Abstract: Methods, systems, and devices for providing remote link failure management using a remote link failure management engine of an artificial intelligence (AI) backend network system are described. Remote link failure management includes hardware-based techniques associated with AI hardware (e.g., an AI accelerator or AI System on Chip “SoC) where the techniques are employed to address malfunctions or breakdowns in components that facilitate the connectivity and communication between AI hardware and other components. The remote link failure management engine supports detecting, mitigating, and recovering from failures in the ports and links in AI hardware. In particular, remote link failure management can be provided for AI hardware based on an Artificial Intelligence Transport Layer Protocol (ATL).

Abstract: Methods, systems, and devices for providing hardware-based failure recovery management using a hardware-based failure recovery management engine of an artificial intelligence (AI) backend network system are described. Hardware-based failure recovery management includes hardware-based techniques associated with AI hardware (e.g., an AI accelerator or AI System on Chip “SoC”) where the techniques and mechanisms are employed to address malfunctions or breakdowns in components that facilitate the connectivity and communication between AI hardware and other components. The hardware-based failure recovery management engine supports detecting, mitigating, and recovering from failures in the ports and links in AI hardware. In particular, in operation, the hardware-based failure recovery management engine supports disabling a port of a plurality of ports in an ANC in AI hardware.

Abstract: A method for computer workload allocation at a system-on-chip (SoC) includes, at a load balancing controller of the SoC, dividing a computer workload for distributed processing between each of a plurality of hardware accelerators of the SoC as a plurality of accelerator-specific data allocations. At a hardware accelerator of the plurality of hardware accelerators, after receiving an accelerator-specific data allocation from the load balancing controller, a resulting dataset output by the hardware accelerator is divided between a plurality of network interface controllers (NICs) of the SoC as a plurality of NIC-specific data allocations. At an NIC of the plurality of NICs an NIC-specific data allocation assigned to the NIC is divided between a plurality of network ports of the NIC for transmission over a computer network.

Abstract: Control of congestion in a network between a transmission source and a transmission destination is provided, including receiving, at the transmission source, a packet transmission acknowledgment associated with a transmission packet of a data stream received at a transmission destination, incrementing a first window resize counter based on a congestion indicator, determining that a window resize condition is satisfied based on the incrementing of the first window resize counter and resizing, based on satisfaction of the window resize condition, a transmission window of the data stream. The packet transmission acknowledgment includes the congestion indicator, which represents whether network traffic experienced by the transmission packet between the transmission source and the transmission destination satisfies a congestion condition.

Abstract: A computing system for transport layer network recovery on a packet-switched computer network includes a source computing device with a processor that executes a network traffic communication module, a load balancing module, and a congestion control module. The network traffic communication module provisions a plurality of source ports to transmit outbound packets to a destination computing device, each source port being associated with a respective network path. The load balancing module assigns each outbound packet to one of the source ports using a port scheduling algorithm to uniformly distribute the packets among the source ports and associated network paths. The congestion control module detects a congestion control condition for a packet transmitted via a source port associated with a congested network path. The load balancing module assigns a next source port for a next outbound packet from a remainder of the source ports not associated with the congested network path.

Abstract: Techniques and algorithms for monitoring network congestion and for triggering a flow to follow a new path through a network. The network is monitored, and network feedback data is acquired, where that data indicates whether the network is congested. If the network is congested, a feedback-driven algorithm can trigger a flow to follow a new path. By triggering the flow to follow the new path, congestion in the network is reduced. To identify congestion, the feedback data is analyzed to determine whether flows are colliding. The feedback-driven algorithm determines that a network remapping event is to occur in an attempt to alleviate the congestion. A flow is then selected to be remapped to alleviate the congestion.

Abstract: Embodiments of the present disclosure include systems and methods for fault detection and recovery over a network. A value of a set of values is stored in packets transmitted during a data transaction between a source and destination. The value corresponds to ports used by one or more switches in the path between the source and destination. The destination includes the value in an acknowledgement packet. Logic circuits in the source device track packets and corresponding values. When a status indicates a particular packet has not received an acknowledgement, the value for the packet may be removed from the set of values. Particular ports that may be congested or down may be detected and the packets re-routed using the logic circuits in the source device.

Abstract: Techniques and algorithms for monitoring network congestion and for triggering a flow to follow a new path through a network. The network is monitored, and network feedback data is acquired, where that data indicates whether the network is congested. If the network is congested, a feedback-driven algorithm can trigger a flow to follow a new path. By triggering the flow to follow the new path, congestion in the network is reduced. To identify congestion, the feedback data is analyzed to determine whether flows are colliding. The feedback-driven algorithm determines that a network remapping event is to occur in an attempt to alleviate the congestion. A flow is then selected to be remapped to alleviate the congestion.

