JP2004102594A - Method for analyzing number of times of execution of parallel execution program and its program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To confirm the distribution volume of a load for parallelization with numerics, and to support the parallelization tuning of a parallel execution program. <P>SOLUTION: A program to be analyzed execution number of time measuring means 12 accumulates the number of times of execution in a count area in response to a count instruction inserted into a program to be analyzed at the time of translation to measure the number of times of execution of an execution sentence by each of parallel execution units. An execution number of time measurement result analyzing means 13 measures the mean number of rotation of a loop from the measurement result of the number of times of execution, and calculates a performance improvement rate when changing the degree of parallelization. Also, the number of times of execution of the execution sentence in each procedure in programs executed in parallel and the number of times of execution of the execution sentence by each of the parallel execution units of each loop are calculated, and the calculation results are displayed in the order of the larger number of times, and information indicating the balance of the number of times of execution between the parallel execution units. <P>COPYRIGHT: (C)2004,JPO

Description

[0001]
TECHNICAL FIELD OF THE INVENTION
The present invention relates to a support function for parallel tuning of programs in the field of HPC (High Performance Computing System) large-scale scientific and technical calculation, and more particularly, to a method of parallel execution of programs written in a high-level language such as Fortran and performing load balancing by parallelization. The present invention relates to a method for analyzing the number of executions of a parallel execution program and a program for analyzing the number of executions, which support a process of improvement.
[0002]
The parallelization of the program to be tuned is targeted at a shared memory type parallel method and a memory distributed type parallelized method.
[0003]
In the shared memory parallel method, a program is executed in parallel on a plurality of parallel execution units called threads. When the user executes the program, the user specifies the range of parallelization in the source code according to the OpenMP parallelization syntax and the amount of parallel load division indicating the distribution of the loop load to the parallel execution unit in the range.
[0004]
For example, in the case of the parallelization target program shown in FIG. 15, the range from the basic block 1 to the basic block 3 is specified by the parallelization syntax (“! @OMP PARALLEL”, “! @OMP END PARALLEL”). I have. Here, the basic block is an execution sentence string whose execution order is not likely to change. For the part corresponding to the loop, the compiler creates an object code generation function corresponding to the parallel load split amount specified by the user, and executes the function in parallel. In the case of the parallelization designation shown in FIG. 15, for the basic blocks 1, 2, and 3, each of the parallel execution units performs redundant execution in which the same processing is performed, and for the loop portion, the load of the loop is distributed according to the parallel load division amount. Is done.
[0005]
In the memory distributed parallelization method, data is divided into a plurality of parallel execution units called processes, and programs are executed in parallel. To specify the parallelization, an MPI (Message Passing Interface) library is inserted into the source code at the position where the parallelization is to be performed, and the parallel load division amount is specified.
[0006]
The present invention outputs information that supports parallelization tuning of these parallelized programs and a parallelized program in which both types are mixed.
[0007]
[Prior art]
In order to support memory-shared parallel and memory-distributed parallel programming, it is particularly necessary to measure the average number of rotations of the loop to be load-balanced. A function that can be checked for each parallel execution unit is required.
[0008]
As a technique for optimizing the parallel processing of a program, for a parallel computer composed of multiple processors, data and programs are allocated to each processor element, and the processing time is measured at the time of execution. There has been a technique of dynamically re-assigning the data so that the data is evenly distributed (for example, see Patent Document 1). In addition, a technique for measuring the average rotation speed of each loop of the sequential execution program has been conventionally known (for example, see Non-Patent Document 1).
[0009]
[Patent Document 1]
JP-A-2-132526 (page 1)
[Non-patent document 1]
FUJITSU, "For UXP / V Analyzer User's Guide V20" (J2U5-0132-02), Fujitsu Limited, September 1999, 2nd edition, p55-56
[0010]
[Problems to be solved by the invention]
However, the conventional program execution number analysis function is for sequential execution programs, and has no function to support program parallelization. That is, there is no function for measuring the average rotation speed of each loop of the parallel execution program for each parallel unit.
