JP7419843B2 - parallel processing device - Google Patents

Description

The present invention relates to a parallel processing device.

A technique for determining parallel processing parameters of a parallel processing program is disclosed. For example, in Patent Document 1, a processing unit determines the number of nodes and the number of processes recommended for a first program to be evaluated. Specifically, the processing unit executes the first program according to the first sample points indicating the number of nodes and the number of processes, and calculates the first statistic based on the rate of change in the evaluation value between each of the first sample points. do. Then, the processing unit generates the first sample points 4 until the first statistic becomes equal to or less than the first threshold. Further, the processing unit adds a second sample point within a predetermined distance from the first sample point, and calculates second statistics for each first sample point using the evaluation value when the first program is executed according to the second sample point. Calculate the amount. The processing unit then generates second sample points until the second statistic becomes equal to or less than the second threshold. The processing unit then determines the recommended number of nodes and processes by complementing the evaluation values of the first sample point and the second sample point.

Furthermore, when a program having a double loop is executed in parallel, it is necessary to set appropriate parallel processing parameters. Parallel processing parameters of the thread parallelization standard OpenMP include scheduling method, chunk size, and processor core allocation. Conventionally, for each combination of parallel processing parameters, the parameters for the combination were set in a program, the program was executed, and the execution results were referenced to adjust the parallelization parameters to be appropriate. Alternatively, the parallelization parameters were adjusted based on the experience of experienced people.

Unexamined Japanese Patent Publication No. 2018-120387 Japanese Patent Application Publication No. 2016-9972

However, when a program having a loop is executed in parallel, there is a problem in that it is difficult to quickly estimate optimal parallel processing parameters to be set in the program.

For example, in a technique for determining the recommended number of nodes and processes, a first program to be evaluated is executed many times according to a first sample point indicating the number of nodes and processes; Determine the recommended number of nodes and processes. Further, the only parallel processing parameters to be determined are the number of nodes and the number of processes. Therefore, it is difficult to quickly estimate optimal parallel processing parameters including scheduling method, chunk size, and processor core allocation.

Furthermore, when determining parallelization parameters when a program having a double loop is executed in parallel, the program is executed for each combination of parallel processing parameters to find the optimal parallel processing parameters. Therefore, even in such a case, it is difficult to estimate optimal parallel processing parameters at high speed.

One aspect of the present invention is to quickly estimate optimal parallel processing parameters when executing a program having a loop in parallel processing.

In one aspect, the parallel processing device includes an execution unit that executes an optimization target program having a double loop only once in parallel processing, and a result executed by the execution unit, in which a predetermined parallelization parameter is set. A calculation unit that calculates the feature value of the program to be optimized from the above, and a database that shows the correlation between the feature value and the optimal parallelization parameter in the sample program having a double loop, and the calculation unit calculates the feature value. and an extraction unit that extracts an optimal parallelization parameter corresponding to the calculated feature value.

According to one embodiment, optimal parallel processing parameters when executing a program having a loop in parallel processing can be estimated at high speed.

FIG. 1 is a block diagram showing the functional configuration of a parallel processing device according to an embodiment. FIG. 2 is a diagram showing an example of a program. FIG. 3 is a diagram illustrating an example of a relational database according to the embodiment. FIG. 4 is a diagram showing an example of hardware that executes a program. FIG. 5 is a diagram illustrating an example of allocation of processor cores. FIG. 6 is a diagram showing an example of execution of a DO loop. FIG. 7 is a diagram showing an example of optimization parameter extraction. FIG. 8 is a diagram showing an example of an optimized program. FIG. 9 is a diagram illustrating an example of a flowchart of generation processing according to the embodiment. FIG. 10 is a diagram illustrating an example of a flowchart of optimization parameter extraction processing according to the embodiment. FIG. 11 is a diagram illustrating an example of a flowchart of the program main execution process according to the embodiment. FIG. 12 is a diagram illustrating an example of a computer that executes a parallel processing program.

Embodiments of the parallel processing device disclosed in the present application will be described in detail below with reference to the drawings. Note that the present invention is not limited to the examples.

