JP4304379B2 - Shift communication method for distributed memory type multiprocessor system and distributed memory type multiprocessor system - Google Patents

Description

  The present invention relates to a shift communication method for a distributed memory type multiprocessor system and a distributed memory type multiprocessor system capable of executing the shift communication method.

  What is important for improving the performance of a distributed memory multiprocessor system is what communication procedure is used when data is transmitted and received between processors. Since the communication speed between processors is considerably slower than the calculation speed of the processor, it is not possible to improve the system performance if inefficient communication is performed. A technique is known that generates an optimal communication schedule when data is transmitted and received between processors constituting a distributed multiprocessor system (see, for example, Patent Document 1).

  In a general parallel computer, communication between a plurality of processors generated when executing a program is executed based on a communication library for asynchronous communication. For example, a communication library such as MPI or EUI prepares a plurality of relatively simple communication patterns that are frequently used, in addition to normal commands such as send and receive.

  Among these communication patterns, there is a communication pattern called shift communication. Shift communication refers to a communication mode in which each processor transmits data in a certain direction to an adjacent processor.

JP-A-9-330304

  In the distributed memory type multiprocessor system, when the same shift communication is performed, there is a technique for speeding up communication by storing the communication schedule at that time and reusing the stored communication schedule. However, when the dimension or direction in which shift communication is performed changes, the communication schedule cannot be reused, and the stored communication schedule must be discarded and newly regenerated.

  The problem to be solved by the present invention is to provide a technology capable of performing shift communication at high speed even if the target array, target dimension, and direction of shift communication change or straddle different procedures.

  The means for solving the problem will be described below using the numbers used in [Best Mode for Carrying Out the Invention]. These numbers are added to clarify the correspondence between the description of [Claims] and [Best Mode for Carrying Out the Invention]. However, these numbers should not be used to interpret the technical scope of the invention described in [Claims].

In order to solve the above problems, a shift communication method of a distributed memory multiprocessor system is executed by the following method. The method includes: [a] a step of analyzing an array arrangement method for performing shift communication; and [b] a shift communication schedule table (7) for the same division arrangement based on an analysis result obtained by the analysis. [C] If there is a shift communication schedule table (7) for the same divided arrangement as a result of the search of the shift communication schedule table (7), [c] A step of searching whether or not the shift communication schedule (7 _{1 to} 7 _n ) for the direction has already been generated; and [d] shift communication that has not been generated as a result of the search of the shift communication schedule (7 _{1 to} 7 _n ). When there is a schedule (7 _{1 to} 7 _n ), the shift communication schedule corresponding to the necessary divided arrangement method, dimension, and direction Joules ₍₇ 1 to _7-n) generates a new, and generating a new shift communication schedule table (7) containing the new shift communication schedule _{(7 1 ~7 n), [} e] the new The shift communication method includes a step of executing shift communication with reference to the shift communication schedule table (7).

In the shift communication method, the step [c] further includes a step of generating a new shift communication schedule table (7) when the shift communication schedule (7 _{1 to} 7 _n ) does not exist. Is preferred.

In the shift communication method, the [e] step further includes the shift communication schedule table (7) or the new shift communication when all necessary shift communication schedules (7 _{1 to} 7 _n ) have been generated. A shift communication method in which shift communication is executed with reference to the schedule table (7) may be used.

In order to solve the above problems, the following distributed memory type multiprocessor system is constructed. The system includes a mapping analysis unit (6-1) for analyzing a divided arrangement method of an array for performing shift communication, and a shift communication schedule table (7) for the same divided arrangement based on an analysis result obtained by the analysis. It is preferable to include a communication schedule reusing means (6-2) for searching whether or not exists and, if not, generating a shift communication schedule table (7). Here, when there is a shift communication schedule table (7) for the same divided arrangement, the communication schedule reusing means (6-2) is configured to use the shift communication schedule (7 _{1 to} 7 _n for the required direction of the required dimension. ) Is already generated, and if there is a shift communication schedule (7 _{1 to} 7 _n ) that has not been generated, a new communication schedule is requested to be generated. In addition, the distributed memory multiprocessor system described above is shifted in response to the request for generating the new communication schedule, and the shift communication schedule (7 _{1 to} 7 _n ) corresponding to the necessary divided arrangement method, dimension, and direction is set. A shift communication schedule generating means (6-3) for generating and updating the shift communication schedule table (7) to generate a new shift communication schedule table (7); and the new shift communication schedule table (7). Reference is made to include shift communication execution means (6-4) for performing actual shift communication.