[0011]
Also, for example, the technique described in Japanese Patent Application Laid-Open No. 2-132526 of Patent Document 1 does not measure the number of loops and displays it for tuning. However, there is a problem that parallelization tuning cannot be performed after confirmation.
[0012]
To be more specific, in the past, functions to predict the performance of a tuning target program by parallelization, to check the balance of the number of executions of procedures and loops between parallel execution units, and to check the There was no function to display the parallel status of executable statements. For this reason, it has been difficult to perform parallelization tuning of the parallelized program based on the load distribution in the parallel execution.
[0013]
SUMMARY OF THE INVENTION It is an object of the present invention to solve the above-mentioned problems of the prior art and to provide a tuning support technique that allows a user to check a load distribution state in parallel execution and then tune a parallelized program.
[0014]
[Means for Solving the Problems]
In order to solve the above problems, the present invention focuses on the following points.
[0015]
(1) The execution load of a program is represented by an execution time. Procedures, loops, and executable statements that are executed more often tend to require more execution time. By utilizing this tendency to express the execution load by the number of executions, a tuning support function that expresses the execution load by a numerical value that can be easily fed back to the source code is realized. Also, since the amount of load distribution for parallelization (mainly the number of rotations of the loop) is often specified by a numerical value, it is useful for tuning to be able to obtain the execution load of the program by a numerical value. Furthermore, if the load of each rotation of the loop is the same regardless of the number of rotations, a performance effect can be exhibited by obtaining the average number of rotations of the loop and dividing the value equally into parallel execution units.
[0016]
(2) If there is performance improvement prediction information when the number of rotations of the loop is divided into parallel execution units in proportion to the increase in the number of parallel execution units, tuning can be effectively advanced.
[0017]
(3) When a program is executed in parallel, if the parallel execution status can be confirmed numerically for each procedure, loop, and execution statement unit, the tuning information becomes easy to understand.
[0018]
The present invention accurately measures the average number of rotations of a loop, which is most noticed for load distribution in parallelization of a program, based on the above noted points.
[0019]
Further, the present invention realizes a performance improvement prediction function when the number of revolutions of the loop is divided in detail in proportion to the increase in the number of parallel execution units. Also, the number of executions of procedures and loops is displayed so that they can be compared between parallel execution units. Also, procedures and loops are displayed in descending order of load so that a tuning target can be easily selected. In addition, it is displayed so that the number of executions and the parallel execution status can be checked in accordance with the execution statement of the source code.
[0020]
That is, the present invention relates to a method for analyzing the number of executions of a parallel execution program, the method comprising translating a program to be analyzed and inserting an instruction for accumulating the number of executions into each procedure during the translation; A program execution number measuring step of accumulating the execution number of the set unit for each parallel unit, and an execution number measurement result analyzing step of analyzing the execution number measurement result of the parallel execution unit and displaying the analysis result on a terminal It is characterized by.
[0021]
Further, in the present invention, in the program translation step, a call instruction of a function having a function of securing a count area for accumulating the number of executions of each execution statement in the procedure is inserted at a start position of each procedure, and the number of executions is reduced. Inserting the count instruction to be accumulated at the start position of each basic block, the start position of each loop, and the position of the first executable statement in each loop, which are executable statement strings in which the execution order in the procedure is not likely to change; In the execution number measurement step, the number of executions is accumulated in the count area by a count instruction that appears during execution of the instruction of each procedure, and when the entire program ends, the number of executions in the count area is stored in the execution information file for each parallel execution unit. The execution number measurement result analysis step extracts the execution number measurement result of each parallel execution unit from the execution information file and outputs the result to each loop execution. And calculates that the average rotational speed of the loop is the average number of executions of the loop in the execution statement for each parallel execution unit.
[0022]
By using the present invention, the state of load distribution in parallel execution can be clarified. Therefore, if the load of each rotation of the loop is the same regardless of the number of rotations, the performance effect can be exerted by equally dividing the value of the average number of rotations of the loop into parallel execution units.
