JPH09167144A - Method and system for program generation - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a means for generating a parallel program for optimizing a communication processing which respective processors execute so that a parallel processing in plural processors can efficiently be executed. SOLUTION: In the method, the parallel program 10 in which plural kinds of operation procedures and plural kinds of communication procedures corresponding to the communication processing among the processors are mentioned and which is for executing the parallel processing is changed. When the communication quantity of the communication processing which is executed in accordance with the communication procedure which is used at present is assumed to be increased, the communication procedure in the parallel program 10 is rearranged and a mentioned content is changed so that more than two communication procedures are synthesized when time from the start to the end of the parallel processing becomes short.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a technique for creating a parallel program used when a multiprocessor system, a distributed memory parallel computer having a local memory for each processor, and the like perform parallel processing.

[0002]

2. Description of the Related Art When performing parallel processing in a multi-processor system or a distributed memory parallel computer having a local memory for each processor, each processor refers to a parallel program and performs so-called processor communication. In this communication processing, since a large amount of data is handled at high speed, it is possible to transfer data that has been collected to some extent in a batch for better communication efficiency. However, when data with an extremely large amount of communication is transferred, parallel processing operation is started As a result, the parallel processing rate becomes worse.

By the way, conventionally, a program conversion system for converting a program for a serial computer into a program for a parallel computer has been proposed. However, iterative calculation such as "Do loop" appearing in the program for a serial computer has been proposed. Data partitioning, which focuses on, is the mainstream.

Here, the data division processing means that the "array" appearing in the program for the sequential computer is assigned according to the number of processors of the parallel computer, and the array assigned to each processor and the other processors are assigned. The allocated array is a process of exchanging data in the array by data communication and performing a parallel processing operation.

In this case, in order to convert into a program for a parallel computer, a communication function is inserted in the program,
Although it is necessary to set the arrangement state of the processors, the allocation pattern of the array, etc., a lot of development man-hours are required for the actual conversion processing including the understanding of the structure of the sequential program and the recognition of the parallelization method.

At present, a "Fortran-D compiler" or the like is used as a conversion means for automatically setting the above-mentioned processor arrangement status, array allocation pattern and the like. Has become possible.

However, actually, when such automatic conversion is executed on a parallel computer, only a specific pattern is parallelized, and the program is parallelized by careful manual work. The conversion performance is inferior to the case. However, when the program is parallelized by hand, in terms of speeding up the parallel operation,
Although a high effect can be expected, a lot of man-hours and labor costs are required. In addition, when performing program conversion processing so that it can be adapted to multiple types of systems, it is necessary to grasp the characteristics of each system and parallelize the program, so it is extremely manual work. It was difficult.

[0008]

By the way, when a parallel processing program is created in order to perform parallel processing using a distributed memory type parallel computer or the like, processing performance such as calculation processing speed and communication speed peculiar to the computer is created. With this in mind, a parallel processing program has not been created, and a program developed for a certain parallel computer may be used on another parallel computer having different performance.

In this case, even if the above-mentioned program developed for a certain parallel computer has been optimized for the parallel computer, this program is different from another parallel computer having greatly different performance. It is not necessarily optimized for the computer. In particular, in each parallel computer, there is such a problem when performance evaluation is performed by the same program.

Further, as an example of a parallel processing optimization method, it is conceivable to collectively transfer communication data for the purpose of reducing the number of times of communication between parallel computers (processors).
In consideration of the parallel computing processing performance, on the contrary, it may be possible to realize high-speed inter-processor communication by performing division transfer of communication data. How can inter-processor communication be performed to improve parallel processing efficiency? There was a problem that it was very difficult to decide whether to improve.

For example, in order to perform high-capacity data communication at high speed, it is preferable to collectively transfer a certain amount of data, but when extremely large data is transferred, parallel processing is started. Therefore, the efficiency of parallel processing deteriorates.

According to the technique disclosed in Japanese Unexamined Patent Publication No. 5-89070, the communication data arranged in discrete locations are collectively transferred to the memory of the processor to improve the efficiency of parallel processing. However, how to set the communication volume,
There is no technology disclosure for speeding up parallel processing by performing a communication processing order determination method, divided transfer, and the like. In addition, the decision of the communication method for executing parallel processing largely depends on the many years of experience and feelings of experts, and it is very difficult to manually create even a parallel program unique to each parallel computer. Met.

