METHOD AND APPARATUS FOR TRANSFORMATION OF MPI PROGRAMS FOR MEMORY CENTRIC COMPUTERS

• US 20200133749A1
• Filed: 10/22/2019
• Published: 04/30/2020
• Est. Priority Date: 10/25/2018
• Status: Active Grant

• Current Assignee: Electronics and Telecommunications Research Institute, Sangmyung University Industry-Academy Cooperation Foundation
• Original Assignee: Electronics and Telecommunications Research Institute, Sangmyung University Industry-Academy Cooperation Foundation
• Inventors: SHIN, Dong Ha, Lee, Je Man, Lee, Seung-Chul, Kim, Kang Ho
• Granted Patent: US 10,963,323 B2
• Time in Patent Office: Days
• CPC Class Codes: G06F 40/151 Transformation
• CPC Class Codes: G06F 40/205 Parsing
• CPC Class Codes: G06F 8/316 Aspect-oriented programming techniques
• CPC Class Codes: G06F 8/427 Parsing
• CPC Class Codes: G06F 8/4441 Reducing the execution time required by the program code
• CPC Class Codes: G06F 9/44563 Sharing
• CPC Class Codes: G06F 9/546 Message passing systems or structures, e.g. queues

Abstract

According to a method of transforming a message passing interface (MPI) program, an MPI function is parsed such that a computer may directly access data created by another computer through a memory when an MPI parallel program is executed on a computer system for distributed processing, a network function corresponding to the MPI function is created for network communication, and a bus line function is created for directly accessing the memory through a bus line.

Claims

1. A method of transforming message passing interface (MPI) programs for memory centric computers, the method comprising:

loading a source code of an MPI parallel program for a high performance computer system in which a plurality of computers are connected in parallel to each other for distributed processing;

parsing the source code to extract an MPI function statement;

creating, from an MPI function corresponding to the MPI function statement, a network function statement for communication with other computers through a network and a bus line function statement for directly accessing a memory shared by a plurality of computers through a bus line; and

storing a transformed source code comprising the network function statement and the bus line function statement.

2. The method of claim 1, wherein the network function statement comprises an absolute address transporter storing absolute address information.

3. The method of claim 2, wherein the absolute address transporter is created by a structure (object) comprising absolute address information about the memory.

4. The method of claim 1, wherein the MPI function comprises at least one of MPI_Send, MPI_Recv, MPI_Bcast, MPI_Reduce, MPI_Scatter, and MPI_Gather.

5. The method of claim 4, wherein a statement comprising the MPI_Send function comprises, as a parameter, an absolute address transporter storing absolute address information, and after the statement comprising the MPI_Send function is executed, a response from a target computer is waited for.

6. The method of claim 4, wherein a statement comprising the MPI_Recv function comprises, as a parameter, an absolute address transporter storing absolute address information, and after the statement comprising the MPI_Recv function is executed, a response signal is sent to a source computer which has executed an MPI_Send function command.

7. The method of claim 4, wherein when a corresponding process is a root process, a statement comprising the MPI_Bcast function executes the MPI_Bcast function to deliver absolute address information to all other processes and then waits for responses from all the other processes, and

when the corresponding process is not the root process, the statement comprising the MPI_Bcast function executes the MPI_Bcast function and then transmits a response signal to the root process.

8. The method of claim 4, wherein when a corresponding process is a root process, a statement comprising the MPI_Reduce function executes the MPI_Reduce function to receive absolute address information from all other processes and perform an operator calculation, and

when the corresponding process is not the root process, the statement comprising the MPI_Reduce function executes the MPI_Reduce function to deliver absolute address information about the corresponding process to the root process.

9. The method of claim 4, wherein when a corresponding process is a root process, a statement comprising the MPI_Scatter function executes the MPI_Scatter function to deliver an array of absolute address information about all other processes, and

when the corresponding process is not the root process, the statement comprising the MPI_Scatter function executes the MPI_Scatter function to store a message transmitted from the root process in absolute address information about the corresponding process.

10. The method of claim 4, wherein when a corresponding process is a root process, a statement comprising the MPI_Gather function executes the MPI_Gather function to store absolute address information about another process at an absolute address of the corresponding process as an array and then store absolute address information about all other processes as an array, and

when the corresponding process is not the root process, the statement comprising the MPI_Gather function executes the MPI_Gather function to deliver absolute address information about the corresponding process.

11. An apparatus for transforming MPI programs for memory centric computers, wherein the apparatus is configured to transform a source code of an MPI parallel program for a computer system in which a plurality of computers are connected in parallel to each other for distributed processing, and the apparatus comprises:

a storage medium storing a source code and a transformed source code; and

a processing device configured to parse the source code to extract an MPI function statement, creates, from an MPI function corresponding to the MPI function statement, a network function statement for communication with other computers through a network and a bus line function statement for directly accessing a memory shared by a plurality of computers through a bus line, store a transformed source code comprising the network function statement and the bus line function statement in the storage medium.

12. The apparatus of claim 11, wherein the network function statement comprises an absolute address transporter storing absolute address information.

13. The apparatus of claim 12, wherein the absolute address transporter is created by a structure (object) comprising absolute address information about the memory.

14. The apparatus of claim 11, wherein the MPI function comprises MPI_Send function, and a statement comprising the MPI_Send function comprises, as a parameter, an absolute address transporter storing absolute address information, and after the statement comprising the MPI_Send function is executed, a response from a target computer is waited for.

15. The apparatus of claim 11, wherein the MPI function comprises MPI_Recv function, and a statement comprising the MPI_Recv function comprises, as a parameter, an absolute address transporter storing absolute address information, and after the statement comprising the MPI_Recv function is executed, a response signal is sent to a source computer which has executed an MPI_Send function command.

16. The apparatus of claim 11, wherein the MPI function comprises MPI_Bcast function, and

when a corresponding process is a root process, a statement comprising the MPI_Bcast function executes the MPI_Bcast function to deliver absolute address information to all other processes and then waits for responses from all the other processes, and

when the corresponding process is not the root process, the statement comprising the MPI_Bcast function executes the MPI_Bcast function and then transmits a response signal to the root process.

17. The apparatus of claim 12, wherein the MPI function comprises MPI_Reduce function, and

when a corresponding process is a root process, a statement comprising the MPI_Reduce function executes the MPI_Reduce function to receive absolute address information from all other processes and perform an operator calculation, and

when the corresponding process is not the root process, the statement comprising the MPI_Reduce function executes the MPI_Reduce function to deliver absolute address information about the corresponding process to the root process.

18. The apparatus of claim 11, wherein the MPI function comprises MPI_Scatter function, and

when a corresponding process is a root process, a statement comprising the MPI_Scatter function executes the MPI_Scatter function to deliver an array of absolute address information about all other processes, and

when the corresponding process is not the root process, the statement comprising the MPI_Scatter function executes the MPI_Scatter function to store a message transmitted from the root process in absolute address information about the corresponding process.

19. The apparatus of claim 11, wherein the MPI function comprises MPI_Gather function, and

when a corresponding process is a root process, a statement comprising the MPI_Gather function executes the MPI_Gather function to store absolute address information about another process at an absolute address of the corresponding process as an array and then store absolute address information about all other processes as an array, and

when the corresponding process is not the root process, the statement comprising the MPI_Gather function executes the MPI_Gather function to deliver absolute address information about the corresponding process.