JP2005078244A - Program execution method and program performing device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To much more quickly and efficiently execute parallel processing by a plurality of PE(processor elements). <P>SOLUTION: A thread #0 transmits data to be transferred to a thread #1 attached with the start instruction of the thread #1 to a PE#1. When the instruction is inputted to the memory of the PE#1, interruption to the PE#1 is generated, and the thread #1 is started. Also, the thread #1 attaches the end notification of the thread #1 to the data(processing result) to be transferred to the thread #0, and when the notification is transmitted from the memory of the PE#1 to the PE#0, interruption to the PE#0 is generated, and the thread #1 is ended. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a program execution method and a program execution device for executing a program in cooperation between a plurality of PEs (Processing Elements).

  A huge program with various functions can be divided into partial programs that realize a group of functions, and only those necessary ones can be selectively combined to improve processing efficiency. The technique to plan has existed conventionally (for example, refer patent document 1).

  In recent computer systems, there is an example in which a plurality of PEs are mounted, and each partial program in the program is distributed to these PEs to improve the processing capacity. The methods are broadly divided into “Shared-Memory Multiprocessors” that share memory between PEs, and “Distributed-Memory Multiprocessors” that provides dedicated memory for each PE. ) ”.

  In such a system, a program based on SPMD (Single-Program, Multiple-Data) programming is often executed by using an inter-processor communication mechanism such as MPI (Message-Passing Interface).

  FIG. 7 is an explanatory diagram showing an example of a program created based on SPMD programming (hereinafter simply referred to as “SPMD program”). The illustrated program is executed by each of a plurality of PEs. Even if the program is the same, the processing branches depending on the ID (number) of the PE, so that parallel processing by a plurality of PEs is realized. For example, in the program shown in the figure, “myrank” is a variable indicating the ID, and Th0 (thread # 0) is used for PE (PE # 0) with myrank = 0, and Th1 (PE # 1) is used for PE with myrank = 1 (PE # 1). When thread # 1) is executed, PE2 with myrank = 2 (PE # 2) executes Th2 (thread # 2), and PE with myrank = 3 (PE # 3) executes Th3 (thread # 3). Become.

  However, at least in the case of the distributed memory type multiprocessor system, each processor executes only a part of the program (only a partial program), but the entire program is distributed to all PEs. Therefore, it is necessary to prepare a memory of that capacity for each PE, which increases costs. Therefore, it is considered to create a program based on MPMD (Multiple-Program, Multiple-Data) programming instead of SPMD.

  In programming based on MPMD, a program for each PE is created rather than a program in which all partial programs executed by each PE are combined as in SPMD. 8 is an explanatory diagram showing an example of a program for PE # 0, FIG. 9 is a PE # 1, FIG. 10 is a PE # 2, and FIG. 11 is an example of a program for PE # 3. As shown in the figure, since the program for each PE does not include a partial program for other PEs, the capacity of the memory of each PE can be reduced by that amount. Reference 1).

JP 2001-282513 A P. Pacheco "MPI Parallel Programming" Baifukan

  However, when a program based on MPMD programming (hereinafter simply referred to as “MPMD program”) is executed under the distributed memory multiprocessor system, the thread of the partner processor is used before / after communication between processors by MPI. (Or task, the same shall apply hereinafter) must be started / stopped.

  Each of the programs shown in FIGS. 8 to 11 is based on the premise that the other party's thread has already been activated. If the other party's thread has not been activated, for example, in the program of FIG. 8, "MPI_Send" Before calling and sending data to another PE, it is necessary to first call a thread activation API to instruct the other PE to start a thread. On the other hand, after receiving the result with “MPI_Recv” from the PE that has passed the data and requested processing, the API for thread termination must be called to instruct the PE to terminate the thread.

  FIG. 12 shows the concept of thread activation / termination and FIG. 13 shows the concept of data transfer in asynchronous remote procedure call using MPI or the like. 12 and 13, the shaded portion indicates that the thread is executing processing, the outlined portion indicates that the thread is temporarily suspending processing, and the straight portion indicates that the thread is processing. It means that each is not activated.