[Functional configuration of parallel processing device according to embodiment]
FIG. 1 is a block diagram showing the functional configuration of a parallel processing device according to an embodiment. A parallel processing device 1 shown in FIG. 1 selects optimal parallelization parameters for an application program that can be parallelized at multiple levels, and tunes execution performance. In the embodiment, it is assumed that OpenMP, which is a thread parallelization standard, is applied as an application program that can perform multiple levels of parallelization. Further, in the embodiment, a case will be described in which, for example, an appropriate parallelization parameter is selected when an application program having a double loop is executed in parallel processing. Note that the OpenMP standard is based on, for example, the description of "OpenMP Application Program Interface Version 5.0", and detailed explanation will be omitted.

The parallel processing device 1 includes a control section 10 and a storage section 20. The control unit 10 corresponds to an electronic circuit such as a CPU (Central Processing Unit). The control unit 10 has an internal memory for storing programs and control data that define various processing procedures, and executes various processes using these. The control unit 10 includes a generation unit 11 , a temporary program execution unit 12 , a feature value calculation unit 13 , an optimization parameter extraction unit 14 , and a main program execution unit 15 . Note that the program temporary execution unit 12 is an example of an execution unit. The feature value calculation unit 13 is an example of a calculation unit. The optimization parameter extraction unit 14 is an example of an extraction unit. The generation unit 11 is an example of a generation unit.

The storage unit 20 is, for example, a semiconductor memory element such as a RAM (Random Access Memory) or a flash memory, or a storage device such as a hard disk or an optical disk. The storage unit 20 has a reference program 21 and a relational database 22.

The reference program 21 refers to a program that has feature values for each combination of a plurality of execution feature values of the program that affect parallelization parameters. The standard program 21 has variables for setting parallelization parameters and has a double loop. Here, a combination of multiple execution feature values of a program is referred to as an "application execution feature value set." A reference program 21 is created for each application execution feature value set.

Here, an example of a program used in the embodiment will be explained with reference to FIG. 2. FIG. 2 is a diagram showing an example of a program. FIG. 2 shows an example of a Fortran language program. As shown in FIG. 2, the program performs two levels of parallelization of the DO loop. L1 and U1 in the program are constants for the variation range of the first-dimensional loop. L2 and U2 are constants of the variation range of the second-dimensional loop.

Such a program has variables for setting parallelization parameters of OpenMP, which is a thread parallelization standard. Parallelization parameters include variables P1, P2, D1, D2, C1, and C2. Variables P1 and P2 are set to values indicating allocation of processor cores to DO loops. Variables D1 and D2 are set to values indicating the scheduling method. Variables C1 and C2 are set to the number of chunks. The scheduling method can be set to, for example, static, dynamic, or guided. For example, any integer can be set for the chunk. Any integer is, for example, 1, 8 or 32. Assigning processor cores to the DO loop, taking 48 cores as an example, (1, 48), (2, 24), (3, 16), (4, 12), (6, 8), (8 , 6), (12, 4), (16, 3), (24, 2), and (48, 1) can be set.

Such a program has a parallel do syntax and executes variable range do loops in parallel. Parallelism here refers to thread parallelism within a program, not process parallelism. The parallel do construct assigns a portion of the do loop iterations to threads according to processor core assignments.

In the above example, there are 810 combinations of parallelization parameters that can be set for such a program. That is, since there are three types of scheduling methods and three types of chunks, there are 81 (=3×3×3×3) ways when applied to a two-level DO loop. In addition, since there are 10 types of allocation of processor cores to DO loops, there are 810 (=10×81) combinations of parallelization parameters that can be set in a program.

In the embodiment, the parallel processing device 1 extracts an optimal parallelization parameter indicating a combination with the shortest execution time from among combinations of parallelization parameters that can be set in a program.