  In the distributed memory multiprocessor system, the shift communication schedule generation means (6-3), when there is no shift communication schedule table (7) for the same divided arrangement, creates a new shift communication schedule table (7 ) Is preferably generated.

  Further, in the distributed memory multiprocessor system, the shift communication schedule generation means (6-3) performs shift communication to the shift communication execution means (6-4) using the current shift communication schedule table (7). The shift communication execution means (6-4) actually refers to the shift communication schedule table (7) or the new shift communication schedule table (7) in response to the instruction. The distributed memory type multiprocessor system is configured to execute the shift communication.

Moreover, it is possible to solve the above problems by causing a computer to execute the following program. The program is a program capable of executing shift communication of a distributed memory multiprocessor system,
Computer
Whether there is a mapping analysis means (6-1) for analyzing the divided arrangement method of the array for performing the shift communication and the shift communication schedule table (7) for the same divided arrangement based on the analysis result obtained by the analysis If it does not exist, the communication schedule reuse means (6-2) for generating the shift communication schedule table (7) and the necessary divided arrangement method in response to the request for generation of the new communication schedule Shift communication schedule generating means for generating a shift communication schedule (7 _{1 to} 7 _n ) corresponding to the dimension, direction, and updating the shift communication schedule table (7) to generate a new shift communication schedule table (7) (6-3) and the new shift communication schedule table (7) with reference to the actual shift communication It is a program for functioning as shift communication execution means (6-4) for performing.
Here, when there is a shift communication schedule table (7) for the same divided arrangement, the communication schedule reusing means (6-2) is configured to use the shift communication schedule (7 _{1 to} 7 _n for the required direction of the required dimension. ) Is already generated, and if there is a shift communication schedule (7 _{1 to} 7 _n ) that has not been generated, it is preferable to request generation of a new communication schedule.

  In the program, the shift communication schedule generation means (6-3) further generates a new shift communication schedule table (7) when there is no shift communication schedule table (7) for the same divided arrangement. It may be a program for causing it to function as a means to do this.

  Further, in the program, the shift communication schedule generation means (6-3) performs shift communication using the current shift communication schedule table (7) to the shift communication execution means (6-4). The shift communication execution means (6-4) is actually operated with reference to the shift communication schedule table (7) or the new shift communication schedule table (7) in response to the instruction. It may be a program for functioning as means for executing the shift communication.

  By executing the above solution, the communication schedule is partially independent by generating and storing the communication schedule in each dimension and in each direction independently in the communication schedule table corresponding one-to-one with the array arrangement method. Can be reused.

  According to the present invention, it is possible to increase the speed of shift communication if the target arrangement, the target dimension, and direction of shift communication are changed, or if the split arrangement method is the same even if the procedure is different. In particular, it is possible to speed up shift communication even when the target array, target dimension, and direction of shift communication change each time.

  In addition, according to the present invention, the communication schedule is partially re-created independently by generating and storing the communication schedule in each dimension and each direction independently in the communication schedule table corresponding to the one-to-one layout arrangement method. This makes it possible to reduce the cost of communication schedule generation.