  Thread # 0 on PE # 0 starts thread # 1 on PE # 1 (FIG. 12 “START”) and waits for the completion of processing by thread # 1 (FIG. 12 “WAIT”), and the processing is completed. Then, thread # 1 is terminated (FIG. 12, “EXIT”).

  Between this start and end (while thread # 1 is starting), thread # 0 transmits data necessary for processing to thread # 1 ("MOVE" on the upper side of FIG. 13), and waits until transmission of all data is completed. ("WAIT" on the upper side of FIG. 13). After the transmission is completed, the thread # 0 instructs the thread # 1 to return the processing result (“MOVE” on the lower side of FIG. 13) and waits until reception of all the data is completed (“WAIT” on the lower side of FIG. 13). "), The processing is started again after the reception is completed.

  As described above, in the prior art, in order to realize parallel processing by a plurality of PEs, four procedures of (1) starting a thread (2) transmitting data (3) receiving data (4) ending a thread are performed. In response to this, the MPMD program also includes at least four types of instructions. For this reason, there is a problem that the overhead for distributing the processing among the processors becomes large, and the processing time is increased accordingly.

  An object of the present invention is to provide a program execution method and a program execution apparatus capable of executing parallel processing by a plurality of PEs more quickly and efficiently in order to eliminate the above-described problems caused by the prior art. .

  In order to solve the above-described problems and achieve the object, a program execution method and a program execution device according to the present invention execute a second program from a first program being executed on a first PE to a second PE. In the program execution method or the program execution device to be executed, when the second PE includes information indicating activation of the second program in the information received from the first program, the second PE The program is started.

  According to the present invention, the first program being executed by the first PE simply transmits the data used by the second program with a value indicating the start of the program attached, that is, the first program in advance. Even if the second program is not activated, the program of the destination PE can be activated.

  In the program execution method and the program execution device according to the present invention, the information transmitted from the second PE to the first program includes information indicating the end of the second program. In this case, the second program is terminated.

  According to the present invention, the first program being executed by the first PE can end the transmission source program only by causing the second program to transmit a value indicating the end of the program. .

  In the program execution method and the program execution device according to the present invention, a value indicating activation or termination of the second program is included in the information received from the first program or transmitted to the first program. It is determined by a detection circuit provided in the memory of the second PE whether or not it is present.

  According to the present invention, the activation or termination of the program can be executed at high speed by dedicated hardware.

  According to the program execution method and the program execution apparatus according to the present invention, the transmission destination and transmission source programs are activated or terminated in response to the data transfer, so that the activation instruction and termination instruction that are necessary separately from the conventional transfer instruction As a result, it is possible to increase the speed and efficiency of parallel processing by a plurality of PEs.

  Exemplary embodiments of a program execution method and a program execution device according to the present invention will be explained below in detail with reference to the accompanying drawings.

  The program execution device according to the present invention is specifically a computer system based on a distributed memory type multiprocessor system, and n PEs composed of processors and the like and n memories dedicated to each PE are mutually connected. Connected via an interconnection network. Each memory is provided with a detection circuit for detecting the fact when a predetermined value indicating the start or end of a thread is input and for interrupting the corresponding processor to start or end the thread. Yes.

  1 and 2 are explanatory diagrams schematically showing a procedure of parallel processing by a plurality of PEs in the program execution device according to the present invention. In the following, for convenience of explanation, it is assumed that n = 2, and the cooperation of PE # 0 and PE # 1 is described as an example, but the same procedure is performed when n> 2.

  First, when thread # 0 executing in PE # 0 requests processing to thread # 1 executed by PE # 1, input data required by thread # 1 is stored in PE # 0 memory (hereinafter referred to as PE # 0). “Memory # 0”) is transmitted to the memory for PE # 1 (“memory # 1”). At this time, a predetermined value indicating the activation of the thread is added to the data (FIG. 1 (1)).