Therefore, an object of the present invention is to perform communication between processors to execute a parallel program, which is used when a multiprocessor system, a distributed memory parallel computer having a local memory for each processor, or the like performs parallel processing. It is to provide a means for creating such that the processing efficiency is improved.

[0014]

In order to solve the above problems and achieve the object of the present invention, there are the following means.

That is, a method for changing a parallel program for performing parallel processing, which describes a plurality of types of arithmetic procedures and a plurality of types of communication procedures corresponding to communication processing between processors, is described. In addition, the communication amount for a certain communication procedure is calculated as the calculation time for each calculation procedure, and
This is a program creation method in which the communication amount for each communication procedure is referred to and determined, and the description of the communication procedure is changed so that communication processing is performed at the determined communication amount.

More specifically, a plurality of types of arithmetic procedures,
And a method for changing a parallel program for performing parallel processing, which describes a plurality of types of communication procedures corresponding to communication processing between processors, and is performed according to the communication procedure currently used. Assuming that the communication processing communication amount is increased, if the time from the start to the end of the parallel processing becomes short, the communication procedures in the parallel program are rearranged so that two or more communication procedures are combined. It is a method of creating a program that changes the description content.

When the communication procedures are rearranged, the predetermined constraint conditions are referred to, and the predetermined constraint conditions are rearranged before or after the specific operation procedure. A program creating method characterized by the above is preferable.

Further, according to another aspect of the present invention, a parallel program for performing parallel processing, which describes a plurality of types of operation procedures and a plurality of types of communication procedures corresponding to communication processing between processors. Is a method of changing
Assuming that the amount of communication processing being performed according to the communication procedure currently in use has been reduced, if the time from the start to the end of the parallel processing becomes short, the communication data to be processed by communication is divided. Thus, there is provided a program creating method for changing the description content of the communication procedure so that the communication processing is performed for each divided data.

Furthermore, according to the present invention, the following system aspects are provided.

That is, it is a system for changing a parallel program for performing parallel processing, which describes a plurality of types of communication procedures corresponding to a plurality of types of arithmetic procedures and communication processing between processors, and is currently used. Assuming that the communication amount of the communication processing performed according to the communication procedure being performed is increased, a means for determining that the time from the start to the end of the parallel processing is shortened, and the means,
When it is determined that the time from the start to the end of the parallel processing is shortened, the communication procedures in the parallel program are rearranged, and the description content is modified so that two or more communication procedures are combined. It is a program creation system provided with.

[0021]

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

FIG. 1 shows the configuration of a communication optimization system which is an embodiment of the present invention.

The communication optimization system 12 includes a communication analysis unit 1
3, a parsing unit 13, a scheduling unit 15,
The communication optimization unit 16 and the target program generation unit 17 are included. The communication optimization system 12 uses the given parallel program 10 to execute the target program 11
Generate Further, the communication database 18 in which information and the like for calculating the communication time and the like are stored is provided.

The communication analysis unit 13 has a function of reading the given parallel program 10 and a function of generating the communication database 18.

Next, the syntactic analysis unit 14 measures the calculation time measuring process for measuring the time of the calculation (calculation) executed based on the parallel program 10 and the communication time executed based on the parallel program 10. Communication time measurement processing,
A dependency relationship checking process, which is executed based on the parallel program 10, is performed to check the dependency relationship indicating the order relationship between calculation and communication.

The scheduling unit 15 also performs a measurement result display process, which is a display process based on the result of each measurement process performed by the syntax analysis unit 14.

The communication optimizing unit 16 optimizes communication, for example, exchanges of a plurality of communication procedures and division processing of communication target data. At this time, the communication optimization unit 16 refers to the communication database 18 to optimize the communication.

The object program generator 17 refers to the result of the optimization performed by the communication optimizer 16 to generate the object program 11.

First, a specific example of the parallel program 10 given to this system will be described with reference to FIG. Here, as an example of a programming language, "FORT
RAN ”is used. That is, FIG. 5 shows FORTRA.
The source program of N is shown. For ease of explanation, an array declaration instruction, a processor calculation range designation instruction, various variable setting instructions, etc., which are usually described at the head position of the source program, are omitted.

Further, hereinafter, assuming that four processors are used for parallel processing, each processor is supposed to perform processing for "1/4" of the total number of array elements.

