Path selection for packet transmission

Network switch apparatus to be communicatively coupled via network links to at least one network, the network switch apparatus being configurable for use in association with at least one graphics processing unit (GPU) and at least one programmable multi-stage packet data processing pipeline, the network links to be associated with network paths in the at least one network, the at least one network being configurable to include at least one fabric and at least one leaf/spine infrastructure, the network switch apparatus comprising:

• at least one network data processor for:
• communicating remote direct memory access over converged Ethernet (RoCE) packet data via the network links;
• scheduling egress ROCE packets for transmission via multiple of the network links, the egress ROCE packets to be associated with portions of egress ROCE packet data;
• and
• selecting the multiple of the network links for the transmission for purpose of evenly distributing egress ROCE packet data bandwidth across the multiple of the network paths;
• wherein:
• the portions of the egress ROCE packet data are associated with a particular flow;
• respective allocations of the portions of the egress ROCE packet data to the multiple of the network paths associated with the multiple of the network links are to be based upon respective of the egress RoCE packets;
• the at least one programmable multi-stage packet data processing pipeline is configurable to:
• be used in ROCE packet data processing;
• and
• comprise at least one parser and at least one match-action stage;
• the network switch apparatus comprises ports;
• the ports are associated with the network links;
• the network switch apparatus is configurable for use in:
• the at least one leaf/spine infrastructure;
• and
• communication via the at least one fabric;
• the at least one network data processor is configurable to select, based upon congestion-related data, an equal-cost multipath (ECMP) load balancing path;
• the ECMP load balancing path is associated with the multiple of the network paths;
• and
• the congestion-related data is to be generated based upon in-band telemetry data.

At least one machine-readable medium storing instructions to be executed by a network switch apparatus, the network switch apparatus to be communicatively coupled via network links to at least one network, the network switch apparatus being configurable for use in association with at least one graphics processing unit (GPU) and at least one programmable multi-stage packet data processing pipeline, the network links to be associated with network paths in the at least one network, the at least one network being configurable to include at least one fabric and at least one leaf/spine infrastructure, the instructions, when executed by the network switch apparatus, resulting in the network switch apparatus being configured to enable performance of operations comprising:

• communicating, via the network switch apparatus, remote direct memory access over converged Ethernet (RoCE) packet data via the network links;
• and
• based upon sizes of portions of egress ROCE packet data to be allocated to multiple of the network paths associated with multiple of the network links:
• scheduling, by the network switch apparatus, egress ROCE packets for transmission via the multiple of the network links, the egress ROCE packets to be associated with the portions of the egress ROCE packet data;
• and
• selecting, by the network switch apparatus, the multiple of the network links for the transmission for purpose of evenly distributing egress ROCE packet data bandwidth across the multiple of the network paths;
• wherein:
• the portions of the egress RoCE packet data are associated with a particular flow;
• respective allocations of the sizes of the portions of the egress ROCE packet data to the multiple of the network paths associated with the multiple of the network links are to be based upon respective of the egress ROCE packets;
• the at least one programmable multi-stage packet data processing pipeline is configurable to:
• be used in ROCE packet data processing;
• and
• comprise at least one parser and at least one match-action stage;
• the network switch apparatus comprises ports;
• the ports are associated with the network links;
• the network switch apparatus is configurable for use in:
• the at least one leaf/spine infrastructure;
• and
• communication via the at least one fabric;
• the network switch apparatus is configurable to select, based upon congestion-related data, an equal-cost multipath (ECMP) load balancing path;
• the ECMP load balancing path is associated with the multiple of the network paths;
• and
• the congestion-related data is to be generated based upon in-band telemetry data.

A method implemented using a network switch apparatus, the network switch apparatus to be communicatively coupled via network links to at least one network, the network switch apparatus being configurable for use in association with at least one graphics processing unit (GPU) and at least one programmable multi-stage packet data processing pipeline, the network links to be associated with network paths in the at least one network, the at least one network being configurable to include at least one fabric and at least one leaf/spine infrastructure, the method comprising:

• communicating, via the network switch apparatus, remote direct memory access over converged Ethernet (RoCE) packet data via the network links;
• scheduling, by the network switch apparatus, egress ROCE packets for transmission via multiple of the network links, the egress ROCE packets to be associated with portions of egress RoCE packet data;
• and
• selecting, by the network switch apparatus, the multiple of the network links for the transmission for purpose of evenly distributing egress ROCE packet data bandwidth across the multiple of the network paths;
• wherein:
• the portions of the egress ROCE packet data are associated with a particular flow;
• respective allocations of the portions of the egress ROCE packet data to the multiple of the network paths associated with the multiple of the network links are to be based upon respective of the egress ROCE packets;
• the at least one programmable multi-stage packet data processing pipeline is configurable to:
• be used in ROCE packet data processing;
• and
• comprise at least one parser and at least one match-action stage;
• the network switch apparatus comprises ports;
• the ports are associated with the network links;
• the network switch apparatus is configurable for use in:
• the at least one leaf/spine infrastructure;
• and
• communication via the at least one fabric;
• the at least one network data processor is configurable to select, based upon congestion-related data, an equal-cost multipath (ECMP) load balancing path;
• the ECMP load balancing path is associated with the multiple of the network paths;
• and
• the congestion-related data is to be generated based upon in-band telemetry data.

At least one data center comprising:

• at least one network comprising network links, at least one fabric, and at least one leaf/spine infrastructure, the network links to be associated with network paths in the at least one network;
• at least one graphics processing unit (GPU);
• at least one programmable multi-stage packet data processing pipeline;
• and
• a network switch apparatus to be communicatively coupled via the network links to the at least one network, the network switch apparatus to be used in association with the at least one graphics processing unit (GPU) and the at least one programmable multi-stage packet data processing pipeline, the network switch apparatus to:
• communicate remote direct memory access over converged Ethernet (RoCE) packet data via the network links;
• and
• based upon sizes of portions of egress ROCE packet data to be allocated to multiple of the network paths associated with multiple of the network links:
• schedule egress ROCE packets for transmission via the multiple of the network links, the egress ROCE packets to be associated with the portions of the egress ROCE packet data;
• and
• select the multiple of the network links for the transmission for purpose of evenly distributing egress ROCE packet data bandwidth across the multiple of the network paths;
• wherein:
• the portions of the egress ROCE packet data are associated with a particular flow;
• respective allocations of the sizes of the portions of the egress ROCE packet data to the multiple of the network paths associated with the multiple of the network links are to be based upon respective of the egress ROCE packets;
• the at least one programmable multi-stage packet data processing pipeline is configurable to:
• be used in ROCE packet data processing;
• and
• comprise at least one parser and at least one match-action stage;
• the network switch apparatus comprises ports;
• the ports are associated with the network links;
• the network switch apparatus is configurable for use in:
• the at least one leaf/spine infrastructure;
• and
• communication via the at least one fabric;
• the network switch apparatus is configurable to select, based upon congestion-related data, an equal-cost multipath (ECMP) load balancing path;
• the ECMP load balancing path is associated with the multiple of the network paths;
• and
• the congestion-related data is to be generated based upon in-band telemetry data.