  Upon receiving this, the detection circuit of the memory # 1 detects the predetermined value and generates an interrupt signal for the PE # 1 (FIG. 1 (2)). Upon receiving this, the PE # 1 starts the thread # 1. To do. The activated thread # 1 reads data transmitted from the memory # 0 to the memory # 1, and executes a predetermined process.

  Thereafter, when the above process ends, the thread # 1 adds a predetermined value indicating the end of the thread # 1 to the processing result (output data from the thread # 1) stored in the memory # 1 to store the memory # 1. Transmit to # 0 (FIG. 2 (3)), and thread # 1 is terminated by PE # 1 (FIG. 2 (4)). The detection circuit of the memory # 0 that detects the predetermined value at the time of reception generates an interrupt signal for the PE # 0 (FIG. 2 (5)).

  FIG. 3 is an explanatory diagram showing the operation status of each PE when parallel processing by a plurality of PEs is realized as shown in FIG. When PE # 0 receiving an instruction from the executing thread # 0 sends the data in thread # 0 and the thread activation instruction to PE # 1 (FIG. 3 (1)), an interrupt is received from memory # 1. Thread # 1 is activated by PE # 1 (FIG. 3 (2)). The data in the memory # 1 is read by the PE # 1 that has received an instruction from the activated thread # 1 (FIG. 3 (3)), and the processing in the thread # 1 is performed based on this data. It is carried out.

  PE # 1 (FIG. 3 (4)) that receives the processing result from thread # 1 transmits the data and thread end notification to PE # 0 (FIG. 3 (5)). Thread # 1 is terminated (FIG. 3 (6)). Therefore, the thread # 1 is activated only during the period from (2) to (6) in FIG. On the other hand, thread # 0 has been running all the time, but is suspended only while waiting for the processing result of thread # 1 (note that normal thread # 0 is not related to the processing of thread # 1 during this waiting period) However, it is simplified here).

  FIG. 4 is an explanatory diagram showing a specific example in which parallel processing by a plurality of PEs as shown in FIG. 1 is realized using MPI of the prior art. As shown in the drawing, “MPI_Send” is called when data is transmitted, and “MPI_Recv” is called and used when data is received. These MPI functions themselves are the same as in the prior art, but when a specific value that means activation is received from another thread, the thread is activated without any special instruction, and a specific value that means termination. Is transmitted to another thread, the feature of the present invention is that the thread being executed is terminated without any special instruction.

  FIG. 5 is an explanatory diagram showing a functional configuration of the program execution device according to the present invention. In the drawing, an execution unit 500a / 501a is a functional unit that is realized by a processor of each PE and executes a thread activated by a thread activation / termination unit 500e / 501e described later.

  The transmission unit 500b / 501b is a functional unit that transmits data in a storage unit 500d / 501d, which will be described later, to another PE in accordance with an instruction from the execution unit 500a / 501a. The receiving unit 500c / 501c is a functional unit that receives data transmitted from the transmitting unit 500b / 501b of another PE. Note that the transmission unit 500b / 501b or the reception unit 500c / 501c is specifically realized by software by “MPI_Send” or “MPI_Recv” called from the thread being executed.

  The storage units 500d / 501d are realized by the memories of the PEs, and are functional units that hold program codes of threads executed by the execution units 500a / 501a, data transmitted to other PEs, data received from other PEs, and the like. The thread activation / termination unit 500e / 501e is realized by a detection circuit provided in the memory of each PE, and detects that a value meaning thread activation is input from the reception unit 500c / 501c to the storage unit 500d / 501d. When the execution unit 500a / 501a is instructed to start the thread, and when it is detected that a value indicating the end of the thread is output from the storage unit 500d / 501d to the transmission unit 500b / 501b, the execution unit 500a / 501a is instructed. This is a functional unit that terminates the running thread.

  FIG. 6 is a flowchart showing the procedure of parallel processing by a plurality of PEs in the program execution device according to the present invention.

  First, the execution unit 500a on the PE # 0 side (the requesting side) creates transmission data to the PE # 1 side and saves it in the storage unit 500d in accordance with an instruction from the executing thread # 0 (step 500). S601). Specifically, this transmission data is composed of (1) a thread activation instruction (a predetermined value meaning the activation instruction) to the partner PE, and (2) input data to the activated thread. .