[0023]
In the present invention, the execution number measurement result analysis step further includes a step of total execution of execution statements in the program converted into sequential execution based on the execution number measurement result of each parallel execution unit extracted from the execution information file. Calculating the number of executions, the total number of executions of the executable statement in the program converted to the calculated sequential execution, the total number of executions of the in-loop execution statement in each parallel execution unit, the parallel frequency, Calculating the total number of executions of the execution statement in the parallel execution unit when dividing the total number of executions of the execution statement in the program converted into the sequential execution into each parallel execution unit based on the overhead; And the calculated total number of executions of the executable statement in the parallel execution unit and the calculated total number of executions of the executable statement in the program converted into the sequential execution. And having a step of calculating the performance improvement rate when changing the column frequencies.
[0024]
By using the present invention, it is possible to realize a performance improvement predicting function when the number of revolutions of the loop is divided in detail in proportion to the increase in the parallelism, and the tuning can be effectively advanced.
[0025]
Further, in the present invention, the execution number measurement result analyzing step further includes an execution statement in each procedure in the program executed in parallel based on the execution number measurement result of each parallel execution unit extracted from the execution information file. Calculates the number of times of execution and the number of executions of executable statements in the parallel execution unit of each loop, displays the calculation result on the terminal in descending order of the number of executions, and displays the balance of the number of executions between the parallel execution units on the terminal And
[0026]
By using the present invention, it becomes possible to balance the load between the parallel units of each procedure and loop and to select a parallelization candidate.
[0027]
In the present invention, the execution count measurement result analysis step further includes, based on the execution count measurement result of each parallel execution unit extracted from the execution information file, a total execution count for each execution statement associated with a source code. The number of times and the degree of parallelism are calculated and displayed on a terminal.
[0028]
By using the present invention, the parallel execution status can be easily recognized, and the feedback to the source code for improving the performance is promoted.
[0029]
The program of the present invention is a program for causing a computer to execute a method for analyzing the number of executions of a parallel execution program, and for accumulating the number of executions at least for each parallel execution unit at the time of translating the parallel execution program to be analyzed. A translation step of inserting an instruction into each procedure, a measuring step of accumulating the number of executions of the instruction set at the time of translation for each parallel unit by executing the translated parallel execution program, and a number of executions of the parallel execution unit Executing a parallel execution program for causing a computer to execute a method for analyzing the number of executions of a parallel execution program, the method comprising: analyzing a measurement result and outputting at least an analysis result indicating a load distribution in the parallel execution. It is a number analysis program.
[0030]
BEST MODE FOR CARRYING OUT THE INVENTION
Hereinafter, embodiments of the present invention will be described with reference to the drawings. FIG. 1 is a diagram showing an example of a hardware configuration for realizing a method for analyzing the number of executions of a parallel execution program according to the present invention.
[0031]
In the figure, reference numeral 1 denotes a parallel execution program execution result analyzer such as a workstation (WS) for analyzing the execution result of the parallel execution program. 2 and 3 are nodes A and B respectively having one or a plurality of CPUs 21 and 31 and memories 22 and 32 and executing parallel execution programs. Reference numeral 4 denotes an execution information file in which a result of measuring the number of executions of the parallel execution programs executed in the nodes A2 and B3 is recorded. Reference numeral 5 denotes a high-speed bus for communication between processors connecting the nodes A2 and B3. Although the execution information file 4 is a shared file in FIG. 1, the present invention can be realized even if the execution information file 4 is not shared.
[0032]
The method for analyzing the number of executions of the parallel execution program according to the present invention is used for parallelization tuning support of a shared memory type parallelization method performed in the node A2 or the node B3, and a memory performed between the node A2 and the node B3. It is also used for parallel tuning support of distributed parallel processing.
[0033]
FIG. 2 is a diagram showing an outline of a method for analyzing the number of executions of a parallel execution program according to the present invention. 6 is an analysis result output device such as a display or a printer, 7 is a source program of the analysis target program in which parallelization instructions are embedded, and 7 'is a load of the analysis target program in which instructions for measuring the number of executions are embedded by translation. A module 11 is an analysis target program translating means for translating (including a link) the analysis target program, a reference numeral 12 is an analysis target program execution number measurement means, and a reference numeral 13 is an execution number measurement result analyzing means for analyzing the execution number measurement result. Note that the analysis target program translating means 11 and the execution count measurement result analyzing means 13 need not always be realized by the same device, and may be realized by different devices.