Returning to FIG. 1, the application execution feature value set will now be described. An application execution feature value set is a combination of multiple execution feature values of a program that affect parallelization parameters. For example, the following x, y, and z can be considered as a plurality of program execution characteristic values that affect the parallelization parameter.
x) Average value of the number of instructions executed in the processing section for each iteration of each DO loop y) Relative standard deviation value of the number of instructions executed for the processing section for each iteration of each DO loop z) Average value of the number of instructions executed for the processing section for each iteration of each DO loop Average cache reuse rate

The reason why x is applied as a program execution characteristic value that affects parallelization parameters is as follows. This is because the number of instructions executed in a processing part affects, for example, the scheduling method, which is one of the parallelization parameters. That is, this is to determine whether the dynamic scheduling method, which is adopted when the processing cost (overhead) is large, can be adopted.

The reason why y is applied as a program execution characteristic value that affects parallelization parameters is as follows. This is because the degree of load balance of the processing parts affects, for example, the scheduling method, which is one of the parallelization parameters. That is, the static scheduling method is used when there is no variation in the load balance of the processing parts, and the dynamic scheduling method is used when there is variation in the load balance.

The reason why z is applied as a program execution characteristic value that affects parallelization parameters is as follows. This is because the cache reuse rate affects, for example, the number of chunks, which is one of the parallelization parameters. That is, when the cache reuse rate is high, a large number of chunks is adopted in order to promote cache reuse by increasing the number of chunks.

In the embodiment, the parallel processing device 1 calculates these three types of execution feature values in the first-dimensional DO loop and the second-dimensional DO loop, and calculates the combination of six execution feature values as the application execution feature. Define as a value set.

The relational database 22 holds the correspondence between application execution feature value sets and optimal parallelization parameters. The relational database 22 is generated by the generation unit 11, which will be described later. Note that an example of the relational database 22 will be described later.

The generation unit 11 generates a relational database 22.

For example, the generation unit 11 generates a plurality of reference programs 21 that change each element of the application execution feature value set in the following increments. The average number of instructions executed in the processing portion for each repetition of each DO loop is 1 to 10 times the number of instructions executed in the runtime library in the case of the dynamic scheduling method. That is, the average value of the number of executed instructions in the processing portion for each repetition of each DO loop takes a value of 1, 2, . . . , 10 for each DO loop. The relative standard deviation value of the number of executed instructions in the processing portion for each repetition of each DO loop is set from 0.1 to 1.0 in 0.1 increments. That is, the relative standard deviation value of the number of executed instructions in the processing portion for each repetition of each DO loop takes a value of 0.1, 0.2, . . . , 1.0 for each DO loop. The average value of the cache reuse rate of the processing portion for each repetition of each DO loop is 0.1 to 1.0 in 0.1 increments. That is, the average value of the cache reuse rate of the processing portion for each repetition of each DO loop takes a value of 0.1, 0.2, . . . , 1.0 for each DO loop. In this example, the application execution feature value set ultimately becomes 10 ⁶ (=1 million) pieces. The generation unit 11 generates each reference program 21 having one million application execution feature value sets.

Then, the generation unit 11 sets and executes a plurality of sets of parallelization parameters for the reference program 21 having one application execution feature value set. As an example, assuming that there are 810 combinations of parallelization parameters that can be set in a program, the generation unit 11 sets and executes each of the 810 parallelization parameters for one reference program 21.

Then, the generation unit 11 determines the set of parallelization parameters with the shortest execution time as the optimal parallelization parameters based on the plurality of execution results. Then, the generation unit 11 associates the application execution feature value set in the standard program 21 with the optimal parallelization parameter and adds them to the relational database 22. Then, the generation unit 11 similarly determines the optimal parallelization parameters for the reference program 21 having the application execution feature value sets for other application execution feature value sets, and adds them to the relational database 22.

Here, an example of the relational database 22 will be explained with reference to FIG. 3. FIG. 3 is a diagram illustrating an example of a relational database according to the embodiment. The relational database 22 shown in FIG. 3 associates the application execution feature value set (X1, Y1, Z1, X2, Y2, Z2) with the optimal parallelization parameters (P1, P2, D1, D2, C1, C2). memorize it.