Device numbers from "0" to "3" are assigned to the four processors, and the respective processors are respectively assigned
Described as "PE0", "PE1", "PE2", and "PE3".

Now, the contents of calculation processing and communication processing performed by the program shown in FIG. 5 will be described.

In steps 50, 53, 56 and 59, the array calculation is performed and the arrays A (I), B (I) and C are calculated.
A calculation process for setting the contents of (I), D (I), and E (I) is performed. In addition, each calculation procedure is described in "CA
L1 ”,“ CAL2 ”,“ CAL3 ”, and“ CAL4 ”.

Further, steps 51, 52, 54, 55,
In 57 and 58, communication processing between processors is performed. In addition, each communication processing procedure, "COM1", "COM
2 ”and“ COM3 ”.

Communication is performed as follows, for example, when a calculation process using an array is performed.

That is, when a processor performs a calculation process using an array and does not own the element of the array, the processor transmits the element transmitted from another processor owning the element. , To receive. This is an example of communication performed when a calculation process using an array is performed.

As a communication function, "SEN" is used for transmission.
"D", "RECV" is used for reception, and the arguments of "SEND" are "type of data to be transmitted, transfer length (assuming data of 8 bytes for one character), processor number of transmission partner". The argument of “RECV” is “type of data to be received, transfer length, processor number of receiving partner”.

In 50 of FIG. 5, a calculation process using an array is performed in which a value obtained by multiplying the array A by Y is assigned to the array A and a value obtained by multiplying the array A by Z is assigned to the array B. I am programming like. The assignment process for the arrays A and B is performed by the "DO statement" command from "ist" to "iend" with an increment of "1".

That is, the assignment process for the arrays A and B is repeated until the value of the variable I changes from "ist" to "iend".

Next, in 51 of FIG. 5, when the PE (processor number) is other than 3 ("NE" is not equal,
Meaning), the processor corresponding to the number obtained by adding “1” to its own processor number (MY) is the closing price of the array A (the closing price of the array A in each processor, not the closing price of the array A in the parallel processing system. ) Is programmed so that the process of sending) is performed.

Further, in 52 of FIG. 5, when the PE is other than 0, the processor corresponding to the number obtained by subtracting "1" from its own processor number, the "initial value-1" (ist-
1) It is programmed to receive the 1st element.

Then, in 53 of FIG. 5, "I" of the array C
The value obtained by multiplying the th value by the "I-1" th value of array A,
The programming is performed so that the process of substituting into the array C is performed.

The substitution process is performed by the "DO statement" command from "ist" to "iend" with an increment of "1".

Next, in 54 of FIG. 5, when PE is other than 0, the initial value of the array A is transmitted to the processor corresponding to the number obtained by subtracting "1" from its own processor number, and
In 55 of FIG. 5, when the PE is other than 3, from the processor corresponding to the number obtained by adding “1” to its own processor number,
The program is programmed to receive the "iend + 1" th element of the array A.

Next, at 56 in FIG. 5, programming is performed so that the value obtained by multiplying the "I" th value of the array D by the "I + 1" th value of the array A is substituted into the array D. There is. This substitution process is performed by changing the "ist" to "ien" by the "DO sentence" command.
Up to "d", the increment is "1".

Next, in 57 of FIG. 5, when the PE is other than 3, the final value of the array B is transmitted to the processor corresponding to the number obtained by adding “1” to its own processor number, and FIG.
58, when the PE is other than 0, the "initial value-1" (ist-1) th element of the array B is received from the processor corresponding to the number obtained by subtracting "1" from its own processor number. Programming to.

In 59 of FIG. 5, programming is performed so that the value obtained by multiplying the "I" th value of the array E by the "I-1" th value of the array B is substituted into the array E. This substitution process is performed by the "DO sentence" command from "ist" to "iend".
The increment is "1".

The above is an explanation of an example of the parallel program 10 described in the source program shown in FIG.

Next, the operation of the communication optimizing system 12 shown in FIG. 1 will be described with reference to FIGS.

First, the communication analysis unit 13 receives the parallel program 10 (an example of the source program is shown in FIG. 5) and determines the number of processors used for parallel processing.

Here, the number of processors used is "4.
Therefore, when the inter-processor communication is performed, one of the inter-processor communication patterns as shown in FIG. 2 is adopted and the inter-processor communication is performed.

The inter-processor communication pattern shown in FIG. 2 will be described. In FIG. 2, four types of patterns are shown. Circled numbers near the arrows indicate the communication order.