[0034]
Hereinafter, processing performed by the parallel execution program execution result analysis device 1 will be described. FIG. 3 is a diagram showing an example of the execution number analysis processing flow of the parallel execution program of the present invention.
[0035]
First, the analysis target program translating means (compiler) 11 translates the analysis target program 7 described in a high-level language such as Fortran (step S1). In the translation of the analysis target program 7, an instruction for calling a function having a function of securing an area (count area) for accumulating the execution count of each execution statement in the procedure is inserted at a start position of each procedure in the object program. I do.
[0036]
Further, the analysis target program translating means 11 inserts an instruction (count instruction) for accumulating the number of executions at a next position in the procedure in the object program.
{Circle around (1)} The starting position of each basic block (executable statement string whose execution order is not likely to change)
(2) Start position of each loop
(3) First executable statement in each loop
These count instructions can be expressed, for example, in the form of source code:
Area name (x) = area name (x) +1
Such an instruction.
[0037]
FIG. 4 shows an example of insertion of a count instruction into an object program when the analysis target program 7 is translated. In the example shown in FIG. 4, a function call instruction for managing the count area is inserted at the start position of the procedure, and a count instruction for counting the number of executions is inserted at the head position of each of the basic blocks 1 to 3. , A loop start count instruction is inserted in the first executable statement in the loop. The insertion of the instruction as described above is performed by selecting a predetermined compiler option at the start of the translation process.
[0038]
Next, the analysis target program execution count measuring means 12 performs an execution count measurement process of the analysis target program 7 (step S2). Specifically, the CPU 21 and 31 of each of the nodes A2 and B3 shown in FIG. 1 execute a load module (analysis target program 7 ') obtained from an object program obtained by translating the analysis target program 7, and The number of executions is accumulated for each parallel execution unit using the set count instruction.
[0039]
The accumulation of the number of executions is performed as follows. At the start of each procedure, a function that allocates a count area is called. This function manages the count area using the function name of the caller and the parallel execution unit as keys. When this function is called, it notifies the caller of the count area corresponding to the key.
[0040]
If a count instruction inserted during translation appears during execution of a procedure instruction, the count instruction accumulates the number of executions in a count area prepared at the start of the procedure. Then, at the end of the entire program, the number of executions of the count area is recorded in the execution information file 4 for each parallel execution unit. FIG. 5 shows an example of a method of measuring the number of executions of the analysis target program 7 shown in FIG. In the execution count measurement shown in FIG. 5, based on the generated function created by the compiler for the parallelization specification in the same manner as the parallelization specification method shown in FIG. And the number of executions is measured for each parallel execution unit.
[0041]
Next, the execution count measurement result analysis means 13 performs an execution count measurement result analysis process (step S3). More specifically, the execution result measurement result of each parallel execution unit is extracted from the execution information file 4, analyzed as follows, and output to the analysis result output device 6 such as a display or a printer.
[0042]
First, the parallel execution units and their average loop speed are calculated for each loop by the following formula.
[0043]
Average loop speed
= Number of executions of the first executable statement in the loop / number of executions of the loop start position
In addition, the following expression is used to calculate the total number of executions of an executable statement in a parallel execution unit when dividing into the respective parallel execution units in accordance with the degree of parallelism, from the total number of executions converted into the sequential execution of the tuning target loop.
[0044]
POTTAL
= STOTAL-LOOP + (LOOP / Para) + POVER
Here, STOTAL is the total number of executions of the execution statement in the program converted to sequential execution, and is composed of a basic block execution cost and LOOP as shown in FIG. LOOP is the total number of executions of the execution statement in the loop to be tuned converted into the sequential execution, and is a total value of the execution number of execution statements in the loop in each parallel execution unit.
[0045]
PTOTAL is the total number of executions of an executable statement in a parallel execution unit, and is composed of redundant execution cost, LOOP / Para, and POVER. The redundant execution cost is the execution cost of a basic block in which each parallel execution unit performs the same processing. FIG. 6 shows a case where the redundant execution cost and the execution cost of the basic block are the same.
