Recording A Communication Pattern and Replaying Messages in a Parallel Computing System

• US 20110010471A1
• Filed: 07/10/2009
• Published: 01/13/2011
• Est. Priority Date: 07/10/2009
• Status: Active Grant

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Heidelberger, Philip, Kumar, Sameer
• Granted Patent: US 8,407,376 B2
• Time in Patent Office: Days
• US Class Current: 710/22
• CPC Class Codes: G06F 13/28 using burst mode transfer, e.g. direct memory access DMA, cycle steal G06F13/32 takes precedence

Abstract

A parallel computer system includes a plurality of compute nodes. Each of the compute nodes includes at least one processor, at least one memory, and a direct memory address engine coupled to the at least one processor and the at least one memory. The system also includes a network interconnecting the plurality of compute nodes. The network operates a global message-passing application for performing communications across the network. Local instances of the global message-passing application operate at each of the compute nodes to carry out local processing operations independent of processing operations carried out at another one of the compute nodes. The direct memory address engines are configured to interact with the local instances of the global message-passing application via injection FIFO metadata describing an injection FIFO in a corresponding one of the memories. The local instances of the global message passing application are configured to record, in the injection FIFO in the corresponding one of the memories, message descriptors associated with messages of an arbitrary communication pattern in an iteration of an executing application program. The local instances of the global message passing application are configured to replay the message descriptors during a subsequent iteration of the executing application program.

Claims

1. A parallel computer system comprising:

a plurality of compute nodes, each of said compute nodes comprising;

at least one processor;

at least one memory; and

a direct memory address engine coupled to said at least one processor and said at least one memory; and

a network interconnecting said plurality of compute nodes;

wherein;

said network operates a global message-passing application for performing communications across said network;

local instances of said global message-passing application operate at each of said compute nodes to carry out local processing operations independent of processing operations carried out at another one of said compute nodes;

said direct memory address engines are configured to interact with said local instances of said global message-passing application via injection FIFO metadata describing an injection FIFO in a corresponding one of said memories;

said local instances of said global message passing application are configured to record, in said injection FIFO in said corresponding one of said memories, message descriptors associated with messages of an arbitrary communication pattern in an iteration of an executing application program; and

said local instances of said global message passing application are configured to replay said message descriptors during a subsequent iteration of said executing application program.

2. The parallel computer system of claim 1, wherein said message descriptors comprise a destination, a source, and a message intent.

3. The parallel computer system of claim 2, wherein said direct memory address engines are configured to assemble said messages in packet form by obtaining data from said memory in accordance with said message descriptors.

4. The parallel computer system of claim 1, wherein a given one of said local instances of said global message passing application records said message descriptors in response to a record command in said executing application program.

5. The parallel computer system of claim 1, wherein a given one of said local instances of said global message passing application replays said message descriptors in response to a replay command in said executing application program.

6. The parallel computer system of claim 1, wherein a given one of said local instances of said global message passing application is configured to automatically record said message descriptors by intercepting persistent application program interface calls of said executing application program.

7. The parallel computer system of claim 1, wherein said injection FIFO metadata comprises head and tail pointers to said message descriptors in said injection FIFO.

8. The parallel computer system of claim 7, wherein a given one of said local instances of said global message passing application is further configured to replay said message descriptors by resetting said head and tail pointers.

9. The parallel computer system of claim 1, wherein said message descriptors specify point-to-point communications.

10. The parallel computer system of claim 1, wherein said message descriptors specify collective communications.

11. The parallel computer system of claim 1, wherein:

said message descriptors comprise first message descriptors;

said arbitrary communication pattern comprises a first arbitrary communication pattern represented by a first persistent identifier;

said local instances of said global message passing application are further configured to record, in said injection FIFO in said corresponding one of said memories, second message descriptors associated with messages of a second arbitrary communication pattern, represented by a second persistent identifier, in another iteration of said executing application program; and

said local instances of said global message passing application are still further configured to replay said second message descriptors during another subsequent iteration of said executing application program, by resetting said injection FIFO metadata.

12. The parallel computer system of claim 11, wherein:

said local instances of said global message passing application are further configured to record, in said injection FIFO in said corresponding one of said memories, sufficient additional message descriptors associated with messages of sufficient additional arbitrary communication patterns, represented by sufficient additional persistent identifiers, in additional iterations of said executing application program, for said executing application program to implement a three dimensional fast Fourier transform operation; and

said local instances of said global message passing application are still further configured to replay said additional message descriptors during other subsequent iterations of said executing application program, by resetting said injection FIFO metadata.

13. The parallel computer system of claim 11, wherein:

said local instances of said global message passing application are further configured to record, in said injection FIFO in said corresponding one of said memories, sufficient additional message descriptors associated with messages of sufficient additional arbitrary communication patterns, represented by sufficient additional persistent identifiers, in additional iterations of said executing application program, for at least one thousand total arbitrary communication patterns; and

said local instances of said global message passing application are still further configured to replay said additional message descriptors during other subsequent iterations of said executing application program, by resetting said injection FIFO metadata.

