System and Method for Performing Collective Operations Using Software Setup and Partial Software Execution at Leaf Nodes in a Multi-Tiered Full-Graph Interconnect Architecture

• US 20090063816A1
• Filed: 08/27/2007
• Published: 03/05/2009
• Est. Priority Date: 08/27/2007
• Status: Active Grant

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Arimilli, Ravi K., Rajamony, Ramakrishnan, Arimilli, Lakshminarayana B., Speight, William E.
• Granted Patent: US 7,958,183 B2
• Time in Patent Office: Days
• US Class Current: 712/30
• CPC Class Codes: G06F 15/17381 Two dimensional, e.g. mesh, torus

Abstract

A method, computer program product, and system are provided for performing collective operations. In software executing on a parent processor in a first processor book, a number of other processors are determined in a same or different processor book of the data processing system that is needed to execute the collective operation, thereby establishing a plurality of processors comprising the parent processor and the other processors. In software executing on the parent processor, the plurality of processors are logically arranged as a plurality of nodes in a hierarchical structure. The collective operation is transmitted to the plurality of processors based on the hierarchical structure. In hardware of the parent processor, results are received from the execution of the collective operation from the other processors, a final result is generated of the collective operation based on the received results, and the final result is output.

Claims

1. A method, in a data processing system, for performing collective operations, the data processing system comprising a plurality of supernodes, the plurality of supernodes comprising a plurality of processor books, and the plurality of processor books comprising a plurality of processors, the method comprising:

determining, in software executing on a parent processor in a first processor book of the data processing system, a number of other processors in a same or different processor book of the data processing system needed to execute the collective operation, thereby establishing a subset of processors comprising the parent processor and the other processors;

logically arranging, in the software executing on the parent processor, the subset of processors as a plurality of nodes in a hierarchical structure;

transmitting the collective operation to the subset of processors based on the hierarchical structure;

receiving, in hardware of the parent processor, results from the execution of the collective operation from the other processors;

generating, in hardware of the parent processor, a final result of the collective operation based on the results received from execution of the collective operation by the other processors; and

outputting the final result.

2. The method of claim 1, wherein outputting the final result comprises transmitting the final result to each of the other processors in the subset of processors.

3. The method of claim 1, wherein the parent processor and at least one of the other processors are in different processor books of the data processing system.

4. The method of claim 3, wherein the parent processor and the at least one of the other processors are in different supernodes of the data processing system.

5. The method of claim 3, wherein the parent processor and the at least one of the other processors are in a same supernode of the data processing system.

6. The method of claim 1, wherein the collective operation is at least one of reduce, multicast, all to one, barrier, or all reduce collective operation.

7. The method of claim 6, wherein the collective operation comprises at least one operand and wherein the operand is at least one of add, min, max, min loc, max loc, or no-op.

8. The method of claim 1, wherein the method is implemented by a collective acceleration unit integrated into each of the plurality of processors and wherein only the processors in the subset of processors utilize the integrated collective acceleration unit to perform the collective operations.

9. A computer program product, for performing collective operations, comprising a computer useable medium having a computer readable program, wherein the computer readable program, when executed in a parent processor in a first processor book of a data processing system, causes the parent processor to:

determining, in software executing on the parent processor, a number of other processors in a same or different processor book of the data processing system needed to execute the collective operation, thereby establishing a subset of processors comprising the parent processor and the other processors;

logically arranging, in the software executing on the parent processor, the subset of processors as a plurality of nodes in a hierarchical structure;

transmitting the collective operation to the subset of processors based on the hierarchical structure;

receiving, in hardware of the parent processor, results from the execution of the collective operation from the other processors;

generating, in hardware of the parent processor, a final result of the collective operation based on the results received from execution of the collective operation by the other processors; and

outputting the final result, wherein the data processing system comprises a plurality of supernodes, the plurality of supernodes comprising a plurality of processor books, and the plurality of processor books comprising a plurality of processors.

10. The computer program product of claim 9, wherein outputting the final result comprises transmitting the final result to each of the other processors in the subset of processors.

11. The computer program product of claim 9, wherein the parent processor and at least one of the other processors are in different processor books of the data processing system.

12. The computer program product of claim 11, wherein the parent processor and the at least one of the other processors are in different supernodes of the data processing system.

13. The computer program product of claim 11, wherein the parent processor and the at least one of the other processors are in a same supernode of the data processing system.

14. The computer program product of claim 9, wherein the collective operation is at least one of reduce, multicast, all to one, barrier, or all reduce collective operation, and wherein the collective operation comprises at least one operand and wherein the operand is at least one of add, min, max, min loc, max loc, or no-op.

15. A data processing system for performing collective operations, comprising:

a parent processor in a first processor book of the data processing system; and

a memory coupled to the parent processor, wherein the memory comprises instructions which, when executed by the parent processor, cause the parent processor to;

determining, in software executing on a parent processor, a number of other processors in a same or different processor book of the data processing system needed to execute the collective operation, thereby establishing a subset of processors comprising the parent processor and the other processors;

logically arranging, in the software executing on the parent processor, the subset of processors as a plurality of nodes in a hierarchical structure;

transmitting the collective operation to the subset of processors based on the hierarchical structure;

receiving, in hardware of the parent processor, results from the execution of the collective operation from the other processors;

generating, in hardware of the parent processor, a final result of the collective operation based on the results received from execution of the collective operation by the other processors; and

outputting the final result, wherein the data processing system comprises a plurality of supernodes, the plurality of sup emodes comprising a plurality of processor books, and the plurality of processor books comprising a plurality of processors.

16. The system of claim 15, wherein outputting the final result comprises transmitting the final result to each of the other processors in the subset of processors.

17. The system of claim 15, wherein the parent processor and at least one of the other processors are in different processor books of the data processing system.

18. The system of claim 17, wherein the parent processor and the at least one of the other processors are in different supernodes of the data processing system.

19. The system of claim 17, wherein the parent processor and the at least one of the other processors are in a same supernode of the data processing system.

20. The system of claim 15, wherein the collective operation is at least one of reduce, multicast, all to one, barrier, or all reduce collective operation, and wherein the collective operation comprises at least one operand, and wherein the operand is at least one of add, min, max, min loc, max loc, or no-op.