Performing synchronized collective operations over multiple process groups

• US 9,665,531 B2
• Filed: 06/13/2012
• Issued: 05/30/2017
• Est. Priority Date: 06/13/2012
• Status: Active Grant

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Jain, Nikhil, Sabharwal, Yogish, George, Thomas, Mittal, Anshul, Kumar, Sameer
• Primary Examiner(s): ABU ROUMI, MAHRAN Y
• Application Number:
  US13/495,647
• Publication Number:
  US 20130339499A1
• Time in Patent Office: 1,812 Days
• Field of Search: 370337, 709217
• CPC Class Codes: G06F 15/17318 Parallel communications techniques, e.g. gather, scatter, reduce, roadcast, multicast, all to all

Abstract

Methods and arrangements for performing synchronized collective operations. Communication calls are accepted from at least two distinct processor groups. Edge disjoint spanning paths are created over a collective comprising the processor groups, and the spanning paths are assigned to the processor groups to facilitate communication within each processor group.

Claims

1. An apparatus comprising:

at least one processor; and

a computer readable non-signal storage medium having code stored therewith, the code being executable by the at least one processor and comprising;

code that performs synchronized collective communication operations across a plurality of processing groups simultaneously, wherein the plurality of processing groups are connected together in a network architecture selected from the group consisting of;

2D mesh, 3D torus, and infiniband and wherein each processing group contains a plurality of high performance computing nodes, each of the plurality of nodes occurring in a single processing group, via;

code that identifies a plurality of characteristics of the network architecture by receiving a call from each of the plurality of nodes, wherein the plurality of characteristics comprises identification of the processing groups via identification of a plurality of sub-communicators and nodes from the plurality of nodes which are connected to each of the plurality of sub-communicators;

code that constructs, based upon the characteristics of the network architecture, a plurality of communication-efficient bidirectional edge disjoint spanning paths for the plurality of processing groups;

code that assigns one of the plurality of edge disjoint spanning paths to each processing group, wherein each node is assigned to only one of the plurality of edge disjoint spanning paths and wherein each edge disjoint spanning path comprises only and all nodes assigned to a particular processing group; and

code that performs synchronized collective communication operations across the plurality of processing groups and reducing communication bottlenecks and preventing contention and interference between communications occurring simultaneously across the plurality of processing groups by routing processing communications for each processing group using only the edge disjoint path assigned to the corresponding processing group.

2. The apparatus of claim 1, wherein the plurality of disjoint spanning paths comprise bandwidth-efficient paths.

3. A computer program product comprising:

a computer readable non-signal storage medium having code stored therewith, the code being executable by at least one processor and comprising;

code that performs synchronized collective communication operations across a plurality of processing groups simultaneously, wherein the plurality of processing groups are connected together in a network architecture accepted as input from the group consisting of;

2D mesh, 3D torus, and infiniband and wherein each processing group contains a plurality of high performance computing nodes, each of the plurality of nodes occurring in a single processing group, via;

code that identifies a plurality of characteristics of the network architecture by receiving a call from each of the plurality of nodes, wherein the plurality of characteristics comprises identification of the processing groups via identification of a plurality of sub-communicators and nodes from the plurality of nodes which are connected to each of the plurality of sub-communicators;

code that constructs, based upon the characteristics of the network architecture, a plurality of communication-efficient bidirectional edge disjoint spanning paths for the plurality of processing groups;

code that assigns one of the plurality of edge disjoint spanning paths to each processing group, wherein each node is assigned to only one of the plurality of edge disjoint spanning paths and wherein each edge disjoint spanning path comprises only and all nodes assigned to a particular processing group; and

code that performs synchronized collective communication operations across the plurality of processing groups and reducing communication bottlenecks and preventing contention and interference between communications occurring simultaneously across the plurality of processing groups by routing processing communications for each processing group using only the edge disjoint path assigned to the corresponding processing group.

4. The computer program product according to claim 3, wherein the spanning paths comprise at least one member selected from the group consisting of:

bandwidth-efficient paths;

load-balancing-efficient paths; and

power-efficient paths.

5. The computer program product according to claim 3, wherein said code is configured to receive calls from nodes in the processor groups.

6. The computer program product according to claim 3, wherein said code is configured to communicate with two sub-communicators.

7. The computer program product of claim 3, wherein the plurality of disjoint spanning paths comprise bandwidth-efficient paths.