | Mar 2003 | WO |
| 2004001615 | Dec 2003 | WO |
| 2005094487 | Oct 2005 | WO |
| 2007034184 | Mar 2007 | WO |
| 2009010461 | Jan 2009 | WO |
| 2009018232 | Feb 2009 | WO |
| 2014092780 | Jun 2014 | WO |
| 2014137382 | Sep 2014 | WO |
| 2014141005 | Sep 2014 | WO |
| 2018004977 | Jan 2018 | WO |
| 2018046703 | Mar 2018 | WO |
| 2019072072 | Apr 2019 | WO |
| Entry |
|---|
| Blumrich, M.A., et al.; “Exploiting Idle Resources in a High-Radix Switch for Supplemental Storage”; Nov. 2018; 13 pages. |
| Hong, Y.; “Mitigating the Cost, Performance, and Power Overheads Induced by Load Variations in Multicore Cloud Servers”; Fall 2013; 132 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US2020/24267, dated Jul. 6, 2020, 9 pages. |
| Awerbuch, B., et al.; “An On-Demand Secure Routing Protocol Resilient to Byzantine Failures”; Sep. 2002; 10 pages. |
| Belayneh L.W., et al.; “Method and Apparatus for Routing Data in an Inter-Nodal Communications Lattice of a Massively Parallel Computer System by Semi-Randomly Varying Routing Policies for Different Packets”; 2019; 3 pages. |
| Bhatele, A., et al.; “Analyzing Network Health and Congestion in Dragonfly-based Supercomputers”; May 23-27, 2016; 10 pages. |
| Chang, F., et al.; “PVW: Designing Vir PVW: Designing Virtual World Ser orld Server Infr er Infrastructur astructure”, 2010; 8 pages. |
| Chang, F., et al.; “PVW: Designing Virtual World Server Infrastructure”; 2010; 8 pages. |
| Chen, F., et al.; “Requirements for RoCEv3 Congestion Management”; Mar. 21, 2019; 8 pages. |
| Cisco Packet Tracer; “packet-tracer;—ping”; https://www.cisco.com/c/en/us/td/docs/security/asa/asa-command-reference/l-R/cmdref2/p1.html; 2017. |
| Cisco; “Understanding Rapid Spanning Tree Protocol (802.1w)”; Aug. 1, 2017; 13 pages. |
| Eardley, Ed, P; “Pre-Congestion Notification (PCN) Architecture”; Jun. 2009; 54 pages. |
| Escudero-Sahuquillo, J., et al.; “Combining Congested-Flow Isolation and Injection Throttling in HPC Interconnection Networks”; Sep. 13-16, 2011; 3 pages. |
| Huawei; “The Lossless Network For Data Centers”; Nov. 7, 2017; 15 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US2020/024248, dated Jul. 8, 2020, 11 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US20/024332, dated Jul. 8, 2020, 13 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US20/24243, dated Jul. 9, 2020, 10 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US20/24253, dated Jul. 6, 2020, 12 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US20/24256, dated Jul. 7, 2020, 11 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US20/24257, dated Jul. 7, 2020, 10 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US20/24258, dated Jul. 7, 2020, 9 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US20/24259, dated Jul. 9, 2020, 13 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US20/24260, dated Jul. 7, 2020, 11 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US20/24268, dated Jul. 9, 2020, 11 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US20/24269, dated Jul. 9, 2020, 11 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US20/24339, dated Jul. 8, 2020, 11 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US2020/024125, dated Jul. 10, 2020, 5 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US2020/024129, dated Jul. 10, 2020, 11 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US2020/024237, dated Jul. 14, 2020, 5 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US2020/024239, dated Jul. 14, 2020, 11 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US2020/024241, dated Jul. 14, 2020, 13 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US2020/024242, dated Jul. 6, 2020, 11 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US2020/024244, dated Jul. 13, 2020, 10 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US2020/024245, dated Jul. 14, 2020, 11 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US2020/024246, dated Jul. 14, 2020, 10 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US2020/024250, dated Jul. 14, 2020, 12 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US2020/024254, dated Jul. 13, 2020, 10 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US2020/024262, dated Jul. 13, 2020, 10 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US2020/024266, dated Jul. 9, 2020, 10 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US2020/024270, dated