Abstract: A computing system for transport layer network recovery on a packet-switched computer network includes a source computing device with a processor that executes a network traffic communication module, a load balancing module, and a congestion control module. The network traffic communication module provisions a plurality of source ports to transmit outbound packets to a destination computing device, each source port being associated with a respective network path. The load balancing module assigns each outbound packet to one of the source ports using a port scheduling algorithm to uniformly distribute the packets among the source ports and associated network paths. The congestion control module detects a congestion control condition for a packet transmitted via a source port associated with a congested network path. The load balancing module assigns a next source port for a next outbound packet from a remainder of the source ports not associated with the congested network path.

Abstract: Control of congestion in a network between a transmission source and a transmission destination is provided, including receiving, at the transmission source, a packet transmission acknowledgment associated with a transmission packet of a data stream received at a transmission destination, incrementing a first window resize counter based on a congestion indicator, determining that a window resize condition is satisfied based on the incrementing of the first window resize counter and resizing, based on satisfaction of the window resize condition, a transmission window of the data stream. The packet transmission acknowledgment includes the congestion indicator, which represents whether network traffic experienced by the transmission packet between the transmission source and the transmission destination satisfies a congestion condition.

Abstract: A method includes dividing a flow of packets between two endpoints into multiple sub-flows and assigning each sub-flow a different hash seed. Packets from the sub-flows are transmitted over a network and performance for each sub-flow is monitored. The hash seed of an underperforming first sub-flow is replaced with a new hash seed based on a performance status of the first sub-flow.

Abstract: Systems and methods for sending and receiving messages, including training data, using a multi-path packet spraying protocol are described. A method includes segmenting a message into a set of data packets comprising training data. The method further includes initiating transmission of the set of data packets to a receiving node. The method further includes spraying the set of data packets across the switch fabric in accordance with the multi-path spraying protocol such that depending upon a value of a fabric determination field associated with a respective data packet, the respective data packet can traverse via any one of a plurality of paths offered by the switch fabric for a connection between the sending node to the receiving node. The method further includes initiating transmission of synchronization packets to the receiving node, where unlike the set of data packets, the synchronization packets are not sprayed across the switch fabric.

Abstract: Techniques and algorithms for monitoring network congestion and for triggering a flow to follow a new path through a network. The network is monitored, and network feedback data is acquired, where that data indicates whether the network is congested. If the network is congested, a feedback-driven algorithm can trigger a flow to follow a new path. By triggering the flow to follow the new path, congestion in the network is reduced. To identify congestion, the feedback data is analyzed to determine whether flows are colliding. The feedback-driven algorithm determines that a network remapping event is to occur in an attempt to alleviate the congestion. A flow is then selected to be remapped to alleviate the congestion.

Abstract: Embodiments of the present disclosure include systems and methods for fault detection and recovery over a network. A value of a set of values is stored in packets transmitted during a data transaction between a source and destination. The value corresponds to ports used by one or more switches in the path between the source and destination. The destination includes the value in an acknowledgement packet. Logic circuits in the source device track packets and corresponding values. When a status indicates a particular packet has not received an acknowledgement, the value for the packet may be removed from the set of values. Particular ports that may be congested or down may be detected and the packets re-routed using the logic circuits in the source device.

Abstract: Systems and methods for sending and receiving messages, including training data, using a multi-path packet spraying protocol are described. A method includes segmenting a message into a set of data packets comprising training data. The method further includes initiating transmission of the set of data packets to a receiving node. The method further includes spraying the set of data packets across the switch fabric in accordance with the multi-path spraying protocol such that depending upon a value of a fabric determination field associated with a respective data packet, the respective data packet can traverse via any one of a plurality of paths offered by the switch fabric for a connection between the sending node to the receiving node. The method further includes initiating transmission of synchronization packets to the receiving node, where unlike the set of data packets, the synchronization packets are not sprayed across the switch fabric.

Abstract: Method and apparatus to perform cyclic redundancy check computations for error detection are described wherein a first stage includes a first set of computation elements, a first multiplexer and a second multiplexer. A latch is connected to the first stage. A second stage is connected to the latch and the second stage includes a second set of computation elements and a third multiplexer. The first stage and the second stage perform cyclic redundancy check computations for a packet, with the first set of computation elements performing cyclic redundancy check computations for a first set of bytes of input data from the packet, and the second set of computation elements performing cyclic redundancy check computations for a second set of bytes of input data from the packet. Other embodiments are described and claimed.

Abstract: Method and apparatus to perform cyclic redundancy check computations for error detection are described.