The application execution feature value set (X1, Y1, Z1, X2, Y2, Z2) corresponds to the above x, y, z of the first-dimensional DO loop and the above x, y, z of the second-dimensional DO loop. . That is, X1 indicates x of the first-dimensional DO loop, that is, the average value of the number of instructions executed in the processing portion for each repetition of the first-dimensional DO loop. Y1 indicates y of the first-dimensional DO loop, that is, the relative standard deviation value of the number of executed instructions of the processing portion for each repetition of the first-dimensional DO loop. Z1 indicates z of the first-dimensional DO loop, that is, the average value of the cache reuse rate of the processing portion for each iteration of the first-dimensional DO loop. X2 indicates x of the second-dimensional DO loop, that is, the average value of the number of instructions executed in the processing portion for each repetition of the second-dimensional DO loop. Y2 indicates y of the second-dimensional DO loop, that is, the relative standard deviation value of the number of executed instructions of the processing portion for each repetition of the second-dimensional DO loop. Z2 indicates z of the second-dimensional DO loop, that is, the average value of the cache reuse rate of the processing portion for each iteration of the second-dimensional DO loop. Here, X1 and X2 take values of 1, 2, . . . , 10. Y1 and Y2 take values of 0.1, 0.2, . . . , 1.0. Z1 and Z2 take values of 0.1, 0.2, . . . , 1.0.

The optimal parallelization parameters (P1, P2, D1, D2, C1, C2) correspond to the optimal parallelization parameters in the standard program 21 having the application execution feature value set (X1, Y1, Z1, X2, Y2, Z2). do. P1 and P2 are values indicating allocation of processor cores to DO loops. D1 is a value indicating the scheduling method of the first-dimensional DO loop. D2 is a value indicating the scheduling method of the second-dimensional DO loop. C1 is the number of chunks of the first-dimensional DO loop. C2 is the number of chunks of the second-dimensional DO loop.

As an example, when the application execution feature value set is (1, 0.1, 0.1, 1, 0.1, 0.1), the optimal parallelization parameter is (6, 8, static, static, 1 , 1) are stored. Furthermore, when the application execution feature value set is (10, 0.1, 0.1, 1, 0.1, 0.1), the optimal parallelization parameters are (6, 8, dynamic, static, 1, 1) is stored.

The program temporary execution unit 12 temporarily executes an optimization target program from which optimization parameters are to be extracted. For example, the program temporary execution unit 12 temporarily executes the optimization target program with predetermined parallelization parameters set only once in order to find the characteristics of the optimization target program. The predetermined parallelization parameters (P1, P2, D1, D2, C1, C2) set in the optimization target program are, for example, (6, 8, static, static, 1, 1). However, the predetermined parallelization parameters (P1, P2, D1, D2, C1, C2) are not limited to these.

The feature value calculation unit 13 calculates each feature value of the application execution feature value set when executed by the program temporary execution unit 12 . Each specific value of the application execution feature value set at the time of execution may be obtained using a performance information acquisition tool such as a specific profile. As a performance information acquisition tool such as a specific profile, for example, an application development support tool for the supercomputer "K" (see "Performance profile of the supercomputer "K") may be used.

The optimization parameter extraction unit 14 refers to the relational database 22 and extracts the optimal parallelization parameter corresponding to the application execution feature value set calculated by the feature value calculation unit 13. For example, the optimization parameter extraction unit 14 obtains, from the relational database 22, an application execution feature value set that is closest in distance to the application execution feature value set calculated by the feature value calculation unit 13. This is because the application execution feature value set stored in the relational database 22 is discrete. Then, the optimization parameter extraction unit 14 extracts the optimal parallelization parameter corresponding to the acquired application execution feature value set from the relational database 22. Thereby, the optimization parameter extraction unit 14 can quickly estimate an optimal parallelization parameter when executing an optimization target program having a loop in parallel processing.

The program execution unit 15 executes the optimization target program. For example, the main program execution unit 15 executes the optimization target program in which the optimal parallelization parameters extracted by the optimization parameter extraction unit 14 are set.

