JPH0916436A - Support method for optimization of parallel program - Google Patents

Abstract

PURPOSE: To decrease the man-hours for the optimization of a parallel program by easily finding a part of the program which needs to be tuned. CONSTITUTION: Optimum timing is determined by performing a display event accepting process 102 which accepts the parallel program 101 and an execution statement for timing display, a 1st program automatic converting process 104 which automatically generates a program displaying the timing of the accepted execution statement, a new event timing accepting process 112 which accepts the alteration of the timing of the displayed execution statement, a 2nd program automatic converting process 114 which automatically generates a program displaying new accepted timing, an event timing display process 109 which displays the execution results of those programs graphically, and a timing evaluating process 116 which evaluates the timing of the new execution statement.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention is intended for a parallel program in an arbitrary message passing type parallel computer,
The present invention relates to an optimization support method for this parallel program.

[0002]

2. Description of the Related Art In a message passing type parallel computer, necessary information is transmitted and received from processor to processor via a communication path, and efficient execution of communication between processors greatly contributes to optimization of programs. When using a general program description language c, FORTRAN, etc., the user describes the instruction regarding this communication on the program. Specifically, a processor sends information using a send instruction, a specified processor receives it using a receive instruction, and a processor sends br
It is common to send information to all processors using the oadcast instruction, and receive it using the receive instruction in all processors. If this interprocessor communication is not programmed by controlling the timing, unnecessary waiting time may occur and deadlock may occur. In a parallel program, optimizing the timing of this inter-processor communication shortens the execution time of the program, and ultimately improves the performance of the program.

One of the conventional techniques for supporting the optimization of communication between processors is XPVM (Oak Rid).
ge National Laboratories, etc.) executes parallel programs and graphically displays the timing of data transmission and reception between processors using a Gantt chart in which the vertical axis is the processor and the horizontal axis is the time. If this is used, the transmission / reception time, waiting time, timing, etc. can be seen at a glance, so it is easy to grasp the points to be corrected in the program, which is useful when optimizing the program.

There is also an advanced system concept. In the paper, "Visual Performance Debugger for Parallel Programs (Information Processing Society of Japan, System Software and Operating System 67-9, pp65-71)" The method is shown. That is, the information in which the program has been executed once is discharged to a file, and based on the information, the case where each processing such as each subroutine in the program is executed by different processors is calculated and presented. Similar to XPVM, this is a Gantt chart that graphically displays the timing of communication between user program processors, and while observing this display, the processor that is less loaded at the timing when the user is executing the program When you instruct to move the process on the screen, the system uses the information obtained by the first execution to simulate a case where the process is executed by a different processor, and the result is displayed in the Gantt chart as above. By using this, it is possible to quantitatively grasp how each processing of the program is distributed to which processor and how the entire processing time changes, which is useful when optimizing the program.

[0005]

If the first prior art is used among the above, it is possible to grasp the timing of each execution statement in the creation program and the data communication. Further, the use of the second conventional technique is useful for optimization in distribution of processing to processors.

However, in either case, the following problems remain in optimizing the parallel program.

(1) When the context of each partial process in each processor in the program is changed, it is unclear how the timing of execution of the program executed after the changing unit changes. Since there is inter-processor communication in parallel programs, processing timing is important,
As a result, understanding how the program will be executed thereafter greatly contributes to optimization of the program in terms of shortening the overall execution time of the program.

(2) If the Gantt chart is output regardless of the content of the program, for example, in the order in which the processors are connected, it may be difficult to grasp the program operation. Even if the number of processors is about 10, if there is communication, the line indicating communication goes from the line corresponding to each processor to the other processor, resulting in a fairly complicated graph depending on the display order of the processors.

An object of the present invention is to solve these problems and provide a support method for efficiently carrying out optimization of a parallel program.

[0010]

[0011]

Means for solving the above-mentioned problems are as follows.

Regarding (1): When the context of a plurality of executable statements is changed in the middle of the program, it is necessary to clarify how the execution timing of the program executed after the changing section changes. This can be achieved by the user being able to easily instruct such a change and by automatically generating and executing the converted program.

Further, it is achieved by automatically evaluating the result of the conversion program and displaying the timing of the optimum executable statement to the user.

Regarding (2): This is achieved by allowing the user to change the output order of the Gantt chart of each processor at any time.

[0015]

The above means has the following effects.

Regarding (1): When the context of a plurality of executable statements is changed in the middle of the program, the user changes the program depending on how the execution timing of the program executed after the changing section changes. By not being clear, the optimization efficiency of the program is improved.