  Next, the transmission unit 500b that receives the instruction from the execution unit 500a transmits the transmission data in the storage unit 500d to the PE # 1 side (step S602). The data received by the receiving unit 501c on the PE # 1 side is stored in the storage unit 501d via the thread activation / termination unit 501e (step S603). At this time, the thread activation / termination unit 501e Waiting for detection of a thread activation instruction in the data and detecting it (step S604: Yes), thread # 1 is activated by the execution unit 501a that has received an instruction from the thread activation / termination unit 501e (step S604). S605).

  Thereafter, when the processing by the thread # 1 is completed, the execution unit 501a creates transmission data to the PE # 0 side according to an instruction from the thread # 1, and stores it in the storage unit 501d (step S606). Specifically, the transmission data includes (1) a thread end notification (predetermined value indicating end notification) to the partner PE, and (2) output data (processing result) from the terminated thread. It consists of one.

  The transmission data in the storage unit 501d is transmitted to the PE # 0 side via the thread activation / termination unit 501e by the transmission unit 501b that receives an instruction from the execution unit 501a (step S607), and the thread # 1 is terminated. (Step S608). At this time, the thread activation / termination unit 500e waits for detection of the thread termination notification in the data and, if detected (step S609: Yes), the execution unit 500a that has received an instruction from the thread activation / termination unit 500e. Thread # 0 is terminated (step S610).

  As described above, according to the program execution method and the program execution device according to the present invention, the thread can be activated or terminated simply by transferring the data. Call) and the prior art that required termination of the partner thread after the data transfer (calling of the termination API), respectively, can increase the processing speed and efficiency.

  In the above-described embodiment, thread execution control is performed by issuing an interrupt from the detection circuit in the memory. Alternatively, when a specific value is detected, a flag is set in a specific area of the memory and the processor is periodically set. Alternatively, the thread activation / termination may be realized by checking this flag.

(Supplementary Note 1) In a program execution method for causing a second PE to execute a second program from a first program being executed in a first PE (Processing Element),
A transmission step of transmitting at least one of information meaning activation of the second program and information handed over to the second program from the first program to the second PE;
A receiving step for receiving the information transmitted in the transmitting step;
A determination step of determining whether or not information indicating activation of the second program is included in the information received in the reception step;
In the determination step, when it is determined that the information received in the reception step includes information meaning the start of the second program, the second PE executes the second program. An activation process to be activated; and
A program execution method comprising:

(Appendix 2) At least one of the information indicating the end of the second program and the information handed over to the first program is transmitted from the second program activated in the activation process to the first program. A second transmission step;
A second determination step of determining whether or not the information transmitted in the second transmission step includes information indicating the end of the second program;
In the second determination step, when it is determined that the information transmitted in the second transmission step includes information indicating the end of the second program, the second PE An ending step of ending the second program;
The program execution method according to appendix 1, wherein:

(Supplementary Note 3) In the determination step and the second determination step, the information of the second PE is included in the information indicating whether the information includes activation or termination of the second program. The program execution method according to the above supplementary note 1 or supplementary note 2, wherein the determination is made by the detected circuit.

(Supplementary Note 4) In a program execution device for causing a second PE to execute a second program from a first program being executed in the first PE,
Transmitting means for transmitting to the second PE at least one of information meaning activation of the second program and information delivered to the second program;
Receiving means for receiving information transmitted by the transmitting means;
Determining means for determining whether or not the information received by the receiving means includes information indicating activation of the second program;
An activation unit that activates the second program when the determination unit determines that the information received by the reception unit includes information indicating the activation of the second program;
A program execution device comprising:

(Additional remark 5) The 2nd transmission means which transmits at least any one of the information which means the end of the 2nd program started by the starting means, and the information handed over to the 1st program,
Second determination means for determining whether or not information indicating the end of the second program is included in the information transmitted by the second transmission means;
When the second determination means determines that the information transmitted by the second transmission means includes information indicating the end of the second program, the second program is Ending means for ending;
The program execution device according to appendix 4, wherein the program execution device is provided.