[Example of hardware that executes the program]
FIG. 4 is a diagram showing an example of hardware that executes a program. As shown in FIG. 4, it is assumed that a computer has 48 processor cores (computing cores) that share one memory. Each computational core has a cache to reduce the cost of accessing memory. Then, a UMA group is generated for each of the 12 calculation cores. The four UMA groups are connected by NUMA (Non-Uniform Memory Access). In addition, in the example of FIG. 4, the case where a computer has 48 calculation cores was explained, but it is not limited to this.

[Example of processor core allocation]
FIG. 5 is a diagram illustrating an example of allocation of processor cores. That is, an example of the allocation of processor cores to the DO loop represented by the variables P1 and P2 of the parallelization parameters is shown. The example in FIG. 5 is a case of 48 processor cores (computation cores).

The upper diagram of FIG. 5 shows a case where the allocation of 48 processor cores is (12,4). In this case, for each core, the program executes the first dimension (first layer) DO loop in 12 divisions, and the second dimension (second layer) DO loop in four divisions.

The lower diagram in FIG. 5 shows a case where the allocation of 48 processor cores is (8,6). In this case, for each core, the program executes the first dimension (first layer) DO loop in eight divisions, and the second dimension (second layer) DO loop in six divisions.

[Execution example of DO loop]
FIG. 6 is a diagram showing an example of execution of a DO loop. In the example of FIG. 6, L1 and L2 of the program shown in FIG. 2 are "1" and U1 and U2 are "100". That is, the variation ranges L1 and U1 of the first dimension (first layer) loop are 1,100. The variation ranges L2 and U2 of the second dimension (second layer) loop are 1,100. Further, this is a case where the processor core allocation (P1, P2) of the optimization parameters is (8, 6).

here, The processor core indicated by CORE#0 is The fluctuation range of the first DO loop (1, 13) Variation range of the second DO loop (1, 17). The processor core indicated by CORE #1 is Variation range of the first DO loop (14, 26) The fluctuation range of the second DO loop (1, 17). The processor core indicated by CORE#2 is Variation range of the first DO loop (27, 39) The fluctuation range of the second DO loop (1, 17). ...The processor core indicated by CORE#7 is Variation range of the first DO loop (92, 100) The fluctuation range of the second DO loop (1, 17). Also, The processor core indicated by CORE #8 is The fluctuation range of the first DO loop (1, 13) Variation range of the second DO loop (18, 34). The processor core indicated by CORE #16 is The fluctuation range of the first DO loop (1, 13) Variation range of the second DO loop (35, 51). The processor core indicated by CORE #24 is The fluctuation range of the first DO loop (1, 13) Variation range of the second DO loop (52, 68). The processor core indicated by CORE #32 is The fluctuation range of the first DO loop (1, 13) Variation range of the second DO loop (69, 85). The processor core indicated by CORE #40 is The fluctuation range of the first DO loop (1, 13) Variation range of the second DO loop (86, 100). in this way, According to processor core allocation, Allocate part of the iterative processing of the do loop to a thread.

[Example of optimization parameter extraction]
FIG. 7 is a diagram showing an example of optimization parameter extraction. As shown in FIG. 7, the relational database 22 generated by the generation unit 11 is displayed. Note that it is assumed that the application execution feature value set corresponding to the optimization target program has been calculated by the feature value calculation unit 13.

The optimization parameter extraction unit 14 refers to the relational database 22 and obtains an application execution feature value set that is closest in distance to the application execution feature value set calculated by the feature value calculation unit 13 . Here, (5, 0.8, 0.2, 1, 0.2, 0.1) is acquired as the application execution feature value set. Then, the optimization parameter extraction unit 14 refers to the relational database 22 and extracts the optimal parallelization parameter corresponding to the acquired application execution feature value set. Here, (12, 4, guided, static, 32, 1) is extracted as the optimal parallelization parameter.