Further, since it is automatically evaluated whether or not the timing of the execution statement is changed in the program, the optimization efficiency of the program is improved.

Regarding (2): Since the user can sequentially change the output order (display order) of the processors, the Gantt chart can be arranged in the processor order suitable for the contents of the program to display the operating status. Therefore, it is easy to understand the operation of the program.

[0019]

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 11 shows a hardware configuration on which the present invention is based. The parallel program 101, which is the target of optimization by the user, is input to the parallel program optimization support processing unit 100 on the workstation, and the parallel computer 1
It is sent to 101 and processed. The parallel computer 1101 has a plurality of cpu, each has a memory, and data is exchanged through a communication path. Hereinafter, the parallel program optimization support processing unit 100 will be described in detail.

FIG. 1 is a processing procedure of a parallel program optimization support method showing an embodiment of the present invention. The flow of processing is shown according to the figure. When the user inputs the parallel program 101 to be optimized into the parallel program optimization support processing unit 100, the processing unit 100 first performs the display event reception processing 102 to display the user information 1011 regarding the timing display of the execution statement by the user. The event is received and the event display table 103 is generated. Next, the first program automatic conversion processing 104 converts the parallel program 101 into the first conversion program 105 that displays the timing of the designated execution statement based on the information in the event display table 103. Next, the compile / link / execute process 107 incorporates the event display function 107 provided by the system into the first conversion program 105 and executes it to generate a first event display result table 1081. Then, the event timing display processing 109 graphically displays 110 the timing of the designated execution statement for each processor.

The user who sees the timing display of the designated execution statement considers which execution statement in the program should be shifted to obtain the optimum program.

Next, a new event timing acceptance process 11
With 0, new timing information 111 such as which user newly adjusts or shifts the timing of which execution statement
Is received and the event timing table 113 is generated. Next, the second program automatic conversion processing 114 converts the first conversion program 105 into the second conversion program 115 based on the information of the event timing table 113 so that the program is executed at a newly designated timing. Next, the event display function 107 provided by the system is incorporated into the second conversion program 115 by the compile / link / execute process 107 and executed.
The second event display result table 1082 is generated. Then, the event timing display processing 109 graphically displays 110 the timing of the designated execution statement for each processor.

Next, the timing evaluation processing 116 which receives two of the first event display result table 1081 output by the execution of the parallel program 101 and the second event display result table 1082 output by the execution of the second conversion program 115. Then, the effect of the timing change on the program is obtained, and the timing evaluation result 117 is displayed (block 118). Further, the optimum timing determination processing 119 which receives the timing evaluation result 117 as an input
The optimum timing of the optimum executable statement is determined by and the optimum timing is displayed (block 120). The above is the outline of the parallel program optimization processing procedure.

The present invention will be described in detail below with reference to examples.

The input information is the parallel program 101 to be optimized, and the user FORTs multiple processors.
It is written in a programming language such as RAN or C. The user inputs this parallel program 101 to the parallel program optimization processing unit 100. In response to this, the parallel program optimization processing unit 100 activates the display event reception processing 102, and the event reception screen 201 shown in FIG.
Is displayed.

The details of the display event reception process 102 will be described. The event reception screen 201 output by the display event reception processing 102 displays the parallel program 101 and receives the user information 1011 regarding the timing display of the execution statement. There are two types of timing display designation methods. One is to directly mark each line in the display program.
The user inputs a command to the command string 202 of the execution statement to be displayed in the timing in the parallel program 101 according to the description 204 of the command corresponding to the program source. Enter S to display the start timing of the executable statement, E to display the end timing, and A to display both. Further, the executable sentence is given a name and input in the name column 203. In this example, row number 50 BROADC
The AST statement is instructed to display the timing at the start of execution, and BRD1 is given as the name (205).

Another timing display designating method is a method of designating according to the type of executable statement. For example, it is used to display all READ statements, SEND statements in a specific subroutine, and so on. The user first selects, from the display timing selection 207 of the sentence type correspondence command 206, whether to display the execution sentence at the start time, the end time, or both. Next, specify the executable statement to be displayed, its selection range, and the name of the executable statement on the command line 2.
Fill out 08-210. After completing the entry, click the next button 2
Press 11 to register all necessary executable statements. In this example, all RECEIVE statements in the subroutine SUB1 are registered under the name RECV1 and are specified to be displayed at the execution start timing.

The above two methods are used to give an instruction, and the end button 212 is pressed when the instruction is completed. The display event reception process 102 ends reception of the timing display of the execution statement by pressing the end button 212, and generates the event display table 103.