[0046]
Para is the degree of parallelism, and LOOP / Para is the parallel execution cost as a load of the number of executions of the in-loop executable statement in each parallel execution unit. POVER is an overhead for parallel execution and is a fixed value.
[0047]
Here, STOTAL is calculated from the count value in the count area for each parallel execution unit shown in FIG. That is, first, LOOP is calculated by calculating the total value of the number of times of execution of an in-loop executable statement in each parallel execution unit. Also, for example, when each parallel execution unit performs the same processing for the basic blocks 1 to 3, the count value of the representative parallel execution unit for the redundant basic execution of the basic blocks 1 to 3 is calculated. Then, the redundant execution cost is calculated by using the calculated LOOP, and the sum of the calculated LOOP and the redundant execution cost is obtained as STOTAL.
[0048]
As shown in FIG. 6, POVER as overhead occurs due to parallel execution, but since LOOP decreases to LOOP / Para, PTOTAL becomes a value smaller than STOTAL.
[0049]
Next, the performance improvement rate when the degree of parallelism is changed is calculated by the following equation using STOTAL and PTOTAL. The unit is a percentage.
[0050]
Performance improvement rate = {(STOTAL-PTOTAL) / STOTAL} × 100
Also, for the loop to be tuned, the execution number of execution statements in the parallel execution unit, the average rotation number of the loop, and the estimated value of the performance improvement rate with respect to the change in the parallelism are displayed. This estimate is based on the assumption that the load on the loop is constant and independent of the loop speed.
[0051]
FIG. 7 shows an example of the display format of the execution number of the execution statement of the specific loop, the average rotation number of the loop, and the estimated value of the performance improvement rate with respect to the change of the parallel degree as an example of the output of the analysis result according to the present invention.
[0052]
In the example shown in FIG. 7, when the degree of parallelism changes from 1 (sequential execution) to 2, the average rotation number (Iteration-count) of the loop changes from 80 to 40, and the total number of executions of the execution statement in the loop whose performance is to be estimated decreases. However, because the overhead of parallel execution increases, the performance is slightly degraded as compared with the case of sequential execution. However, when the degree of parallelism increases to 4,8, the average number of rotations of the loop becomes 20,10, respectively, and the performance improvement rate is improved to 14.5% and 22.2%.
[0053]
Further, in the present embodiment, after calculating the loop average rotation speed in the sequential execution, the calculated average rotation speed information of the parallel loop is used for estimating the parallel load division amount. The average rotation number of the loop in the sequential execution is obtained by multiplying the average rotation number of the measured loop by the parallel frequency, and the average rotation number 80 of the loop when the parallel frequency shown in FIG. It is an average rotation speed.
[0054]
FIG. 8 shows a method of estimating the parallel load division amount based on the average rotation speed of the sequentially executed loop. As shown in FIG. 8, the parallel load division amount is calculated by dividing the average rotation speed of the loop in the sequential execution by the parallel frequency.
[0055]
Also, the total value of the number of executions for each procedure and the balance of the number of executions in each parallel execution unit are displayed. Procedure information is displayed in the order of the number of executions. In the case of the mixed type of the memory-distributed parallelization method and the memory-sharing parallelization method, each parallel execution unit of the memory-distributed parallelization method is displayed, and the memory-shared parallelism is used as detailed information of the parallel execution unit. Display of conversion.
[0056]
FIG. 9 shows an example of a display format of the procedure execution count total value. FIG. 10 shows an example of a form of a balance display of the number of procedure executions in each parallel execution unit.
[0057]
For example, in the procedure of “main._OMP_1_”, the total execution count (Total-count) is 1042, and the ratio (Run) of the total execution count to the entire program is 0.88%, and the procedure execution count is 0.88%. 9 is eight times.
[0058]
This procedure “main._OMP_1_” is executed in parallel by four threads, and FIG. 10 shows a deviation ratio of the number of executions of the corresponding parallel execution unit to the average number of executions. From the example of FIG. 10, it can be seen that the load of the fourth thread (Thread 3) is + 11%, which is larger than that of the other threads. Therefore, after recognizing the balance display shown in FIG. 10, the user may need to reduce the load on the fourth thread in order to perform tuning to equalize the load balance divided into each parallel execution unit. Understand.
