Performing an allreduce operation on a plurality of compute nodes of a parallel computer

1. A method of performing an allreduce operation on a plurality of compute nodes of a parallel computer, each compute node comprising at least two processing cores, each processing core having contribution data for the allreduce operation, the method comprising:

establishing one or more logical rings among the compute nodes, each logical ring including at least one processing core from each compute node, wherein in a logical ring, all data that is transferred between processing cores travels from one core of one node to a next core in a next node in a circular manner;

performing, for each logical ring, a global allreduce operation using the contribution data for the processing cores included in that logical ring, yielding a global allreduce result for each processing core included in that logical ring; and

performing, for each compute node, a local allreduce operation using the global allreduce results for each processing core on that compute node.

2. The method of claim 1 wherein performing, for each logical ring, a global allreduce operation using the contribution data for the processing cores included in that logical ring, yielding a global allreduce result for each processing core included in that logical ring further comprises:

performing, for each logical ring, a reduce scatter operation using the global allreduce result for each processing core included in that logical ring, yielding reduce scatter result for each processing core; and

performing, for each logical ring, an allgather operation using the reduce scatter result for each processing core included in that logical ring, yielding the global allreduce result for each processing core.

3. The method of claim 1 further comprising assigning, on each compute node, the processing cores on that compute node to one of a plurality of logical ring domains.

4. The method of claim 3 wherein:

establishing one or more logical rings among the compute nodes, each logical ring including at least one processing core from each compute node further comprises selecting, for each logical ring domain, one of the processing cores assigned to that logical ring domain as the representative core for that logical ring domain; and

performing, for each logical ring, a global allreduce operation using the contribution data for the processing cores included in that logical ring, yielding a global allreduce result for each processing core included in that logical ring further comprises: performing, for each logical ring, a reduce scatter operation using the contribution data for each representative core included in that logical ring, yielding a reduce scatter result for each representative core, and performing, for each logical ring, an allgather operation using the reduce scatter result for each representative core included in that logical ring, yielding the global allreduce result for each representative core; and

performing, for each compute node, a local allreduce operation using the global allreduce results for each processing core on that compute node further comprises performing, for each logical ring domain for each compute node, the local allreduce operation using the global allreduce result for each processing core assigned to that logical ring domain.

5. The method of claim 3 wherein:

each logical ring domain corresponds to one of the logical rings;

establishing one or more logical rings among the compute nodes, each logical ring including at least one processing core from each compute node further comprises assigning, for each logical ring domain for each compute node, all of the processing cores of that logical ring domain to the logical ring corresponding to that logical ring domain;

performing, for each logical ring, a global allreduce operation using the contribution data for the processing cores included in that logical ring, yielding a global allreduce result for each processing core included in that logical ring further comprises: performing, for each logical ring, a reduce scatter operation using the contribution data for each processing core included in that logical ring, yielding a reduce scatter result for each processing core, and performing, for each logical ring, an allgather operation using the reduce scatter result for each processing core included in that logical ring, yielding the global allreduce result for each processing core; and

performing, for each compute node, a local allreduce operation using the global allreduce results for each processing core on that compute node further comprises performing, for each compute node, the local allreduce operation using the global allreduce result for processing cores assigned to different logical ring domains for that compute node.

6. The method of claim 1 wherein the compute nodes are connected using a data communications network having multiple dimensions and each logical ring corresponds to one of the dimensions of the data communications network.

7. A parallel computer for performing an allreduce operation on a plurality of compute nodes, each compute node comprising at least two processing cores, each processing core having contribution data for the allreduce operation, the parallel computer comprising computer memory operatively coupled to the processing cores of the parallel computer, the computer memory having disposed within it computer program instructions capable of:

establishing one or more logical rings among the compute nodes, each logical ring including at least one processing core from each compute node, wherein in a logical ring, all data that is transferred between processing cores travels from one core of one node to a next core in a next node in a circular manner;

performing, for each logical ring, a global allreduce operation using the contribution data for the processing cores included in that logical ring, yielding a global allreduce result for each processing core included in that logical ring; and

performing, for each compute node, a local allreduce operation using the global allreduce results for each processing core on that compute node.

8. The parallel computer of claim 7 wherein performing, for each logical ring, a global allreduce operation using the contribution data for the processing cores included in that logical ring, yielding a global allreduce result for each processing core included in that logical ring further comprises:

performing, for each logical ring, a reduce scatter operation using the global allreduce result for each processing core included in that logical ring, yielding reduce scatter result for each processing core; and

performing, for each logical ring, an allgather operation using the reduce scatter result for each processing core included in that logical ring, yielding the global allreduce result for each processing core.

9. The parallel computer of claim 7 wherein the computer memory has disposed within it computer program instructions capable of assigning, on each compute node, the processing cores on that compute node to one of a plurality of logical ring domains.