An example of the event display table 103 is shown in FIG.
Shown in (A) and (B). The event display table 103 consists of two tables, one of which is the name-to-ID table 10
The user-specified name 311 and the corresponding system ID 312 automatically assigned by the system in 3A are listed in one-to-one correspondence. The other is the basic event display table 103B, in which all the execution statements for which timing is displayed are listed in the order of line numbers. Each has a line number 313, a system ID 312, an execution statement 314, and a display type 315 as information. The event display table 103 is generated, and the display event reception process 102 ends.

Next, the first program automatic conversion processing 104, which receives the parallel program 101 and the event display table 103, is executed. FIG. 4 shows a processing example of the first program automatic conversion processing 104. In the first program automatic conversion processing 104, while looking at the event display table 103, the event display function is inserted into the parallel program 101 before and after the execution statement to be timing-displayed. If the timing display is specified at the beginning of the executable statement, it will be inserted before the executable statement, at the end it will be displayed after the executable statement. Further, the initial setting section is inserted at the head of the parallel program 101.

In the illustrated example, the first program automatic conversion processing 104 converts the parallel program 101 into the first conversion program 105 for displaying the timing of the designated execution statement, based on the information in the event display table 103. .
Since the call statement 401 of the subroutine INITEVENT which is the initialization section is automatically inserted at the beginning of the first conversion program 105, and CALL RECEIVE (...) on the 10th line of the parallel program is registered in the event display table 103. , First
In the conversion program 105, the line next to CALL RECEIVE (...)
CALL EVENT (2,3) is inserted (402). CALL EVENT (2,
The second argument 3 of 3) indicates that the type of display is after the end of the executable statement, and the second argument 3 is the system ID of the name RECV1 given by the user. Similarly, CALL BROAD on line 50
CAL before CAST () and after CALL RECEIVE () on line 53
Insert L EVENT (...) (403, 404).

This CALL EVENT () is an event display function 10
6 is a function provided by this parallel program optimization support process, and is for recording the timing of an executable statement. The source code is shown in FIG. As shown in Figure 4,
The arguments of this function are the FLG indicating the timing of the designated execution statement and the variable ID indicating the system ID (501). As the processing contents, the time is obtained by the CLOCK function (502), the number of the executed sentence of the system ID is obtained (503), and these are output to the file of the first event display result table 1081 (504).

When the first program automatic conversion processing 104 is completed, the output first conversion program 105 and the event display function 106 are input, the compile / link / execute processing 107 is executed, and the first event display result is obtained. A table 1081 is generated. Then, the contents of the first event display result table 1081 are graphically displayed by the event timing display processing 109 (110).
An event display screen 1100 by the event timing display processing 109 is shown in FIG.

In the first event display result table 1081, the results of all the execution statements instructed to display the event are registered in time series for each process. Registration item is system
ID, execution order with the same system ID, and execution start or end time. Event timing display processing 109
Displays the contents of the event display result table 1081 graphically. Specifically, a line on the time axis is displayed for each processor (601), and the time at which each execution statement is executed is displayed. Further, the color of the display timing, that is, whether the execution statement starts or ends is displayed. When both are displayed, the width from the start to the end is shown thickly. In addition, by pointing each execution statement on the time axis with the mouse, detailed time is displayed. For example, in the case of the figure, an executable statement named BRD1 is executed in the processor 1 (602), and data from this is received by the executable statement of the name RECV2 of the processors 2 and 3 (60
3), and this RECV2 is executed only during the time indicated by the thick line (604), and so on, can be read at a glance. Also, the user can freely change the order of the time axis lines for each processor to make it easier to see.

Next, the user looks at the timings of the executable statements shown in FIG. 6 and specifies different timings for them. This is the new event timing acceptance process 1 of FIG.
Accepted by 12. FIG. 7 shows an example of the new event timing reception screen 700 output by the new event timing reception process 112. The initial screen is an output screen of the event timing display processing 109 of FIG. On the other hand, the user gives the user information 1011 regarding the timing display of the execution statement. Specifically, the timings of a plurality of execution statements executed by different processors can be set at the same time, and the context of the execution statement timings can be changed. For example, in the case of the figure, the name SU of the processor 2
It is instructed to match the timing of B1 and the name SUB2 of the processor 3 (701), and to execute the name SEND1 of the processor 2 after the name SEND1 of the processor 3 (702). These are the timing buttons 7 at the bottom of the screen.
Use 03 to indicate. Instructions can also be made interactively on the screen in this way. When the instruction is completed, the user presses the end button. New event timing acceptance process 1
10 terminates the reception of the new event timing by pressing the end button, and registers the data of all the execution statements whose context is registered in the event timing table 111.