[0059]
For example, when the movement of the loop changes depending on the number of revolutions, such as in a loop including an if statement, the load usually varies for each parallel execution unit. The balance display shown in FIG. 10 is a useful information display in the sense that the user can recognize the balance display and change the load balance.
[0060]
Also, the number of times of execution of the loop is displayed in the same manner as in the case of the procedure. FIG. 11 shows an example of the display format of the loop execution count total value. FIG. 12 shows an example of a form of a balance display of the number of loop executions in each parallel execution unit.
[0061]
In the display shown in FIG. 11, Mark is identification of parallel execution or sequential execution, and indicates that OMP is parallel execution and S is sequential execution.
[0062]
In addition, the number of executions and the number of parallel executions of each execution statement are extracted from the execution information file, and the values are displayed in association with the source code. The location of the source code is notified by an argument to this analysis process. FIGS. 13 and 14 show an example of a format for displaying the number of executions and the degree of parallelism for each execution statement.
[0063]
The processing performed by the execution number measurement result analysis means 13 and the like can be realized by a computer and a software program, and the program is executed by a computer-readable recording medium such as a portable medium memory, a semiconductor memory, or a hard disk. And read it from there to be executed by the computer.
[0064]
As can be understood from the above, the features of the embodiment of the present invention are as follows.
[0065]
(Supplementary Note 1) A method of analyzing the number of executions of a parallel execution program,
A translation process of inserting an instruction for accumulating the number of executions at least for each parallel execution unit into each procedure when translating the parallel execution program to be analyzed;
By executing the translated parallel execution program, accumulating the number of executions of instructions set at the time of translation for each parallel unit;
Analyzing an execution count measurement result of the parallel execution unit and outputting at least an analysis result indicating a load distribution in the parallel execution.
A method for analyzing the number of executions of a parallel execution program, characterized in that:
[0066]
(Supplementary note 2) In the method for analyzing the number of executions of a parallel execution program according to supplementary note 1,
In the translation step, a call instruction of a function having a function of securing a count area for accumulating the number of executions of each execution statement in the procedure is inserted at a start position of each procedure, and a count instruction for accumulating the number of executions is added to the procedure. Is inserted at the start position of each basic block, the start position of each loop, and the position of the first executable statement in each loop, which are executable statement strings whose execution order is not likely to change.
In the measurement step, the number of executions is accumulated in the count area by a count instruction appearing during execution of the instruction of each procedure, and the number of executions of the count area is recorded in an execution information file at the end of the entire program.
In the analysis process, the result of measuring the number of executions of each parallel execution unit is extracted from the execution information file, and the average rotation number of each loop, which is the average of the number of executions of the in-loop execution statement in each loop execution, is calculated for each parallel execution unit. Do
A method for analyzing the number of executions of a parallel execution program, characterized in that:
[0067]
(Supplementary note 3) In the method for analyzing the number of executions of a parallel execution program according to supplementary note 2,
The analysis process includes:
Calculating the total number of executions of the executable statement in the program converted to sequential execution based on the execution number measurement result of each parallel execution unit extracted from the execution information file;
Based on the calculated total number of executions of the executable statement in the program converted to the sequential execution, the total number of executions of the in-loop executable statement in each parallel execution unit, the parallelism, and the overhead for the parallel execution, Calculating the total number of executions of the executable statement in the parallel execution unit when dividing the total number of executions of the executable statement in the program converted into the serial execution into each parallel execution unit according to the parallelism frequency;
Based on the calculated total number of executions of the execution statement in the parallel execution unit and the calculated total number of executions of the execution statement in the program converted to the sequential execution, a performance improvement rate when the parallelism is changed is calculated. Outputting a calculation result
A method for analyzing the number of executions of a parallel execution program, characterized in that:
[0068]
(Supplementary note 4) In the method for analyzing the number of executions of the parallel execution program according to supplementary note 2 or 3,
The analysis process includes:
Based on the execution count measurement result of each parallel execution unit extracted from the execution information file, the execution count of the execution statement in each procedure and the execution count of the execution statement in the parallel execution unit of each loop in the program executed in parallel Calculating and outputting the calculation result in descending order of the number of times;
Outputting information indicating the balance of the number of executions between the parallel execution units.