Jul. 10, 2020, 13 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US2020/024271, dated Jul. 9, 2020, 10 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US2020/024272, dated Jul. 9, 2020, 10 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US2020/024276, dated Jul. 13, 2020, 9 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US2020/024304, dated Jul. 15, 2020, 11 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US2020/024311, dated Jul. 17, 2020, 8 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US2020/024321, dated Jul. 9, 2020, 9 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US2020/024324, dated Jul. 14, 2020, 10 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US2020/024327, dated Jul. 10, 2020, 15 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US2020/24158, dated Jul. 6, 2020, 18 pages. |
| International Search Report and Written Opinion received for PCT Application No. PCT/US2020/24251, dated Jul. 6, 2020, 11 pages. |
| International Search Report and Written Opinion received for PCT Patent Application No. PCT/US20/24303, dated Oct. 21, 2020, 9 pages. |
| International Search Report and Written Opinion received for PCT Patent Application No. PCT/US20/24340, dated Oct. 26, 2020, 9 pages. |
| International Search Report and Written Opinion received for PCT Patent Application No. PCT/US20/24342, dated Oct. 27, 2020, 10 pages. |
| International Search Report and Written Opinion received for PCT Patent Application No. PCT/US2020/024192, dated Oct. 23, 2020, 9 pages. |
| International Search Report and Written Opinion received for PCT Patent Application No. PCT/US2020/024221, dated Oct. 26, 2020, 9 pages. |
| International Search Report cited in PCT/US2020/024170 dated Dec. 16, 2020; 3 pages. |
| Maabi, S., et al.; “ERFAN: Efficient reconfigurable fault-tolerant deflection routing algorithm for 3-D Network-on-Chip”; Sep. 6-9, 2016. |
| Maglione-Mathey, G., et al.; “Scalable Deadlock-Free Deterministic Minimal-Path Routing Engine for InfiniBand-Based Dragonfly Networks”; Aug. 21, 2017; 15 pages. |
| Mamidala, A. R., et al.; “Efficient Barrier and Allreduce on Infiniband clusters using multicast and adaptive algorithms”; Sep. 20-23, 2004; 10 pages. |
| Mammeri, Z; “Reinforcement Learning Based Routing in Networks: Review and Classification of Approaches”; Apr. 29, 2019; 35 pages. |
| Mollah; M. A., et al.; “High Performance Computing Systems. Performance Modeling, Benchmarking, and Simulation: 8th International Workshop”; Nov. 13, 2017. |
| Open Networking Foundation; “OpenFlow Switch Specification”; Mar. 26, 2015; 283 pages. |
| Prakash, P., et al.; “The TCP Outcast Problem: Exposing Unfairness in Data Center Networks”; 2011; 15 pages. |
| Ramakrishnan, K., et al.; “The Addition of Explicit Congestion Notification (ECN) to IP”; Sep. 2001; 63 pages. |
| Roth, P. C., et al; “MRNet: A Software-Based Multicast/Reduction Network for Scalable Tools1”; Nov. 15-21, 2003; 16 pages. |
| Silveira, J., et al.; “Preprocessing of Scenarios for Fast and Efficient Routing Reconfiguration in Fault-Tolerant NoCs”; Mar. 4-6, 2015. |
| Tsunekawa, K.; “Fair bandwidth allocation among LSPs for AF class accommodating TCP and UDP traffic in a Diffserv-capable MPLS network”; Nov. 17, 2005; 9 pages. |
| Underwood, K.D., et al.; “A hardware acceleration unit for MPI queue processing”; Apr. 18, 2005; 10 pages. |
| Wu, J.; “Fault-tolerant adaptive and minimal routing in mesh-connected multicomputers using extended safety levels”; Feb. 2000; 11 pages. |
| Xiang, D., et al.; “Fault-Tolerant Adaptive Routing in Dragonfly Networks”; Apr. 12, 2017; 15 pages. |
| Xiang, D., et al.; “Deadlock-Free Broadcast Routing in Dragonfly Networks without Virtual Channels”, submission to IEEE transactions on Parallel and Distributed Systems, 2015, 15 pages. |
| Extended European Search Report and Search Opinion received for EP Application No. 20809930.9, dated Mar. 2, 2023, 9 pages. |
| Extended European Search Report and Search Opinion received for EP Application No. 20810784.7, dated Mar. 9, 2023, 7 pages. |
| Ramakrishnan et al, RFC 3168, “The addition of Explicit Congestion Notification (ECN) to IP”, Sep. 2001 (Year: 2001). |
| Number | Date | Country | |
|---|---|---|---|
| 20220210094 A1 | Jun 2022 | US |
| Number | Date | Country | |
|---|---|---|---|
| 62852203 | May 2019 | US | |
| 62852289 | May 2019 | US | |
| 62852273 | May 2019 | US |