An example of the event timing table 111 is shown in FIG. The table items include a PE number 801, which indicates the processor number, a system ID 802, an execution order 804 indicating the number of execution of the system ID, and
The execution number uniquely determined from these pieces of information (the above is the target execution statement), and further, the processor number 805, the execution number 806, and the contextual relationship 807 of the related execution statement related to this execution statement. In the example of FIG. 8, two contexts designated by the new event reception screen 700 of FIG. 7 are registered. One is information for simultaneously executing SUB1 of processor 2 and SUB2 of processor 3. The first line is from PE number to execution number for SUB1 and SU related to SUB1.
The PE number of B2, the execution statement number, and the relationship of simultaneous execution are described (808). SUB2 in reverse on the third line
The relationship of SUB1 with respect to is described (810). 2
In the line, S of processor 3 for SEND1 of processor 2
The relationship of END1 is described (809), and conversely, the relationship of SEND1 of processor 4 to SEND1 of processor 3 is described on the fourth line (811). The new event timing reception process 110 uses this event timing table 11
Generate 1 and exit.

Next, the second program automatic conversion processing 114, which receives the first conversion program 101, the first event display result table 1081 and the new event timing table 111, is executed. FIG. 9 shows an example of the second program automatic conversion processing 114. In the second program automatic conversion processing 114, while looking at the first event display result table 1081 and the new event timing table 111 with respect to the first conversion program 101, an execution statement whose timing is to be controlled waits for an event before and after the execution statement. Insert a function etc. When issuing multiple executable statements at the same time, issue a barrier synchronization function for those executable statements and receive them with the RECEIVE function immediately before the executable statement. In addition, the SEND and RECEIVE functions are placed between the execution statements that have context, to control the timing. For example, in order to synchronize the names SUB1 and SUB2 instructed on the new event timing reception screen 700 of FIG. 7, SUB1 and SUB2 are changed to be executed immediately after receiving a signal from the same barrier function. Specifically, as shown in the second conversion program 115 in FIG. 9, CALL SUB1 ()
Before that, the RECEIVE function is automatically inserted by the second conversion program processing 114 (901). Further, in order to execute the order relation of the name SEND1 designated in FIG. 7, when the SEND1 of the processor 2 is finished, the SEND function is sent to the processor 2, and after the processor 2 receives this, SEND1 is executed. Also in the second conversion program 115, SEND1 of the processor 1 executes after receiving the data from the processor 2 by the automatically inserted RECEIVE function (902).

When the second program automatic conversion process 114 is completed, the second conversion program 115 and the event display function 106 that have been output are used as inputs to execute the compile / link / execute process 107 to generate the second event display result table 1082. To do. Then, the event display result table 1082 is displayed by the event timing display processing 109.
The content of is displayed graphically (110). This second
The event display result table 1082 has the same format as the first event display result table 1082, and the results of all the execution statements instructed to display the event are registered in time series for each process. The registered items are the system ID, the execution order with the same system ID, and the execution start or end time.

FIG. 10 shows the event display screen 110 which is the execution result of the new event timing information 111.
SUB1 and SUB2 are executed simultaneously as specified (10
02), and processor 1 after SEND1 of processor 2
Is being executed (1003). The waits are shown as small boxes, respectively, for timing (1001, 1004). If the timing cannot be adjusted, it is indicated by a cross mark (1005).

Finally, FIG. 12 shows an example of the processing contents of the timing evaluation processing 116 and the optimum timing determination processing 119 for evaluating the obtained result and displaying the optimum timing. First, the first event display result table 1081 output by the execution result of the original parallel program 101 and the second conversion program 115 executed at the designated timing.
Second event display result table 1 output as the execution result of
The influence of the timing change on the program is obtained by the timing evaluation processing 116 with 082 as an input, and the timing evaluation result 117 is displayed (118). As shown in the figure, the content shows a new event timing 1201 instructed by the user and an evaluation 1202 of the event timing obtained from the processing result. The specific evaluation contents are as follows. (A) and (B) show waiting times required to execute each designated event timing. (C) shows the event timing in each PE changed by the designated event timing. (D), (E), and (F) show the processing time of each PE changed by the designated event timing. (G) shows the processing time of the entire program that changes depending on the designated event timing. When this processing is finished, finally, the optimum timing of the optimum executable statement is obtained by the optimum timing determination processing 119 with the timing evaluation result 117 as an input, and the optimum timing is displayed (120). Specifically, the difference between the waiting time and the processing time due to the event timing change calculated above is calculated,
We will show whether the processing time will be shorter if the program is changed to a new timing, and how much the waiting time by changing the event timing needs to be shortened.