14. A method comprising the steps of

providing a parallel computer system comprising:

a plurality of compute nodes, each of said compute nodes comprising;

at least one processor;

at least one memory; and

a direct memory address engine coupled to said at least one processor and said at least one memory; and

a network interconnecting said plurality of compute nodes;

wherein;

said network operates a global message-passing application for performing communications across said network;

local instances of said global message-passing application operate at each of said compute nodes to carry out local processing operations independent of processing operations carried out at another one of said compute nodes; and

said direct memory address engines are configured to interact with said local instances of said global message-passing application via injection FIFO metadata describing an injection FIFO in a corresponding one of said memories;

recording, with said local instances of said global message passing application, in said injection FIFO in said corresponding one of said memories, message descriptors associated with messages of an arbitrary communication pattern in an iteration of an executing application program; and

replaying said message descriptors, with said local instances of said global message passing application, during a subsequent iteration of said executing application program.

15. The method of claim 14, wherein:

said message descriptors recorded in said recording step comprise a destination, a source, and a message intent; and

said direct memory address engines provided in said providing step are configured to assemble said messages in packet form by obtaining data from said memory in accordance with said message descriptors.

16. The method of claim 14, further comprising receiving a record command from said executing application program, wherein a given one of said local instances of said global message passing application records said message descriptors in response to said record command in said executing application program.

17. The method of claim 14, further comprising receiving a replay command from said executing application program, wherein a given one of said local instances of said global message passing application replays said message descriptors in response to said replay command in said executing application program.

18. The method of claim 14, wherein a given one of said local instances of said global message passing application records said message descriptors automatically by intercepting persistent application program interface calls of said executing application program.

19. The method of claim 14, wherein said message descriptors recorded in said recording step specify point-to-point communications.

20. The method of claim 14, wherein said message descriptors recorded in said recording step specify collective communications.

21. The method of claim 14, wherein:

said message descriptors comprise first message descriptors; and

said arbitrary communication pattern comprises a first arbitrary communication pattern represented by a first persistent identifier;

further comprising;

recording, with a given one of said local instances of said global message passing application, in said injection FIFO in said corresponding one of said memories, second message descriptors associated with messages of a second arbitrary communication pattern, represented by a second persistent identifier, in another iteration of said executing application program; and

replaying, with said given one of said local instances of said global message passing application, said second message descriptors during another subsequent iteration of said executing application program, by resetting said injection FIFO metadata.

22. The method of claim 21, further comprising:

with said local instances of said global message passing application, recording, in said injection FIFO in said corresponding one of said memories, sufficient additional message descriptors associated with messages of sufficient additional arbitrary communication patterns, represented by sufficient additional persistent identifiers, in additional iterations of said executing application program, for said executing application program to implement a three dimensional fast Fourier transform operation; and

with said local instances of said global message passing application, replaying said additional message descriptors during other subsequent iterations of said executing application program, by resetting said injection FIFO metadata.

23. The method of claim 21, further comprising:

with said local instances of said global message passing application, recording, in said injection FIFO in said corresponding one of said memories, sufficient additional message descriptors associated with messages of sufficient additional arbitrary communication patterns, represented by sufficient additional persistent identifiers, in additional iterations of said executing application program, for at least one thousand total arbitrary communication patterns; and

with said local instances of said global message passing application, replaying said additional message descriptors during other subsequent iterations of said executing application program, by resetting said injection FIFO metadata.

24. A computer program product comprising a computer readable storage medium having computer readable program code embodied therewith, the computer readable program code comprising:

computer readable program code configured to implement a global message-passing application for performing communications across the network, with local instances of said global message-passing application operating at each of the compute nodes to carry out local processing operations independent of processing operations carried out at another one of the compute nodes;

computer readable program code configured to facilitate interaction between said direct memory address engines and said local instances of said global message-passing application via injection FIFO metadata describing an injection FIFO in a corresponding one of said memories;

computer readable program code configured to facilitate recording message descriptors associated with messages of an arbitrary communication pattern in an iteration of an executing application program; and

computer readable program code configured to facilitate replaying said message descriptors during a subsequent iteration of said executing application program.

25. An apparatus comprising:

a plurality of compute nodes, each of said compute nodes comprising;

at least one processor;

at least one memory; and

a direct memory address engine coupled to said at least one processor and said at least one memory;

means for interconnecting said plurality of compute nodes;

means for global message-passing communications across said means for interconnecting, said means for global message passing operating at each of said compute nodes to carry out local processing operations independent of processing operations carried out at another one of said compute nodes; and

means for facilitating interaction between said direct memory address engines and said means for global message-passing;

means for recording message descriptors associated with messages of an arbitrary communication pattern in an iteration of an executing application program; and

means for replaying said message descriptors during a subsequent iteration of said executing application program.