Mechanism to support generic collective communication across a variety of programming models

• US 7,984,448 B2
• Filed: 06/26/2007
• Issued: 07/19/2011
• Est. Priority Date: 06/26/2007
• Status: Expired due to Fees

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Dozsa, Gabor, Almasi, Gheorghe, Kumar, Sameer
• Primary Examiner(s): Nguyen; Van H
• Application Number:
  US11/768,669
• Publication Number:
  US 20090006810A1
• Time in Patent Office: 1,484 Days
• Field of Search: 719/313, 719/328, 718/100, 718/102, 712/35
• US Class Current: 719/313
• CPC Class Codes: G06F 9/54 Interprogram communication

Abstract

A system and method for supporting collective communications on a plurality of processors that use different parallel programming paradigms, in one aspect, may comprise a schedule defining one or more tasks in a collective operation, an executor that executes the task, a multisend module to perform one or more data transfer functions associated with the tasks, and a connection manager that controls one or more connections and identifies an available connection. The multisend module uses the available connection in performing the one or more data transfer functions. A plurality of processors that use different parallel programming paradigms can use a common implementation of the schedule module, the executor module, the connection manager and the multisend module via a language adaptor specific to a parallel programming paradigm implemented on a processor.

Claims

1. A system for supporting collective communications on a plurality of processors that use different parallel programming paradigms, comprising:

a processor;

a schedule module for defining one or more tasks in a collective operation, the collective operation being a function supported at least in a message passing interface (MPI) including at least a reduce operation, all-reduce operation, and all-to-all operation, in which a group of processors in parallel computer system participate to perform a global operation where each of the processors contribute to the global operation;

an executor module for executing said one or more tasks;

a multisend module for performing one or more data transfer functions in said one or more tasks;

a connection manager for controlling one or more connections and identifying an available connection,

the multisend module further for using the available connection in performing the one or more data transfer functions; and

a language adaptor for interfacing at least to the multisend module,

wherein the plurality of processors that use different parallel programming paradigms can use a common implementation of the schedule module, the executor module, the connection manager and the multisend module via the language adaptor specific to a parallel programming paradigm implemented on the processor,

wherein the multisend module sends many messages at the same time, each message or a group of messages identified by a connection identifier.

2. The system of claim 1, wherein the schedule module, the executor module, the connection manager and the multisend module are created on each processor of a multiple processor system, the multiple processor system being connected by a tree network.

3. The system of claim 1, wherein the schedule module, the executor module, the connection manager and the multisend module are created on each processor of a multiple processor system, the multiple processor system being connected by a torous network with deposit bit sends.

4. The system of claim 1, wherein the schedule module defines one or more tasks in the collective operation per each phase of the collective operation.

5. The system of claim 1, wherein the executor module performs the one or more tasks using one or more optimizations.

6. The system of claim 5, wherein said one or more optimizations include pipelining, phase independence, or multi-color routes or combinations thereof.

7. The system of claim 1, wherein the multisend module performs data transfer using point-to-point communication.

8. The system of claim 1, wherein the multisend module is invoked for each phase of the collective operation.

9. The system of claim 1, wherein the executor module is created for each collective operation being performed.

10. A method of supporting collective communications on a plurality of processors that use different parallel programming paradigms, comprising:

receiving an application call invoking a collective operation from a processor of the plurality of processors via a language adaptor, the collective operation being a function supported at least in a message passing interface (MPI) including at least a reduce operation, all-reduce operation, and all-to-all operation, in which a group of processors in parallel computer system participate to perform a global operation where each of the processors contribute to the global operation;

creating a schedule module for defining one or more tasks in the collective operation;

creating an executor module for executing said one or more tasks;

invoking a multisend module to perform one or more data transfer functions associated with said one or more tasks,

creating a connection manager to control one or more connections and identifying an available connection; and

the multisend module further uses the available connection in performing the one or more data transfer functions,

wherein the plurality of processors that use different parallel programming paradigms can use a common implementation of the schedule module, the executor module, the connection manager and the multisend module via the language adaptor specific to a parallel programming paradigm implemented on the processor, and

wherein the multisend module sends many messages at the same time, each message or a group of messages identified by a connection identifier.

11. The method of claim 10, further including:

invoking the multisend module for each phase of the collective operation.

12. The method of claim 10, further including:

performing one or more optimizations during executing said one or more tasks.

13. The method of claim 12, wherein said one or more optimizations include pipelining, phase independence, or multi-color routing, or combinations thereof.

14. The method of claim 10, wherein the plurality of processors is connected by a tree network.

15. The method of claim 10, wherein the plurality of processors is connected by a torus network with deposit bit sends.

16. A non-transitory program storage device readable by a machine, tangibly embodying a program of instructions executable by the machine to perform a method of supporting collective communications on a plurality of processors that use different parallel programming paradigms, comprising:

receiving an application call invoking a collective operation from a processor of the plurality of processors via a language adaptor, the collective operation being a function supported at least in a message passing interface (MPI) including at least a reduce operation, all-reduce operation, and all-to-all operation, in which a group of processors in parallel computer system participate to perform a global operation where each of the processors contribute to the global operation;

creating a schedule module for defining one or more tasks in the collective operation;

creating an executor module for executing said one or more tasks;

invoking a multisend module to perform one or more data transfer functions associated with said one or more tasks;

creating a connection manager to control one or more connections and identifying an available connection; and

the multisend module further uses the available connection in performing the one or more data transfer functions,

wherein the plurality of processors that use different parallel programming paradigms can use a common implementation of the schedule module, the executor module, the connection manager and the multisend module via the language adaptor specific to a parallel programming paradigm implemented on the processor, and

wherein the multisend module sends many messages at the same time, each message or a group of messages identified by a connection identifier.

17. The program storage device of claim 16, further including:

invoking the multisend module for each phase of the collective operation.