According to this embodiment, the following effects can be obtained.

First, it was possible to instruct the display of the timing of the target executable statement by the graphical screen display shown in FIG. 2 and simple input in an interactive manner, and the result could be displayed as shown in FIG. Next, the timings of the execution statements were shifted by the simple instruction shown in FIG. 7, and the result could be displayed as shown in FIG. Further, as shown in FIG. 12, the original program and the case where the timing is shifted are evaluated as to how the execution time changes, and the optimum timing is displayed, so that the future optimization policy can be simplified. Be taken.
In this example, if the waiting time for SUB2 and SEND1 can be kept within s-sec, the new timing can shorten the program execution time. Since it is important to shorten the s-sec in this example, we decided to debug so that the program can be executed at a new timing as a future optimization policy. In this way, since the optimization policy is shown, the man-hours for optimizing the parallel program can be reduced.

[0044]

According to the present invention, executable statements can be synchronized among the processors in a parallel program by a graphical screen display and simple input in an interactive manner. Therefore, the part of the program that needs to be tuned can be easily displayed, and the number of steps for optimizing the parallel program can be reduced. In addition, since the guideline for program evaluation and how much to change the parallel program when the execution statements are synchronized is shown, the man-hours for optimizing the parallel program can be reduced.

[0045]

[Brief description of the drawings]

FIG. 1 is a flowchart of a parallel program creation support procedure showing an embodiment of the present invention.

FIG. 2 is a diagram showing an example of an event display acceptance screen.

FIG. 3 is a diagram showing an example of an event display table.

FIG. 4 is a diagram showing a conversion example of a first program conversion process.

FIG. 5 is a diagram showing a source program of an event display function.

FIG. 6 is a diagram showing an example of an event display screen.

FIG. 7 is a diagram showing an example of a new event reception screen.

FIG. 8 is a diagram showing an example of a new event timing table.

FIG. 9 is a diagram showing a conversion example of second program conversion processing.

FIG. 10 is a diagram showing an example of an event display screen after executing a program using new timing information.

FIG. 11 is a hardware configuration diagram.

FIG. 12 is a diagram showing an example of timing evaluation processing and optimum timing determination processing.

[Explanation of symbols]

100 ... Parallel program optimization support processing unit, 102
Display event reception process, 103 event display table, 104 first program automatic conversion process, 105 first conversion program, 106 event display function, 1081 first Event display table 1, 1082 ... Second event display table 2,
109 ... Event timing display processing, 112 ...
..New event timing reception processing, 113 ... Event timing table, 114 ... Second program automatic conversion processing, 115 ... Second conversion program, 1
16 ... Timing evaluation processing, 117 ... Timing evaluation, 118 ... Evaluation result display, 119 ... Optimal timing determination processing, 120 ... Optimal timing display.

Claims (6)

[Claims] 1. A user specifies a parallel program to be optimized and at least one execution statement in the parallel program, and the parallel program is automatically converted so as to display the timing of the specified execution statement. The specified conversion statement is displayed in time series for each processor by executing the specified first conversion program, and the user further specifies the new timing for each execution statement by observing the display and newly specifies it. A parallel program optimization support method comprising automatically converting a first conversion program to be executed at a specified timing, executing the obtained second conversion program, and displaying the timing of an executable statement. 2. The parallel program optimization support method according to claim 1, wherein the process of designating the executable statement comprises a process of directly designating the executable statement in the parallel program displayed on the screen. 3. The parallel program optimization support method according to claim 1, wherein the process of receiving the timing designation input of the execution statement is a process performed in the form of direct instruction from a timing display screen. 4. The parallel program optimization support method according to claim 1, wherein the process of displaying the timing of the execution statement comprises a process of sequentially rearranging and displaying the display order of each processor in an order specified by the user. . 5. A parallel program to be optimized, a plurality of execution statements in the parallel program, and timings different from the original program of those execution statements are designated by a user, and the designated execution statements are executed at designated timings. The parallel program is automatically converted so that the conversion program is executed to evaluate how the specified timing affects the subsequent programs, and the optimum timing is determined based on the evaluation result. A parallel program optimization support method characterized by displaying timing. 6. The parallel program optimization support method according to claim 5, wherein the evaluation process includes a process of displaying an evaluation result.