Network interface device-based computations

Patent metadata

Field	Value
Publication number	US-12413516-B2
Application number	US-202117515222-A
Country	US
Kind code	B2
Filing date	Oct 29, 2021
Priority date	Oct 29, 2021
Publication date	Sep 9, 2025
Grant date	Sep 9, 2025

Abstract

Examples described herein relate to a switch comprising: circuitry, when operational, to receive a packet comprising a header and a payload and in conjunction with performance of computation on the packet payload, forward the packet header, but not the payload, to a destination endpoint. In some examples, the destination endpoint of the packet is to perform management of reliable transport. In some examples, the circuitry includes programmable data plane circuitry comprising ingress pipeline or egress pipeline and one or more match action units (MAUs) to perform processing of the payload, wherein the programmable data plane circuitry is to perform computation on the packet payload.

First claim

What is claimed is:

1. An apparatus comprising:
a switch comprising:
circuitry, when operational, to:
receive a packet comprising a header and a payload and
in conjunction with performance of computation on the packet payload, forward the packet header, but not the payload, to a destination endpoint;

wherein:
the circuitry comprises one or more match action units (MAUs) to perform one or more actions based upon match table data and the packet header;
the one or more actions comprise (1) the performance of the computation on the packet payload, and (2) insertion of computed results data, based upon the performance of the computation on the packet payload, in packet data to be sent to at least one worker;
the switch comprises circuitry to store a computed result data from the performance of computation on the packet payload and circuitry to forward at least one packet with the computed result data to a compute node that executes a worker; and
the circuitry to store a computed result data from the performance of computation on the packet payload and the circuitry to forward at least one packet with the computed result data to a compute node that executes a worker are to insert the computed result data in at least one packet received from a server and forward the at least one packet received from the server with the inserted computed result data.

2. The apparatus of claim 1 , wherein the destination endpoint of the packet is to perform management of reliable transport.

3. The apparatus of claim 1 , further comprising:
programmable data plane circuitry comprising ingress pipeline or egress pipeline and the one or more match action units (MAUs) to perform processing of the payload, wherein the programmable data plane circuitry is to perform computation on the packet payload.

4. The apparatus of claim 1 , wherein the computation on the packet payload comprises one or more of: aggregation operation in a reduce-scatter phase, write operation in an Allgather phase, or read operation in an Allgather phase and wherein the aggregation operation comprises one or more of summation or min-max.

5. The apparatus of claim 1 , wherein the packet is to identify a total number of workers in a work group in the header and the switch is to provide processed data to the destination endpoint after the total number of workers have sent data to the switch.

6. The apparatus of claim 1 , wherein the switch comprises:
circuitry to
receive a packet comprising a computed result data and to store the computed result data,
form a second packet based on the stored computed result data and at least one received header, and
cause transmission of the second packet.

7. The apparatus of claim 1 , wherein the switch comprises one or more of: network interface controller (NIC), SmartNIC, router, forwarding element, infrastructure processing unit (IPU), or data processing unit (DPU).

8. The apparatus of claim 1 , comprising a server to execute a parameter server that is to receive computed result data from the performance of computation on the packet payload and perform connection maintenance and management with at least one compute node that executes a worker.

9. The apparatus of claim 8 , comprising a data center that includes the at least one compute node that executes a worker and the parameter server, wherein the at least one compute node is to send data to the parameter server via the switch for neural network training and/or inference operations.

10. A non-transitory computer-readable medium comprising instructions stored thereon, that if executed by one or more processors, cause:
configuration of a network interface device to:
based on receipt of a packet from a worker, perform packet payload processing and offload transport header processing and connection maintenance to a server, wherein the packet indicates a worker that contributes data for processing;

wherein:
the network interface device comprises one or more match action units (MAUs) to perform one or more actions based upon match table data and packet header data;
the one or more actions comprise (1) performing, at least in part, the packet payload processing, and (2) insertion of computed results data, based upon the packet payload processing, in packet data to be sent to at least one worker; and
the packet identifies a total number of workers in a work group in a header and the instructions cause the network interface device to be configured to provide processed data to the server after the total number of workers have sent data to the network interface device.

11. The non-transitory computer-readable medium of claim 10 , wherein the packet payload processing comprises one or more of: aggregation operation in a reduce-scatter phase, write operation in an Allgather phase, or read operation in an Allgather phase.

12. The non-transitory computer-readable medium of claim 10 , wherein the network interface device comprises circuitry to store the computed results data and to forward at least one packet to a compute node that executes a worker with the computed results data.

13. The non-transitory computer-readable medium of claim 10 , comprising instructions stored thereon, that if executed by one or more processors, cause:
configuration of the network interface device to store the computed results data and to forward at least one packet to a compute node that executes a worker with the computed results data by insertion of the computed results data in a packet with no payload.

14. A method comprising:
performing packet payload processing at a network interface device while offloading network connection maintenance to a server and
forwarding packets related to network connection maintenance from the server to one or more compute nodes that execute workers;
wherein:
the network interface device comprises one or more match action units (MAUs) to perform one or more actions based upon match table data and packet header data;
the one or more actions comprise (1) performing, at least in part, the packet payload processing, and (2) insertion of computed results data, based upon the packet payload processing, in packet data to be sent to at least one worker; and
the packet identifies a total number of workers in a work group in a header and the method further comprises:
the network interface device providing processed data to the server after the total number of workers have sent data to the network interface device.

15. The method of claim 14 , wherein the packet payload processing comprises one or more of: aggregation operation in a reduce-scatter phase, write operation in an Allgather phase, or read operation in an Allgather phase and wherein the aggregation operation comprises one or more of summation or min-max.

16. The method of claim 14 , comprising:
at the network interface device, storing the computed results data and
forward at least one packet to a compute node that executes a worker with the computed results data.

17. The method of claim 14 , wherein the network interface device comprises one or more of: network interface controller (NIC), SmartNIC, router, switch, forwarding element, infrastructure processing unit (IPU), or data processing unit (DPU).

Assignees

• Intel Corp

Inventors

• Le Yanfang
• Alvarez Daniel A
• Sapio Amedeo
• Fingerhut John Andrew

Classifications

• H04L45/38

  Flow based routing · CPC title

• H04L45/74 Primary

  Address processing for routing · CPC title