Abstract: Distributed machine learning systems and other distributed computing systems are improved by embedding compute logic at the network switch level to perform collective actions, such as reduction operations, on gradients or other data processed by the nodes of the system. The switch is configured to recognize data units that carry data associated with a collective action that needs to be performed by the distributed system, referred to herein as “compute data,” and process that data using a compute subsystem within the switch. The compute subsystem includes a compute engine that is configured to perform various operations on the compute data, such as “reduction” operations, and forward the results back to the compute nodes. The reduction operations may include, for instance, summation, averaging, bitwise operations, and so forth. In this manner, the network switch may take over some or all of the processing of the distributed system during the collective phase.

Abstract: A network device includes ingress queues for storing data units while the data units are being processed by ingress packet processors, and a plurality of egress buffer memories for storing data units received from the ingress queues while the data units are being processed by the egress packet processors. First circuitry controls respective rates at which data units are transferred from ingress queues to egress buffer memories. Second circuitry monitors the egress buffer memories for congestion and sends, to the first circuitry, flow control messages related to congestion resulting of egress buffer memories. The first circuitry progressively increases over time a rate at which data from each ingress queue are transferred to an egress buffer memory in response to receiving a flow control message that indicates that congestion corresponding to the egress buffer memory has ended.

Abstract: A network device obtains measurement data for one or more device attributes or environmental factors, and compares the measurement data to respective ranges specified for the device attributes or the environmental factors. Different ranges for the device attributes or the environmental factors are associated with different operating regions (OREs) classified for the device. The operating state of the network device corresponds to a first ORE of the different OREs, and various tasks performed by the device in the operating state are based on configurations specified by the first ORE. Based on comparing the measurement data, the network device identifies a second ORE that includes ranges for the device attributes or the environmental factors that match the measurement data. The network device transitions the operating state to correspond to the second ORE, and adjusting the tasks performed by the device according to configurations specified by the second ORE.

Abstract: The present disclosure describes apparatuses and methods for machine learning-enabled (ML-enabled) queue management for network devices. In some aspects, an ML-enabled queue manager of a network device initializes a queue management setting with a randomized value and the device processes packets through the queue based on the queue management setting. The ML-enabled queue manager measures a performance metric of the queue and provides, to an ML algorithm, an indication of the queue management setting and an indication of the performance metric of the queue. The ML-enabled queue manager then receives, from the machine learning algorithm, an updated queue management setting and configures the queue with the updated queue management setting to process subsequent packets based on the updated queue management setting. By so doing, the ML-enabled queue manager may tune one or more queue management settings of the queue to optimize performance of the network device.

Abstract: Packet-switching operations in a network device are managed based on the detection of excessive-rate traffic flows. A network device receives a data unit, determines the traffic flow to which the data unit belongs, and updates flow tracking information for that flow. The network device utilizes the tracking information to determine when a rate at which the network device is receiving data belonging to the flow exceeds an excessive-rate threshold and is thus an excessive-rate flow. The network device may enable one or more excessive-rate policies on an excessive-rate traffic flow. Such a policy may include any number of features that affect how the device handles data units belonging to the flow, such as excessive-rate notification, differentiated discard, differentiated congestion notification, and reprioritization. Memory and other resource optimizations for such flow tracking and management are also described.

Abstract: Automatic load-balancing techniques in a network device are used to select, from a multipath group, a path to assign to a flow based on observed state attributes such as path state(s), device state(s), port state(s), or queue state(s) of the paths. A mapping of the path previously assigned to a flow or group of flows (e.g., on account of having then been optimal in view of the observed state attributes) is maintained, for example, in a table. So long as the flow(s) are active and the path is still valid, the mapped path is selected for subsequent data units belonging to the flow(s), which may, among other effects, avoid or reduce packet re-ordering. However, if the flow(s) go idle, or if the mapped path fails, a new optimal path may be assigned to the flow(s) from the multipath group.

Abstract: Packet metadata for incoming packets are buffered in queue selection buffers associated with a port of a network node. Packet data for outgoing packets are buffered in a port selection buffer associated with the port. At a selection clock cycle, while a port scheduler of the network node selects a subset of the packet data for a subset of the outgoing packets from the port selection buffer, a queue scheduler of the port concurrently selects a subset of the packet metadata for a subset of the incoming packets from the queue selection buffers and adds new packet data for new outgoing packets to the port selection buffer of the port. The new packet data are derived based at least in part on the subset of the packet metadata for the subset of the incoming packets.

