Method and apparatus for compute end point based collective operations
AssigneeIntel Corporation
InventorsFrancesc GUIM BERNAT, Kshitij A. DOSHI, Daniel RIVAS BARRAGAN, Alejandro DURAN GONZALEZ
Life of this patent
Estimating the term…
Abstract
Examples may include techniques for collective operations in a distributed architecture. A collective operation request message from a computing node causes collective operations at one or more target computing nodes communicatively coupled with the computing node through a network switch. The collective operation request message also causes the network switch to perform collective operations on collective operation results received from the one or more target computing nodes.
Claims22
An apparatus comprising:
- circuitry at a computing node coupled with a plurality of target computing nodes communicatively coupled with the computing node through a network switch, the circuitry to:
- receive a request to obtain data from the plurality of target computing nodes;
- generate a collective operation request message based on the request to obtain the data;
- send the collective operation request message to the network switch, the collective operation request message to include information to cause the network switch to send unicast messages to respective target computing nodes of the plurality of target computing nodes to direct the respective target computing nodes to perform collective operations to generate separate results and to send the separate results to the network switch, the collective operation request message to also include information to direct the network switch to perform a collective operation via use of received separate results to generate a final result and to send the final result to the computing node;
- receive the final result from the network switch;
- and
- provide the final result to the requester to fulfill the request to obtain the data from the plurality of target computing nodes.
An apparatus comprising:
- circuitry at a network switch, the circuitry to:
- receive a collective operation request message from a request computing node communicatively coupled with the network switch, the collective operation request message to include information for sending unicast messages to respective target computing nodes of a plurality of computing nodes communicatively coupled with the request computing node through the network switch, the unicast messages to direct the respective target computing nodes to perform collective operations to generate separate results and to send the separate results to the network switch;
- send the unicast messages to the plurality of computing nodes;
- receive the separate results;
- perform a collective operation via use of the received separate results to generate a final result, the collective operation performed based on additional information included in the collective operation request message;
- and
- send the final result to the request computing node.
15-23. (canceled)
At least one machine readable medium comprising a plurality of instructions that in response to being executed by a system at a computing node cause the system to:
- receive a unicast message from a network switch communicatively coupled to the computing node, the unicast message to include information to perform a collective operation to generate a result via use one or more operands obtained from one or more respective targets located at or accessible to the computing node;
- and
- send the result to the network switch for the network switch to use the result along with one or more other results sent from other computing nodes as part of a separate collective operation to generate a final result, the final result generated responsive to a collective operation request message received by the network switch from a request computing node communicatively coupled to the computing node through the network switch, the collective operation request message sent by the request computing node to the network switch responsive to a request to obtain data from the computing node and the other computing nodes.
Dependencies are inferred from the claim text.