In the pattern of FIG. 2A, the processor 0 (PE0) has all the other processors (PE1 and PE0).
2, PE3) is a pattern for receiving data.
An example of the reception order is shown by circled numbers.

Next, the pattern of FIG.
This is a pattern for performing a communication operation opposite to the pattern of (A), and the processor 0 (PE0) transmits data to all the other processors (PE1, PE2, PE3).
An example of the transmission order is shown by circled numbers.

Further, the pattern of FIG. 2C is for performing transmission and reception between the processors of the adjacent comrades.
In the illustrated example, first, each processor performs transmission processing to the processor arranged on the right side of itself (),
Next, each processor performs transmission processing to the processor arranged on the left side of itself ().

Further, in FIG. 2D, each processor sequentially performs transmission processing. Specifically, first, processor 0 (PE0) is processor 1 (PE).
1), the processor 1 (PE1) completes the reception, and then the processor 1 (PE1) transmits to the processor 2 (PE2) and the processor 2 (PE2) After the reception is completed, finally, the processor 2 (PE2) performs the transmission process to the processor 3 (PE3).

These are merely examples of communication patterns.

The communication analysis unit 13 determines the communication time for each pattern of inter-processor communication for each data transfer length (0, 8, 16, ..., 8n bytes (n is a natural number)).
In addition, it has a function of measuring in advance and creating a database. This database is the communication database 18 shown in FIG. Therefore, the communication database 18
By referring to, the communication time corresponding to the adopted communication pattern and data transfer length is determined.

Specifically, the communication database 18
The information of the communication pattern name, the data transfer length, and the communication time corresponding to the type of the communication pattern is stored as a pair.

Here, as an example of the values stored in the communication database 18, FIG. 3 shows the relationship between the data transfer length and the communication time in each communication pattern of processor 0. (A), (B), (C), shown on the right side of FIG.
(D) corresponds to the communication pattern described in (A) to (D) of FIG. As shown in FIG. 3, in each communication pattern, the communication time corresponding to the data transfer length is preset in the communication database 18.

As shown in FIG. 3, (A) to (D) of FIG.
According to the communication pattern indicated by, the communication time does not change much when the data transfer length is small to some extent (generally 1024 Kbytes or less).

Next, the processing performed by the syntax analysis unit 14 will be described with reference to FIGS.

FIG. 4 is a flow chart showing the processing contents performed by the syntax analysis unit 14. The processing performed by the syntax analysis unit 14 will be described below.

First, in step 40, the parallel program 10 received by the communication analysis unit is obtained, and the Do loop,
The existence of GoTo sentence, Subroutine, etc. is searched, and the parallel program 10 is divided into a calculation processing unit and a communication processing unit.

Here, the calculation processing unit is a portion where a series of calculation processing is performed in the parallel program. For example, 50, 53 in FIG.
56 and 59.

The communication processing unit is a portion of the parallel program in which communication processing is performed, and examples thereof include 51 and 52 in FIG. 5 showing the communication processing procedure.

Then, the parallel program is obtained by dividing,
A name is given to each of the calculation processing unit and the communication processing unit (steps 42 and 46).

Specifically, the following may be done. Since “50” in FIG. 5 performs the calculation using the array, it is named CAL1 (abbreviation of “Calculation”), and 51 and 52 in FIG. 5 perform inter-processor communication. COM1 ("Commuu
abbreviation of "nication: communication").

Similarly, 53 in FIG. 5 is designated as “CAL2”, and FIG.
54 and 55 of "COM2", 56 of FIG. 5 is "CAL
3 ”, 57 and 58 in FIG. 5 are“ COM3 ”, and 59 in FIG. 5 is“ CAL4 ”.

First, the communication time measuring process 41 will be described. In this processing, the communication time is calculated for each communication processing unit named in step 42. That is, in step 43, the data transfer length and communication pattern for each communication processing unit are checked, and in step 44,
The communication time is calculated for each communication processing unit by referring to the communication database 18.

Next, the calculation time measuring process 45 will be described.