Abstract: When a measure of buffer space queued for garbage collection in a network device grows beyond a certain threshold, one or more actions are taken to decreasing an enqueue rate of certain classes of traffic, such as of multicast traffic, whose reception may have caused and/or be likely to exacerbate garbage-collection-related performance issues. When the amount of buffer space queued for garbage collection shrinks to an acceptable level, these one or more actions may be reversed. In an embodiment, to more optimally handle multi-destination traffic, queue admission control logic for high-priority multi-destination data units, such as mirrored traffic, may be performed for each destination of the data units prior to linking the data units to a replication queue. If a high-priority multi-destination data unit is admitted to any queue, the high-priority multi-destination data unit can no longer be dropped, and is linked to a replication queue for replication.

Abstract: Packet-switching operations in a network device are managed based on the detection of excessive-rate traffic flows. A network device receives a data unit, determines the traffic flow to which the data unit belongs, and updates flow tracking information for that flow. The network device utilizes the tracking information to determine when a rate at which the network device is receiving data belonging to the flow exceeds an excessive-rate threshold and is thus an excessive-rate flow. The network device may enable one or more excessive-rate policies on an excessive-rate traffic flow. Such a policy may include any number of features that affect how the device handles data units belonging to the flow, such as excessive-rate notification, differentiated discard, differentiated congestion notification, and reprioritization. Memory and other resource optimizations for such flow tracking and management are also described.

Abstract: Distributed machine learning systems and other distributed computing systems are improved by compute logic embedded in extension modules coupled directly to network switches. The compute logic performs collective actions, such as reduction operations, on gradients or other compute data processed by the nodes of the system. The reduction operations may include, for instance, summation, averaging, bitwise operations, and so forth. In this manner, the extension modules may take over some or all of the processing of the distributed system during the collective phase. An inline version of the module sits between a switch and the network. Data units carrying compute data are intercepted and processed using the compute logic, while other data units pass through the module transparently to or from the switch. Multiple modules may be connected to the switch, each coupled to a different group of nodes, and sharing intermediate results. A sidecar version is also described.

Abstract: A traffic manager is shared amongst two or more egress blocks of a network device, thereby allowing traffic management resources to be shared between the egress blocks. Schedulers within a traffic manager may generate and queue read instructions for reading buffered portions of data units that are ready to be sent to the egress blocks. The traffic manager may be configured to select a read instruction for a given buffer bank from the read instruction queues based on a scoring mechanism or other selection logic. To avoid sending too much data to an egress block during a given time slot, once a data unit portion has been read from the buffer, it may be temporarily stored in a shallow read data cache. Alternatively, a single, non-bank specific controller may determine all of the read instructions and write operations that should be executed in a given time slot.

Abstract: A pre-scaled accumulated byte count of a port of a network node over a sampling period is scaled with a scaling factor to generate a scaled accumulated byte count. The pre-scaled accumulated byte count represents a total number of bytes in packets transferred by the port. The scaling factor represents a first port-specific attribute of the port and scales a port-specific maximum throughput of the port to a specific maximum port throughput of the network node. An iterative vector encoding method is applied to the scaled accumulated byte count to generate an encoded bit vector comprising bits respectively ordered bit positions. Each set bit of the encoded bit vector represents a respective weighted value of port utilization of the port. The encoded bit vector is stored, at a map location, in an operational statistics map.

Abstract: Certain hash-based operations in network devices and other devices, such as mapping and/or lookup operations, are improved by manipulating a hash key prior to executing a hash function on the hash key and/or by manipulating outputs of a hash function. A device may be configured to manipulate hash keys and/or outputs using manipulation logic based on one or more predefined manipulation values. A similar hash-based operation may be performed by multiple devices within a network of computing devices. Different devices may utilize different predefined manipulation values for their respective implementations of the manipulation logic. For instance, each device may assign itself a random mask value for key transformation logic as part of an initialization process when the device powers up and/or each time the device reboots. In an embodiment, described techniques may increase the entropy of hashing function outputs in certain contexts, thereby increasing the effectiveness of certain hashing functions.