A method for analyzing the number of executions of a parallel execution program, characterized in that:
[0069]
(Supplementary note 5) In the method for analyzing the number of times of execution of a parallel execution program according to any one of supplementary notes 2 to 4,
The analysis process includes:
A step of calculating the total number of executions and the parallelism for each execution statement in association with the source code based on the execution count measurement result of each parallel execution unit extracted from the execution information file, and outputting the calculation result
A method for analyzing the number of executions of a parallel execution program, characterized in that:
[0070]
(Supplementary Note 6) A program for causing a computer to execute a method of analyzing the number of executions of a parallel execution program,
At the time of translating the parallel execution program to be analyzed, a translation process for inserting an instruction for accumulating the number of executions at least for each parallel execution unit into each procedure;
By executing the translated parallel execution program, a measurement process for accumulating the number of executions of the instruction set at the time of translation for each parallel unit;
An analysis process of analyzing the execution count measurement result of the parallel execution unit and outputting an analysis result indicating at least a load distribution in the parallel execution;
An execution count analysis program for a parallel execution program to be executed by a computer.
[0071]
(Supplementary Note 7) A recording medium storing a program for causing a computer to execute the method of analyzing the number of executions of a parallel execution program,
At the time of translating the parallel execution program to be analyzed, a translation process for inserting an instruction for accumulating the number of executions at least for each parallel execution unit into each procedure;
By executing the translated parallel execution program, a measurement process for accumulating the number of executions of the instruction set at the time of translation for each parallel unit;
An analysis process of analyzing the execution count measurement result of the parallel execution unit and outputting an analysis result indicating at least a load distribution in the parallel execution;
A recording medium that stores a program for analyzing the number of executions of a parallel execution program to be executed by a computer.
[0072]
【The invention's effect】
With the method for analyzing the number of times of execution of the parallel execution program according to the present invention, the load distribution situation can be accurately confirmed by measuring the average rotation number of the loop for each parallel execution unit. In addition, it is possible to obtain a specific numerical reference for determining the number of load divisions. In addition, by displaying the predicted value of the parallel effect when the parallel frequency is increased, it is possible to determine a measure of performance improvement and contribute to improvement of tuning efficiency.
[0073]
Also, by displaying the number of executions of procedures and loops in descending order, extraction of parallelization candidates becomes easy. Further, by displaying the balance between the parallel execution units, the load status between the parallel execution units can be confirmed. In addition, by displaying the number of executions and the parallelism of each execution statement in association with the source code, it is easy to recognize the parallel execution status, and the feedback of the performance improvement to the source code is promoted.
[Brief description of the drawings]
FIG. 1 is a diagram illustrating an example of a hardware configuration that implements a method for analyzing the number of executions of a parallel execution program according to the present invention.
FIG. 2 is a diagram showing an outline of a method for analyzing the number of executions of a parallel execution program according to the present invention.
FIG. 3 is a diagram illustrating an example of an execution number analysis processing flow of a parallel execution program according to the present invention.
FIG. 4 is a diagram showing an example of insertion of a count command at the time of translation.
FIG. 5 is a diagram illustrating an example of a method of measuring the number of executions of an analysis target program.
FIG. 6 is a diagram showing the relationship between the total number of executions of a statement executed in a program (STOTAL) converted to sequential execution and the total number of executions of a statement executed in a parallel execution unit (PTOTAL).
FIG. 7 is a diagram illustrating an example of a format of an execution value execution count of a specific loop, an average rotation speed of a loop, and an estimated value of a performance improvement rate with respect to a change in a parallelism;
FIG. 8 is a diagram illustrating a method of estimating a parallel load split amount based on an average rotation speed of a sequentially executed loop.
FIG. 9 is a diagram illustrating an example of a display format of a procedure execution count total value.
FIG. 10 is a diagram showing an example of a format of a balance display of the number of times of procedure execution in each parallel execution unit.