  An outline of an embodiment of the present invention will be described. In the distributed memory multiprocessor system 1 of the present embodiment, array data is divided and arranged on the storage device group 4 of the parallel processing node 2, and a program for using / defining such array data is a computing device. Run on group 3 At that time, the distributed memory type multiprocessor system 1 includes an arithmetic processing node communication unit 5 between a plurality of arithmetic processing nodes (2-n: n is an arbitrary natural number) in which adjacent elements of the array are divided and arranged. When the shift communication via is repeatedly performed, information necessary for the shift communication is stored in the shift communication schedule table 7 independently for each dimension and direction, and reused later. As a result, the distributed memory multiprocessor system 1 of the present invention has a high-speed operation even when the shift communication dimension, the shift communication direction, the shift communication target array changes, or the shift communication occurs in different procedures. It communicates. In the following embodiments, the arithmetic processing node is preferably an information processing device represented by a personal computer or a server device. Further, the computing device (3-1 to 3-n) is preferably a CPU provided in the personal computer. Furthermore, it is preferable that the storage device (4-1 to 4-n) is a RAM provided in the personal computer.

[Configuration of the embodiment]
The configuration of the embodiment of the present invention will be described in detail below with reference to the drawings. FIG. 1 is a block diagram illustrating a configuration of a distributed memory multiprocessor system 1 according to the present embodiment. As shown in FIG. 1, array data is divided and arranged on the storage device group 4 of the parallel processing node 2 on the distributed memory multiprocessor system 1, and such array data is used. When the program to be defined is executed on the computing device group 3, shift communication via the inter-processing node communication unit 5 is performed between the parallel processing nodes 2 in which adjacent elements of the array are divided and arranged. .

  In other words, the distributed memory type multiprocessor system 1 of the present embodiment is configured to include the parallel operation processing node 2. As shown in FIG. 1, the parallel processing node 2 is composed of a plurality of processing nodes (2-1 to 2-n: n is an arbitrary natural number). Each arithmetic processing node (2-1 to 2-n) includes a computing device (3-1 to 3-n) and a storage device (4-1 to 4-n). Further, as shown in FIG. 1, the plurality of arithmetic processing nodes (2-1 to 2-n) are connected to be communicable with each other via the inter-processing node communication unit 5. In the distributed memory multiprocessor system 1 of the present embodiment, a plurality of operation processing nodes (2-1 to 2-n) connected to each other constitute a parallel operation processing node 2, and the parallel operation processing node 2 Performs information processing in cooperation with the computing device group 3 and the storage device group 4.

  A configuration for executing shift communication in the present embodiment will be described below. FIG. 2 is a block diagram illustrating the configuration of the shift communication execution unit 6 in the present embodiment. In the present embodiment, the shift communication executed between the operation processing nodes (2-1 to 2-n) is a communication interface (illustrated) provided in each of the operation processing nodes (2-1 to 2-n). Preferably). The communication schedule between the arithmetic processing nodes (2-1 to 2-n) is determined by the shift communication execution unit 6. This communication schedule determination operation is preferably performed by a communication library called at the time of execution.

  Referring to FIG. 2, the shift communication unit 6 includes a mapping analysis unit 6-1, a communication schedule reuse unit 6-2, a communication schedule generation unit 6-3, and a shift communication execution unit 6-4. It is configured. The mapping analysis unit 6-1 is an information processing function block that analyzes an array arrangement method for performing shift communication. The communication schedule reuse unit 6-2 is an information search function block that searches whether there is a shift communication schedule table 7 for the same divided arrangement. In addition, if the predetermined shift communication schedule table 7 does not exist as a result of the search, the communication schedule reuse unit 6-2 newly generates the shift communication schedule table 7. Further, if there is a shift communication schedule table 7 for the same divided arrangement, it is searched whether a shift communication schedule for a required direction of a required dimension has already been generated. As a result of the search, if the shift communication schedule to be generated is not generated, the communication schedule generating unit 6-3 is instructed to generate. When all necessary shift communication schedules have been generated, the shift communication execution unit 6-4 is instructed to execute shift communication.

  The communication schedule generation unit 6-3 is an information processing function block that generates a shift communication schedule when there is a shift communication schedule that needs to be newly generated. The communication schedule generation unit 6-3 generates a shift communication schedule corresponding to a necessary division arrangement method, dimension, and direction, and stores the shift communication schedule as a corresponding dimension and direction in the shift communication schedule table 7.