[An example of an optimized program]
FIG. 8 is a diagram showing an example of an optimized program. Here, it is assumed that (12, 4, guided, static, 32, 1) is extracted as the optimal parallelization parameter of the optimization target program. As shown by the symbol p1, 12.4 is set as a value indicating the allocation of the processor core to the DO loop. As shown by symbol p2, "guided" is set as the scheduling method for the first-dimensional DO loop. As shown by symbol p3, "static" is set as the scheduling method for the second-dimensional DO loop. As shown by symbol p4, "32" is set as the number of chunks of the first-dimensional DO loop. As shown by symbol p5, "1" is set as the number of chunks of the second-dimensional DO loop.

Then, the program execution unit 15 executes the optimization target program for which the optimal parallelization parameters have been set.

[Example of flowchart of generation process]
FIG. 9 is a diagram illustrating an example of a flowchart of generation processing according to the embodiment. It is assumed that the application execution feature value set is (X1, Y1, Z1, X2, Y2, Z2) shown in FIG.

As shown in FIG. 9, the generation unit 11 generates a plurality of reference programs 21 that can take values arranged in a grid in the six-dimensional space of the application execution feature value set (step S11).

Then, the generation unit 11 executes all combinations of parallelization parameters in each reference program 21. The generation unit 11 stores the parallelization parameter with the shortest execution time as the optimal parallelization parameter in the relational database 22 in association with the application execution feature value set corresponding to the reference program 21 (step S12).

[Example of flowchart of optimization parameter extraction process]
FIG. 10 is a diagram illustrating an example of a flowchart of the optimization parameter extraction process according to the embodiment.

As shown in FIG. 10, the program temporary execution unit 12 that has received the optimization target program executes the received optimization target program once. Then, the feature value calculation unit 13 obtains an application execution feature value set (step S21). Each specific value of the application execution feature value set during execution may be obtained using a performance information acquisition tool such as a specific profile.

Then, the optimization parameter extraction unit 14 rounds the acquired application execution feature value set (step S22).

Then, the optimization parameter extraction unit 14 refers to the relational database 22 using the rounded application execution feature value set as a key and extracts the optimal parallelization parameter (step S23).
[An example of a flowchart of program execution processing]
FIG. 11 is a diagram illustrating an example of a flowchart of the program main execution process according to the embodiment.

The program main execution unit 15 sets the optimum parallelization parameter to the optimization target program and executes it (step S31).

[Effects of Examples]
According to the above embodiment, the parallel processing device 1 executes the optimization target program having a double loop only once in parallel processing, for which a predetermined parallelization parameter has been set. The parallel processing device 1 calculates feature values of the program to be optimized from the executed results. The parallel processing device 1 refers to a database showing the correlation between feature values and optimal parallelization parameters in sample programs having double loops, and extracts optimal parallelization parameters corresponding to the calculated feature values. . According to this configuration, when the parallel processing device 1 executes an optimization target program having a double loop in parallel processing, it is possible to quickly obtain the optimal parallelization parameters by executing the program only once. can.

Further, according to the above embodiment, the parallel processing device 1 sets the characteristic values of the sample program having characteristic values and the optimal Generate a database showing the correlation with parallelization parameters. The parallel processing device 1 refers to the generated database and extracts the optimal parallelization parameter corresponding to the calculated feature value of the program to be optimized. According to this configuration, the parallel processing device 1 can quickly determine the optimal parallelization parameter by generating a correlation between the feature value and the parallelization parameter in a program having a double loop.

Further, according to the above embodiment, the parallel processing device 1 generates a database showing the following correlations. The correlations are based on multiple feature values that indicate the average number of executed instructions, the relative standard deviation of the number of executed instructions, and the average cache reuse rate for each loop, as well as processor core allocation and parallelization scheduling for each loop. This is a correlation between the method and a plurality of parallelization parameters indicating the number of chunks of the parallelization scheduling method. According to this configuration, the parallel processing device 1 can quickly determine the optimal parallelization parameter by generating a correlation between the feature value and the parallelization parameter in a program having a double loop.

[others]
In the above embodiment, the Fortran language was used as an example of the program. However, the program may be in C language or C++ language, as long as it can be handled by the OpenMP standard.