FIG. 11 is a diagram illustrating an example of a display format of a total number of times of loop execution.
FIG. 12 is a diagram illustrating an example of a form of a balance display of the number of loop executions in each parallel execution unit.
FIG. 13 is a diagram showing an example of a format of display of the number of executions and the number of parallelisms for each execution statement.
FIG. 14 is a diagram showing an example of a format of a display of the number of executions and the number of parallelisms in execution statement units.
FIG. 15 is a diagram illustrating an example of designation of parallelization;
[Explanation of symbols]
1 Parallel execution program execution result analyzer
2 Node A
3 Node B
4 Execution information file
5 Express bus
6 Analysis result output device
7, 7 'analysis target program
11 Analysis target program translation means
12 means for measuring the number of executions of the program to be analyzed
13. Execution count measurement result analysis means
21,31 CPU
22, 32 memories

Claims (5)

An execution number analysis method for a parallel execution program,
A translation process of inserting an instruction for accumulating the number of executions at least for each parallel execution unit into each procedure when translating the parallel execution program to be analyzed;
By executing the translated parallel execution program, accumulating the number of executions of instructions set at the time of translation for each parallel unit;
Analyzing the execution count measurement result of the parallel execution unit and outputting at least an analysis result indicating a load distribution in the parallel execution. 2. The method according to claim 1, wherein the number of executions of the parallel execution program is analyzed.
In the translation step, a call instruction of a function having a function of securing a count area for accumulating the number of executions of each execution statement in the procedure is inserted at a start position of each procedure, and a count instruction for accumulating the number of executions is added to the procedure. Is inserted at the start position of each basic block, the start position of each loop, and the position of the first executable statement in each loop, which are executable statement strings whose execution order is not likely to change.
In the measurement step, the number of executions is accumulated in the count area by a count instruction appearing during execution of the instruction of each procedure, and the number of executions of the count area is recorded in an execution information file at the end of the entire program.
In the analysis process, the result of measuring the number of executions of each parallel execution unit is extracted from the execution information file, and the average rotation number of each loop, which is the average of the number of executions of the in-loop execution statement in each loop execution, is calculated for each parallel execution unit. A method for analyzing the number of executions of a parallel execution program. 3. The method for analyzing the number of executions of a parallel execution program according to claim 2,
The analysis process includes:
Calculating the total number of executions of the executable statement in the program converted to sequential execution based on the execution number measurement result of each parallel execution unit extracted from the execution information file;
Based on the calculated total number of executions of the executable statement in the program converted to the sequential execution, the total number of executions of the in-loop executable statement in each parallel execution unit, the parallelism, and the overhead for the parallel execution, Calculating the total number of executions of the executable statement in the parallel execution unit when dividing the total number of executions of the executable statement in the program converted into the serial execution into each parallel execution unit according to the parallelism frequency;
Based on the calculated total number of executions of the execution statement in the parallel execution unit and the calculated total number of executions of the execution statement in the program converted into the sequential execution, a performance improvement rate when the parallelism is changed is calculated. And analyzing the number of executions of the parallel execution program. A method for analyzing the number of executions of a parallel execution program according to claim 2 or 3,
The analysis process includes:
Based on the execution count measurement result of each parallel execution unit extracted from the execution information file, the execution count of the execution statement in each procedure and the execution count of the execution statement in the parallel execution unit of each loop in the program executed in parallel Calculating and outputting the calculation result in descending order of the number of times;
Outputting information indicating a balance of the number of executions among the parallel execution units. A program for causing a computer to execute a method for analyzing the number of executions of a parallel execution program,
At the time of translating the parallel execution program to be analyzed, a translation process for inserting an instruction for accumulating the number of executions at least for each parallel execution unit into each procedure;
By executing the translated parallel execution program, a measurement process for accumulating the number of executions of the instruction set at the time of translation for each parallel unit;
An analysis process of analyzing the execution count measurement result of the parallel execution unit and outputting an analysis result indicating at least a load distribution in the parallel execution;
An execution count analysis program for a parallel execution program to be executed by a computer.

Images