  The shift communication execution unit 6-4 is a communication processing unit that performs actual shift communication with reference to the shift communication schedule table 7.

FIG. 3 is a table illustrating the configuration of the shift communication schedule table 7 generated by the shift communication execution unit 6. Referring to FIG. 3, the shift communication schedule table 7 includes a first communication schedule 7 ₁ to an m-th communication schedule 7 _n .

The first communication schedule ₇₁ retains the one-dimensional instantaneous direction of the shift communication schedule in the shift communication. Second communication schedule 7 ₂ holds a one-dimensional superior direction of the shift communication schedule in the shift communication. Third communication schedule 7 ₃ holds the 2-dimensional instantaneous direction of the shift communication schedule in the shift communication. Fourth communication schedule 7 ₄ holds a 2-dimensional superior direction of the shift communication schedule in the shift communication. Thus, each table constituting the shift communication schedule table 7 holds a schedule in which each dimension and direction are associated with each other. The (m-1) th communication schedule 7 _n-1 holds the shift communication schedule in the n-dimensional down direction in the shift communication, and the m- _th communication schedule 7 _n that finally forms the table is the _n-th communication schedule in the shift communication. Holds the upward shift communication schedule (m corresponds to an arbitrary natural number).

[Operation of the embodiment]
Below, the outline | summary of operation | movement of this Embodiment is demonstrated. For example, referring to the example of the program in the HPF language shown in FIG. 4, each element of the four-element processor configuration P (2, 2) corresponds to one arithmetic processing node (for example, the arithmetic processing node 2-1). If so, the arrays A, B, and C are divided and arranged on the storage device group 4 of the parallel processing node 2 as shown in FIG.

  The loop in the subroutine SUB1 activated in the time evolution loop is calculated in parallel on the parallel processing node 2 by the owner compute rule for the array A (the operation is performed on the processing node where the elements of the array A are arranged). In this case, since the subscript of the second dimension of the array B is shifted by one element in the downward direction and two elements in the upward direction with respect to the array A, the parallel processing node in which the adjacent elements of the array B are divided and arranged Between the two elements, it is necessary to communicate one element downward in the second dimension of the array B and two elements upward. Such communication between adjacent elements of array data is called shift communication.

  Similarly, the loop in the subroutine SUB2 that is activated in the time evolution loop is changed by the owner compute rule for the array A on the parallel operation processing node 2 (the operation is performed on the operation processing node in which the elements of the array A are arranged). When performing parallel calculation, the first dimension subscript of array C is shifted one element downward, one element upward, and the second dimension subscript is shifted one element downward relative to array A. Between the parallel processing nodes 2 in which adjacent elements of C are divided and arranged, it is necessary to communicate one element downward in the first dimension of the array C, one element upward, and one element downward in the second dimension.

  In this way, even if the target array, the target dimension and the direction of the required shift communication change each time, or if the array split layout method is the same even if the procedure is different, it is one-to-one with the split layout method. In the corresponding shift communication schedule table 7, a communication schedule is generated independently for each dimension and each direction, and is reused later. This reduces the cost of communication schedule generation and speeds up.

  The operation of this embodiment will be described in detail below. In the following description of the operation, the operation will be described corresponding to the distributed memory multiprocessor system 1 described above. In the distributed memory multiprocessor system 1, the array data is divided and arranged on the storage device group 4 of the parallel processing node 2, and a program for using / defining such array data is executed on the computer device group 3. An example of the execution is performed, and the operation corresponding to the case where shift communication is performed between the parallel processing nodes 2 in which adjacent elements of the array are divided and arranged via the inter-processing node communication unit 5 is performed. explain. Further, the operation described below is executed corresponding to the program in the HPF language exemplified in FIG.