Further, in the above embodiment, the OpenMP standard, which is a thread parallelization standard, is used as an example of an application program that can perform multiple levels of parallelization. However, the present invention is not limited to the Open standard, and any predetermined standard that is a thread parallelization standard may be used as an application program that can perform multiple levels of parallelization.

Further, each component of the illustrated parallel processing device 1 does not necessarily need to be physically configured as illustrated. In other words, the specific manner of distributing and integrating the parallel processing device 1 is not limited to what is shown in the diagram, and all or part of it can be functionally or physically distributed in arbitrary units depending on various loads and usage conditions. It can be configured in a distributed/integrated manner. For example, the feature value calculation unit 13 and the optimization parameter extraction unit 14 may be integrated as one unit. Further, the generation unit 11 may be distributed into a first generation unit that generates the reference program 21 and a second generation unit that generates the relational database 22. Further, the storage unit 20 may be connected as an external device to the parallel processing device 1 via a network.

Moreover, the various processes described in the above embodiments can be realized by executing a program prepared in advance on a computer such as a personal computer or a workstation. Therefore, an example of a computer that executes an analysis program that implements the same functions as the parallel processing device 1 shown in FIG. 1 will be described below. FIG. 12 is a diagram illustrating an example of a computer that executes a parallel processing program.

As shown in FIG. 12, the computer 200 includes a CPU 203 that executes various calculation processes, an input device 215 that receives data input from a user, and a display control unit 207 that controls a display device 209. The computer 200 also includes a drive device 213 that reads programs and the like from a storage medium, and a communication control unit 217 that exchanges data with other computers via a network. Further, the computer 200 includes a memory 201 that temporarily stores various information, and an HDD 205. The memory 201, CPU 203, HDD 205, display control section 207, drive device 213, input device 215, and communication control section 217 are connected via a bus 219.

The drive device 213 is, for example, a device for the removable disk 211. The HDD 205 stores a parallel processing program 205a and parallel processing related information 205b.

The CPU 203 reads the parallel processing program 205a, develops it in the memory 201, and executes it as a process. Such processes correspond to each functional unit of the parallel processing device 1. The parallel processing related information 205b corresponds to the reference program 21 and the relational database 22. For example, the removable disk 211 stores information such as the parallel processing program 205a.

Note that the parallel processing program 205a does not necessarily have to be stored in the HDD 205 from the beginning. For example, the program is stored in a "portable physical medium" such as a flexible disk (FD), CD-ROM, DVD disk, magneto-optical disk, or IC card that is inserted into the computer 200. Then, the computer 200 may read out and execute the parallel processing program 205a from these programs.

1 Parallel processing device 10 Control unit 11 Generation unit 12 Program temporary execution unit 13 Feature value calculation unit 14 Optimization parameter extraction unit 15 Program main execution unit 20 Storage unit 21 Reference program 22 Relational database

Claims (3)

an execution unit that executes an optimization target program having a double loop only once in parallel processing, with predetermined parallelization parameters set;
a calculation unit that calculates a feature value of the optimization target program from the result executed by the execution unit;
an extraction unit that refers to a database showing the correlation between feature values and optimal parallelization parameters in a sample program having a double loop, and extracts an optimal parallelization parameter corresponding to the feature value calculated by the calculation unit; and,
A parallel processing device characterized by having: Generates a database that shows the correlation between the feature values and the optimal parallelization parameters for a sample program with feature values based on the results of setting and executing multiple parallelization parameters for the sample program. further comprising:
The extraction unit refers to the database generated by the generation unit and extracts an optimal parallelization parameter corresponding to the feature value of the optimization target program calculated by the calculation unit. The parallel processing device according to claim 1. The correlation is based on multiple characteristic values indicating the average number of executed instructions, the relative standard deviation of the number of executed instructions, and the average cache reuse rate for each loop, and processor core allocation and parallelization for each loop. The parallel processing device according to claim 1 or 2, characterized in that the correlation is between a scheduling method and a plurality of parallelization parameters indicating the number of chunks of the parallelization scheduling method.

Images