(Supplementary note 6) The program execution device according to supplementary note 4 or supplementary note 5, wherein the determination unit and the second determination unit are realized by a detection circuit provided in a memory of the second PE. .

  The program execution method and the program execution apparatus according to the present invention are useful for a computer system that performs processing in cooperation with a plurality of PEs, and is particularly suitable for a computer system that performs communication between PEs by MPI calls.

It is explanatory drawing which shows typically the procedure (especially thread | sled starting procedure) of the parallel processing by multiple PE in the program execution apparatus concerning this invention. It is explanatory drawing which shows typically the procedure (especially completion | finish procedure of a thread | sled) of the parallel processing by multiple PE in the program execution apparatus concerning this invention. It is explanatory drawing which shows the operation condition of each PE in the case of implement | achieving parallel processing by multiple PE like FIG. It is explanatory drawing which shows the specific example which implement | achieves parallel processing by multiple PE like FIG. 1 using MPI of a prior art. It is explanatory drawing which shows the function structure of the program execution apparatus concerning this invention. It is a flowchart which shows the procedure of the parallel processing by multiple PE in the program execution apparatus concerning this invention. It is explanatory drawing which shows an example of the SPMD program by a prior art. It is explanatory drawing which shows an example of the MPMD program (for PE # 0) by a prior art. It is explanatory drawing which shows an example of the MPMD program (for PE # 1) by a prior art. It is explanatory drawing which shows an example of the MPMD program (for PE # 2) by a prior art. It is explanatory drawing which shows an example of the MPMD program (for PE # 3) by a prior art. It is explanatory drawing which shows notionally the start / end of the thread | sled between PE in a prior art. It is explanatory drawing which shows notionally the data transfer between PE in a prior art.

Explanation of symbols

500a, 501a execution unit 500b, 501b transmission unit 500c, 501c reception unit 500d, 501d storage unit 500e, 501e thread start / end unit

Claims (5)

In a program execution method for causing a second PE to execute a second program from a first program being executed by a first PE (processor element),
A transmission step of transmitting at least one of information meaning activation of the second program and information handed over to the second program from the first program to the second PE;
A receiving step for receiving the information transmitted in the transmitting step;
A determination step of determining whether or not information indicating activation of the second program is included in the information received in the reception step;
In the determination step, when it is determined that the information received in the reception step includes information meaning the start of the second program, the second PE executes the second program. An activation process to be activated; and
A program execution method comprising: A second transmission that transmits at least one of information indicating the end of the second program and information delivered to the first program from the second program activated in the activation step to the first program. Process,
A second determination step of determining whether or not the information transmitted in the second transmission step includes information indicating the end of the second program;
In the second determination step, when it is determined that the information transmitted in the second transmission step includes information indicating the end of the second program, the second PE An ending step of ending the second program;
The program execution method according to claim 1, further comprising:   In the determination step and the second determination step, a detection circuit provided in the memory of the second PE as to whether or not the information includes information indicating activation or termination of the second program. The program execution method according to claim 1, wherein the determination is made by the following method. In a program execution device for causing a second PE to execute a second program from a first program being executed on the first PE,
Transmitting means for transmitting to the second PE at least one of information meaning activation of the second program and information delivered to the second program;
Receiving means for receiving information transmitted by the transmitting means;
Determining means for determining whether or not the information received by the receiving means includes information indicating activation of the second program;
An activation unit that activates the second program when the determination unit determines that the information received by the reception unit includes information indicating the activation of the second program;
A program execution device comprising: Second transmitting means for transmitting at least one of information indicating the end of the second program started by the starting means and information delivered to the first program;
Second determination means for determining whether or not information indicating the end of the second program is included in the information transmitted by the second transmission means;
When the second determination means determines that the information transmitted by the second transmission means includes information indicating the end of the second program, the second program is Ending means for ending;
The program execution device according to claim 4, further comprising:

Images