10. The parallel computer of claim 9 wherein:

establishing one or more logical rings among the compute nodes, each logical ring including at least one processing core from each compute node further comprises selecting, for each logical ring domain, one of the processing cores assigned to that logical ring domain as the representative core for that logical ring domain; and

performing, for each logical ring, a global allreduce operation using the contribution data for the processing cores included in that logical ring, yielding a global allreduce result for each processing core included in that logical ring further comprises: performing, for each logical ring, a reduce scatter operation using the contribution data for each representative core included in that logical ring, yielding a reduce scatter result for each representative core, and performing, for each logical ring, an allgather operation using the reduce scatter result for each representative core included in that logical ring, yielding the global allreduce result for each representative core; and

performing, for each compute node, a local allreduce operation using the global allreduce results for each processing core on that compute node further comprises performing, for each logical ring domain for each compute node, the local allreduce operation using the global allreduce result for each processing core assigned to that logical ring domain.

11. The parallel computer of claim 9 wherein:

each logical ring domain corresponds to one of the logical rings;

establishing one or more logical rings among the compute nodes, each logical ring including at least one processing core from each compute node further comprises assigning, for each logical ring domain for each compute node, all of the processing cores of that logical ring domain to the logical ring corresponding to that logical ring domain;

performing, for each logical ring, a global allreduce operation using the contribution data for the processing cores included in that logical ring, yielding a global allreduce result for each processing core included in that logical ring further comprises: performing, for each logical ring, a reduce scatter operation using the contribution data for each processing core included in that logical ring, yielding a reduce scatter result for each processing core, and performing, for each logical ring, an allgather operation using the reduce scatter result for each processing core included in that logical ring, yielding the global allreduce result for each processing core; and

performing, for each compute node, a local allreduce operation using the global allreduce results for each processing core on that compute node further comprises performing, for each compute node, the local allreduce operation using the global allreduce result for processing cores assigned to different logical ring domains for that compute node.

12. The parallel computer of claim 7 wherein the compute nodes are connected using a data communications network having multiple dimensions and each logical ring corresponds to one of the dimensions of the data communications network.

13. A computer program product for performing an allreduce operation on a plurality of compute nodes of a parallel computer, each compute node comprising at least two processing cores, each processing core having contribution data for the allreduce operation, the computer program product disposed upon a computer readable recordable medium, the computer program product comprising computer program instructions capable of:

establishing one or more logical rings among the compute nodes, each logical ring including at least one processing core from each compute node, wherein in a logical ring, all data that is transferred between processing cores travels from one core of one node to a next core in a next node in a circular manner;

performing, for each logical ring, a global allreduce operation using the contribution data for the processing cores included in that logical ring, yielding a global allreduce result for each processing core included in that logical ring; and

performing, for each compute node, a local allreduce operation using the global allreduce results for each processing core on that compute node.

14. The computer program product of claim 13 wherein performing, for each logical ring, a global allreduce operation using the contribution data for the processing cores included in that logical ring, yielding a global allreduce result for each processing core included in that logical ring further comprises:

performing, for each logical ring, a reduce scatter operation using the global allreduce result for each processing core included in that logical ring, yielding reduce scatter result for each processing core; and

performing, for each logical ring, an allgather operation using the reduce scatter result for each processing core included in that logical ring, yielding the global allreduce result for each processing core.

15. The computer program product of claim 13 further comprising computer program instructions capable of assigning, on each compute node, the processing cores on that compute node to one of a plurality of logical ring domains.

16. The computer program product of claim 15 wherein:

establishing one or more logical rings among the compute nodes, each logical ring including at least one processing core from each compute node further comprises selecting, for each logical ring domain, one of the processing cores assigned to that logical ring domain as the representative core for that logical ring domain; and

performing, for each logical ring, a global allreduce operation using the contribution data for the processing cores included in that logical ring, yielding a global allreduce result for each processing core included in that logical ring further comprises: performing, for each logical ring, a reduce scatter operation using the contribution data for each representative core included in that logical ring, yielding a reduce scatter result for each representative core, and performing, for each logical ring, an allgather operation using the reduce scatter result for each representative core included in that logical ring, yielding the global allreduce result for each representative core; and

performing, for each compute node, a local allreduce operation using the global allreduce results for each processing core on that compute node further comprises performing, for each logical ring domain for each compute node, the local allreduce operation using the global allreduce result for each processing core assigned to that logical ring domain.

17. The computer program product of claim 15 wherein:

each logical ring domain corresponds to one of the logical rings;

establishing one or more logical rings among the compute nodes, each logical ring including at least one processing core from each compute node further comprises assigning, for each logical ring domain for each compute node, all of the processing cores of that logical ring domain to the logical ring corresponding to that logical ring domain;

performing, for each logical ring, a global allreduce operation using the contribution data for the processing cores included in that logical ring, yielding a global allreduce result for each processing core included in that logical ring further comprises: performing, for each logical ring, a reduce scatter operation using the contribution data for each processing core included in that logical ring, yielding a reduce scatter result for each processing core, and performing, for each logical ring, an allgather operation using the reduce scatter result for each processing core included in that logical ring, yielding the global allreduce result for each processing core; and

performing, for each compute node, a local allreduce operation using the global allreduce results for each processing core on that compute node further comprises performing, for each compute node, the local allreduce operation using the global allreduce result for processing cores assigned to different logical ring domains for that compute node.

18. The computer program product of claim 13 wherein the compute nodes are connected using a data communications network having multiple dimensions and each logical ring corresponds to one of the dimensions of the data communications network.