In this processing, the calculation time, which is the time required for the calculation, is calculated for each calculation processing unit named in step 46. That is, in step 47, the calculation amount for each calculation processing unit is calculated, and further, in step 48, the calculation amount calculated by referring to the calculation processing performance (one calculation, a predetermined time) is calculated. The calculation time is calculated based on. The calculation amount is the number of iterations in the case of Do loop, the number of branches in the case of GoTo statement,
In the case of Subroutine, it may be defined as the number of calls, or may be defined as the total number of program steps for each calculation processing unit, and there is no limitation to these defining methods. For each calculation processing unit, any parameter may be adopted as long as it can calculate the time required for the calculation.

Further, the arithmetic processing performance may be determined in advance by determining a routine for checking the performance and executing the routine only once in step 48.

Finally, in the dependency relationship investigation processing 49, the dependency relationship between the communication processing unit and the calculation processing unit is investigated. Here, the dependency means, when focusing on a certain communication processing unit,
Due to the existence of calculation processing units that must be placed before and after the communication processing unit, and because there is a communication processing unit that is a reception partner or a transmission partner of a certain communication processing unit, a specific calculation processing unit calculation It means that there is an arrangement relationship in the program that must be protected with a specific communication processing unit.

In the dependency relation investigation processing 49, it is examined whether or not there is a dependency relation between the communication processing unit and the calculation processing unit, and the information on the existing dependency relation is calculated.
Pass to 5.

Next, the scheduling unit 15 performs a result display process for displaying the measurement and investigation results obtained by the syntax analysis unit 14. This display may be displayed on a display device not shown in FIG. The scheduling unit 15 also has a function of passing the measurement and investigation results obtained by the syntax analysis unit 14 to the communication optimization unit 16.

FIG. 6 shows the display data, processor 0.
Shows the data in the schedule format. The display data here is received by the communication optimizing unit 16 and is processed based on the received data, which will be described later.

Now, with reference to FIGS. 4, 5 and 6, the schedule format data in FIG. 6 will be described.

First, "NAME" 62 in FIG. 6 is a column showing a list of names named in steps 42 and 46 in FIG. 4 in order. In addition, in FIG.
In the column of "IME" 63, the communication time for each communication processing unit calculated in step 44 of FIG. 4 or the calculation time for each calculation processing unit calculated in step 48 of FIG. 4 is shown.

The "calculation amount" 64 of FIG. 6 is a column showing the calculation amount calculated in step 47 of FIG. 4, and the "transfer length" 65 of FIG. 6 is checked in step 43 of FIG. It is a column showing the data transfer length. The "pre-calculation" of the "dependency" 66 of FIG. 6 indicates the name of the calculation processing unit that needs to perform the calculation processing before performing a certain communication processing, while the "post-calculation" This indicates the name of a calculation processing unit that needs to perform calculation processing after performing a certain communication processing. In addition, “reception” indicates from which processor data is received, and “transmission” is
It indicates to which processor the data is transmitted.

Referring to FIG. 6 in more detail,
It looks like this:

It can be seen that "CAL1" has a calculation time (TIME) of "5.00 (μsec)" and a calculation amount of "200". Since "CAL1" is not a communication processing procedure, the fields of "transfer length" and "dependency" are blank.

"COM1" is the communication time (TIM
E) is “412.00 (μsec)”, and “transfer length” is “8
"Byte""and" dependency "indicate that" CAL1 "must be calculated before" COM1 "communication processing and" CAL2 "must be calculated after" COM1 "communication processing. ing. Furthermore, since "transmission" is "1", it means that processor 0 transmits to processor 1. Further, since "receive" is blank, processor 0 does not perform the receiving operation.

Since "CAL2" is the same as CAL1, its description is omitted.

"COM2" is a communication time (TIME) of "1025.00 (μsec)", "transfer length" is "8 (byte)", and "dependency" is "COM2". It indicates that it is necessary to calculate “CAL1” before performing the above, and perform “CAL3” after performing the communication processing for “COM2”. Furthermore, since "reception" is "1", it means that processor 0 receives the data sent from processor 1. Further, since “Send” is blank, processor 0 does not perform the send operation.

Further, "CAL3" and "CAL4" are
It is the same as "CAL1", and "COM3" is COM except that "post-calculation" of "dependency" is "CAL4".
Since it is the same as that of 1, "CAL3", "CAL4",
Description of “COM3” is omitted.

Now, referring to FIG. 6, it can be seen that the communication time (TIME) of "COM2" is significantly increased as compared with "CAL1". This is because processor 0 is used in the operation of receiving the data sent from processor 1.
This is because there is a waiting time.