Abstract: An improved buffer for networking and other computing devices comprises multiple memory instances, each having a distinct set of entries. Transport data units (“TDUs”) are divided into storage data units (“SDUs”), and each SDU is stored within a separate entry of a separate memory instance in a logical bank. One or more grids of the memory instances are organized into overlapping logical banks. The logical banks are arranged into views. Different destinations or other entities are assigned different views of the buffer. A memory instance may be shared between logical banks in different views. When overlapping logical banks are accessed concurrently, data in a memory instance that they share may be recovered using a parity SDU in another memory instance. The shared buffering enables more efficient buffer usage in a network device with a traffic manager shared amongst egress bocks. Example read and write algorithms for such buffers are disclosed.

Abstract: A network device organizes packets into various queues, in which the packets await processing. Queue management logic tracks how long certain packet(s), such as a designated marker packet, remain in a queue. Based thereon, the logic produces a measure of delay for the queue, referred to herein as the “queue delay.” Based on a comparison of the current queue delay to one or more thresholds, various associated delay-based actions may be performed, such as tagging and/or dropping packets departing from the queue, or preventing addition enqueues to the queue. In an embodiment, a queue may be expired based on the queue delay, and all packets dropped. In other embodiments, when a packet is dropped prior to enqueue into an assigned queue, copies of some or all of the packets already within the queue at the time the packet was dropped may be forwarded to a visibility component for analysis.

Abstract: A switch or other network device may be configured as an ingress edge telemetry node in a telemetry domain. The ingress edge telemetry node may clone certain data units it processes, for example in response to certain telemetry triggers being met. The ingress edge telemetry node may further inject telemetry and/or other data into the cloned data unit. The cloned data unit continues along the same path as the original data unit until it reaches an egress edge telemetry node in the telemetry domain. The second node extracts the telemetry data from the cloned data unit and sends telemetry information based thereon to a telemetry collector, while the original data unit continues to its final destination. Nodes along the path between the first node and the second node may be configured as transit telemetry nodes that insert or otherwise update the telemetry data.

Abstract: A network device accesses, from a queue corresponding to a port of the device, a packet for processing. The device identifies a present operating region (ORE) of one or more OREs specified for the device, an ORE being associated with at least one of (i) one or more device attributes, or (ii) one or more environmental factors associated with an environment in which the device is operational. The device determines a number of power credits available for processing one or more packets. In response to determining that the number of power credits available is non-negative, the device completes processing of the packet. The device computes, based at least on the present ORE, a power credit reduction for the packet, which corresponds to an amount of power for processing the packet, and reduces the number of power credits available by the power credit reduction for the packet.

Abstract: A packet to be forwarded over a computer network to a destination is received. A group of multiple network paths is available to forward to the packet to the destination. One or more path selection factors are determined to be used to identify a specific network load balancing algorithm to select a specific network path from the group of multiple network paths. The one or more path selection factors include at least one path selection factor determined based at least in part on a dynamic state of the computer network or a network node in the computer network. In response to selecting, by the specific network load balancing algorithm, the specific network path from among the group of multiple network paths, the packet is forwarded over the specific network path.

Abstract: Approaches, techniques, and mechanisms are disclosed for improving operations of a network switching device and/or network-at-large by utilizing queue delay as a basis for measuring congestion for the purposes of Automated Queue Management (“AQM”) and/or other congestion-based policies. Queue delay is an exact or approximate measure of the amount of time a data unit waits at a network device as a consequence of queuing, such as the amount of time the data unit spends in an egress queue while the data unit is being buffered by a traffic manager. Queue delay may be used as a substitute for queue size in existing AQM, Weighted Random Early Detection (“WRED”), Tail Drop, Explicit Congestion Notification (“ECN”), reflection, and/or other congestion management or notification algorithms. Or, a congestion score calculated based on the queue delay and one or more other metrics, such as queue size, may be used as a substitute.

Abstract: Efficient scaling of in-network compute operations to large numbers of compute nodes is disclosed. Each compute node is connected to a same plurality of network compute nodes, such as compute-enabled network switches. Compute processes at the compute nodes generate local gradients or other vectors by, for instance, performing a forward pass on a neural network. Each vector comprises values for a same set of vector elements. Each network compute node is assigned to, based on the local vectors, reduce vector data for a different a subset of the vector elements. Each network compute node returns a result chunk for the elements it processed back to each of the compute nodes, whereby each compute node receives the full result vector. This configuration may, in some embodiments, reduce buffering, processing, and/or other resource requirements for the network compute node or network at large.