FIG. 7 shows an example of “calculation time”, “communication time”, and “waiting time” for each processor based on the data of the scheduling format as shown in FIG.

The "calculation time" is the time during which processing is performed for the calculation processing unit, and the "communication time" is
Is the time during which processing is performed for the communication processing unit. The time other than the “calculation time” and the “communication time” is the “waiting time”.

FIG. 7 shows “calculation time”, “communication time”, and “waiting time” for each of the four processors. Referring to the data in FIG. 6, for example, the white part 70 of the processor 0 (PE0) corresponds to the waiting time in the process of “COM2”. In the present invention, the communication processing procedure is optimized so that such "wait time"
It is characterized by suppressing the scattering of.

Now, the contents of the communication optimizing process performed by the communication optimizing unit 16 will be described with reference to FIG. The data in the scheduling format before the optimization is shown in FIG. 6, and the data in the scheduling format after the optimization is shown in FIGS. 9 and 10.

First, the calculation amount or transfer length before and after each calculation processing unit or communication processing unit in FIG. 6 is checked (step 80). Specifically, for “CAL1” in FIG. 6 (which is a unit of calculation processing), it is “0” because there is no processing existing before, and the transfer length of the communication processing procedure existing after is “0”. Check that it becomes "8".

Further, in the case of "COM1" (this is a communication processing unit), the calculation amount of the calculation process existing before is "200", and the calculation amount existing after is also "200". Can be investigated.

Similarly, "CAL2", "CAL"
For each of "3", "CAL4", "COM2", and "COM3", the calculation amount or transfer length before and after that is checked.

Next, from the relationship between the amount of calculation and the transfer length, it is determined whether the increase in the current transfer length enables the faster parallel processing (step 81). Here, since the transfer lengths of “COM1”, “COM2”, and “COM3” are small, increasing the transfer length in all results in faster parallel processing. It should be noted that the communication time with respect to the transfer length is predetermined as shown in FIG. 2, and it is possible to determine whether or not the increase in the transfer length allows high-speed parallel processing.

When it is determined in step 81 that the current transfer length is not increased, the high-speed parallel processing is not performed. In step 84, the communication data is divided and the communication processing is performed. To do so. The data division unit may be determined in advance.

Next, in steps 82 and 83, "COM
It is also determined whether the order of "1" to "COM3" can be changed, that is, whether the arrangement position in FIG. 6 can be moved, and if it is possible, the movable range. “COM1” to “COM3” need only exist after “CAL1” and before “CAL2” if the dependency is referred to. Therefore, each communication processing procedure has a dependency. It is possible to move within the range of satisfaction.

Such processing is easy to understand by comparing FIG. 6 and FIG.

However, FIG. 9 and FIG. 10 show the data in the scheduling format after optimization, and are not in the middle of optimization.

That is, "COM2" and "CO" in FIG.
An example of movement of "M3" is shown in FIG. "COM
Both "2" and "COM3" have been moved so as to satisfy the dependency, and the movement is also possible by "C".
It is performed so that the OMs are continuously arranged. In addition, in FIG. 9, "COM1" and "COM3"
Is shown as a communication processing unit 91, and "CO
“M2” is shown as the communication processing unit 92.

"COM1", which is the communication processing unit 91,
Both “COM3” perform a transmission process to the processor 1. On the other hand, “COM2” performs the receiving process from the processor 1.

In FIG. 10, "COM1" and "CO1"
M3 ”together with the communication processing unit 101“ TRA1 ”
In addition, the communication processing unit 10 is defined as "COM2".
It is 2.

Next, of the moved communication processing units,
Determine what can be communicated at the same time.

This is performed until the target transfer length is reached in step 85 of FIG. The “target transfer length” here means that the data transfer length is as long as possible. The communication processing unit 91 shown in FIG. 9 corresponds to a state in which the predetermined transfer length is reached.

Since the communication processing unit 91 has the same processor 1 as the transmission partner, the two communication processing procedures "COM1" and "COM3" can be simultaneously executed.

However, the communication processing unit 92 of FIG.
Unlike the other communication processing procedure, "COM2" is another communication processing procedure "COM1", "COM" for receiving processing.
3 ”cannot be executed at the same time. Here, the communication processing procedure 91 that can be executed at the same time is the same as the communication processing procedure 101 shown in FIG.
1 ”and“ COM3 ”as one communication processing procedure“ TRA
1 ”, the transfer length is a total of“ 16 ”of the two communication processing procedures, and the communication time of“ TRA1 ”is obtained with reference to FIG.

Then, in step 86, it is judged whether or not the above-mentioned judgment has been carried out for all the programs for each processor. If all the programs have been judged, the first rearrangement is completed and each processor is judged. Processing time is calculated, for example, the maximum value among them is extracted, and the processing time for the rearrangement is set to this maximum value Tn (n is a variable indicating processing repetition).
(Step 87).

For example, the maximum value is extracted from the processing time of each processor, and the processing time after the first rearrangement is set to T1.

Next, the program is returned to the beginning again (step 88), it is judged whether the communication processing procedure position can be changed to another position, and if it can be changed, the second rearrangement is carried out (step 89) Similarly, the processing time after rearrangement is T2.

As described above, FIGS. 9 and 10 show data in the scheduling format after optimization, and are not in the middle of optimization. Therefore, in actuality, in step 89, in the program for each processor, each "COM" is moved (rearranged) so as to satisfy the dependency, and the optimum communication pattern is obtained.

As described above, if the position of the communication processing procedure can be changed, a new possible rearrangement is performed, the processing time for this is set to Tn, and in the processing time T1 to Tn after the rearrangement, The one having the minimum value is set as the optimum communication pattern (step 810).

That is, in the program for each processor, all the patterns in which "COM" can be moved so as to satisfy the dependency relationship are examined, and the pattern having the smallest Tn among them is determined as the optimum communication pattern. And FIG.
In the scheduling format data shown in, "COM"
However, in reality, there are so many "COMs" that it is necessary to examine all patterns in which "COMs" can be moved.

The communication pattern in this case is represented by scheduling data as shown in FIG.

Finally, the target program generator 17 performs a process of changing the scheduling format data created by the communication optimizer 16 to the target program 11.

Here, the scheduling format data before optimization and the scheduling format data after optimization are compared to determine the position and contents of the communication processing procedure "COM" in the program before optimization. Change
Create a target program that is a parallel program after optimization.

Regarding this processing content, as shown in FIGS.
This will be specifically described with reference to FIG.

First, FIG. 10 shows the communication processing procedure "CO" in the scheduling format data shown in FIG.
“M3” and “COM2” have moved, and “COM2” 102 is arranged next to the new communication processing procedure “TRA1” 101. That is, the communication processing procedure 101 is “CO
The "M1" and "COM2" are collectively referred to as "TRA1".

Therefore, the communication processing procedure 5 in FIG.
7, 58, 54, 55 are moved to be placed under the communication processing procedure 52. Then, the communication processing procedures 51, 5
The result of replacing 2, 57, 58 with another communication procedure is
The communication processing procedures 120 and 121 shown in FIG. As can be seen by comparing FIG. 12 and FIG. 5, there is no change in other parts.

By the way, here, a local array "BUF" is used, and the data to be transmitted or received in the communication process is temporarily stored. That is, FIG.
Specifically, in the communication processing procedure 120, the closing price of the array A and the closing price of the array B are respectively set in the array BUF.
Of the array BUF, and the value of the array BUF is transmitted at a time to the processor corresponding to the number (MY + 1) obtained by adding “1” to the processor number (MY).

On the other hand, in the communication processing procedure 121, contrary to the communication processing procedure 120, the value of the array BUF is a processor corresponding to the processor number (MY) minus "1" (MY-1). From the array A,
"Initial value-1" th element and "Initial value-1" of array B
As the second element, the first and second array elements of the array BUF are substituted, respectively.

With such a program description, the communication process performed at "51, 52, 57, 58" in FIG.
This is performed at "120, 121" in FIG.

FIG. 11 shows scheduling patterns of “calculation time”, “communication time”, and “waiting time” for each processor with reference to the data in the scheduling format after optimization shown in FIG. It is a thing.

As can be seen from the comparison with FIG. 7, the waiting time and the communication time are shorter than those before the optimization processing. this is,
This is because the time required for the communication processing is reduced because the communication processing procedure is put together and the communication data can be collectively sent.

If the relationship between the transfer length and the amount of calculation is examined and the transfer length is long and "it is judged that the parallel processing does not become faster when the transfer length is increased", the transfer is reversed. By shortening the length, that is, by dividing the communication data,
The time required for communication processing is reduced.

As described above, a plurality of types of arithmetic procedures and a plurality of types of communication procedures corresponding to communication processing between processors are described, and are described in a parallel program for performing parallel processing. The communication amount for the communication procedure is determined with reference to the calculation time for each calculation procedure and the communication amount for each communication procedure, and further, the communication procedure of the communication procedure is determined so that the communication process is performed with the determined communication amount. A method for creating a program for changing the description is provided.

It should be noted that the system shown in FIG. 1 is provided in a parallel processing computer or the like, and each processor performs a communication process or a calculation procedure by referring to a target program generated by giving a parallel program. By doing so, a parallel processing system for high-speed interprocessor communication can be realized.

[0127]

As described above, according to the present invention, in consideration of the communication time required for inter-processor communication and the calculation time, the communication processing pattern is changed to be optimum and the communication processing is performed. Therefore, it is possible to create a parallel program that can reduce the communication waiting time, improve the communication data transfer efficiency, and improve the parallel processing efficiency.

[Brief description of the drawings]

FIG. 1 is a configuration diagram of a communication optimization system.

FIG. 2 is an explanatory diagram of a communication pattern.

FIG. 3 is an explanatory diagram showing the relationship between the data transfer length and communication time in processor 0.

FIG. 4 is an explanatory diagram showing a process performed by a syntax analysis unit.

FIG. 5 is an explanatory diagram describing a parallel program at a source level.

FIG. 6 is an explanatory diagram of data handled by a scheduling unit.

FIG. 7 is an explanatory diagram of a pattern example of calculation time, communication time, and waiting time of each processor.

FIG. 8 is an explanatory diagram illustrating a process performed by a communication optimization unit.

FIG. 9 is an explanatory diagram of data handled in communication optimization processing.

FIG. 10 is an explanatory diagram of data handled in communication optimization processing.

FIG. 11 is an explanatory diagram of an example of a pattern of calculation time, communication time, and waiting time of each processor after the communication optimization process.

FIG. 12 is an explanatory diagram in which a parallel program (objective program) after the communication optimization process is described at the source level.

[Explanation of symbols]

10 ... Parallel program, 11 ... Objective program, 12 ...
Communication optimization system, 13 ... Communication analysis unit, 14 ... Syntax analysis unit, 15 ... Scheduling unit, 16 ... Communication optimization unit, 17 ... Object program generation unit, 18 ... Communication database

Claims (5)

[Claims] 1. A method for changing a parallel program for performing parallel processing, which describes a plurality of types of arithmetic procedures and a plurality of types of communication procedures corresponding to communication processing between processors, the method being described. Also, the communication amount for a certain communication procedure is determined by referring to the calculation time for each calculation procedure and the communication amount for each communication procedure, and the communication procedure is performed so that the communication process is performed at the determined communication amount. How to change the description of the program. 2. A method for changing a parallel program for performing parallel processing, which describes a plurality of types of arithmetic procedures and a plurality of types of communication procedures corresponding to communication processing between processors, and is currently used. Assuming that the communication amount of the communication processing being performed according to the communication procedure being performed is increased, if the time from the start to the end of the parallel processing becomes short, the communication procedures in the parallel program are rearranged. A method of creating a program in which the description contents are changed so that two or more communication procedures are combined. 3. The rearrangement of communication procedures according to claim 2, wherein the predetermined constraint conditions are referred to, and the predetermined constraint conditions are before or after a specific operation procedure. A method of creating a program, characterized in that it is rearranged later. 4. A method for changing a parallel program for performing parallel processing, which describes a plurality of types of communication procedures and a plurality of types of communication procedures corresponding to communication processing between processors, and is currently used. Assuming that the amount of communication processing performed according to the communication procedure is reduced, if the time from the start to the end of the parallel processing becomes short, the communication data to be processed by communication is divided, A method of creating a program, which changes the description content of the communication procedure so that communication processing is performed for each divided data. 5. A system for changing a parallel program for performing parallel processing, which describes a plurality of kinds of arithmetic procedures and a plurality of kinds of communication procedures corresponding to communication processing between processors, and is currently used. Assuming that the communication amount of the communication processing being performed according to the communication procedure being performed is increased, means for determining that the time from the start to the end of the parallel processing is shortened, and the means for performing parallel processing by the means. When it is determined that the time from the start to the end of the communication is shortened, the communication procedure in the parallel program is rearranged, and the description content is modified so as to combine two or more communication procedures. Creation system.