  Referring to the example program of FIG. 4, subroutines SUB1 and SUB2 including shift communication are repeatedly called in the time evolution loop. Here, in the arrays A, B, and C appearing in FIG. 4, it is assumed that each element of the four-element processor configuration P (2, 2) corresponds to one arithmetic processing node, as shown in FIG. Further, it is assumed that they are divided and arranged on the storage device group 4 of the parallel processing node 2.

  When the loop in the subroutine SUB1 in FIG. 4 is calculated in parallel on the calculation means by the owner compute rule for the array A (the operation is performed on the operation processing node in which the elements of the array A are arranged), The second dimension subscript of the array B is shifted by one element in the downward direction and by two elements in the upward direction. For this reason, communication is required between the processing nodes (2-1 to 2-n) in which the adjacent elements of the array B are divided and arranged, one element in the second dimension of the array B and two elements in the upward direction. It becomes.

  Similarly, the loop in the subroutine SUB2 activated in the time evolution loop is calculated in parallel on the calculation means by the owner compute rule for the array A (the operation is performed on the operation processing node where the elements of the array A are arranged). In this case, the first dimension subscript of the array C is shifted by one element downward, one element upward, and the second dimension subscript is shifted downward by one element with respect to the array A. Therefore, communication is required between the parallel processing nodes 2 in which adjacent elements of the array C are divided and arranged, one element downward in the first dimension, one element upward, and one element downward in the second dimension. It becomes.

  When the subroutine SUB1 in FIG. 4 is activated for the first time, the mapping analysis unit 6-1 of the shift communication unit 6 is activated immediately before the loops therein are executed in parallel. At this time, the mapping analysis unit 6-1 analyzes the division arrangement method of the array data B to be shifted.

The communication schedule reuse unit 6-2 recognizes that the shift communication schedule table 7 corresponding to the specified divided arrangement of the array data B has not been generated by the analysis of the mapping analysis unit 6-1. Therefore, the communication schedule reuse unit 6-2 recognizes that the shift communication schedule for the second element in the second dimension required for the array data B has not been generated for the second element in the second dimension. To do. Based on this recognition, the communication schedule reuse unit 6-2 generates the shift communication schedule table 7 corresponding to the array having the same divided arrangement method as the divided arrangement method of the array data B. In other words, communication schedule reuse unit 6-2, a third communication schedule 7 _3, to generate a fourth communication schedule 7 ₄ capable of storing a new shift communication schedule table 7.

The communication schedule generation unit 6-3 generates a shift communication schedule for the second element in the second dimension required for the array data B, and generates a shift communication schedule for the second element in the second dimension. Registered in the third communication schedule 7 ₃ and the fourth communication schedule 7 ₄ of the shift communication schedule table 7 (the above-described new shift communication schedule table 7) for an array having the same divided arrangement method as that of B.

Next, the shift communication execution unit 6-4 in the shift communication schedule table 7 is generated, the third communication schedule 7 _3, based on the fourth communication schedule 7 ₄ performs the shift communication to sequence data B.

  When the subroutine SUB2 in FIG. 4 is activated for the first time, the mapping analysis unit 6-1 of the shift communication unit 6 is activated immediately before the loop is executed in parallel. At this time, the mapping analysis unit 6-1 analyzes the division arrangement method of the array data C to be shifted.

The communication schedule reuse unit 6-2 searches the shift communication schedule table 7 for the same divided arrangement method as the arrangement method of the array data C specified by the mapping analysis unit 6-1. As a result of this search, the communication schedule reuse unit 6-2 uses the same divided arrangement as C for the shift communication schedule table 7 (new shift communication schedule table 7) generated when the subroutine SUB1 of FIG. A shift communication schedule corresponding to the method is detected. Further, the communication schedule reusing unit 6-2 detects that the shift communication schedule in the first dimension downward and the first dimension upward required in the loop of the subroutine SUB2 has not been generated, and detects the second dimension downward. detecting that a third communication schedule 7 ₃ are already generated.

The communication schedule generation unit 6-3 generates a shift communication schedule for one element in the first dimension in the first direction and one element in the first direction in the first dimension, which is necessary in the loop of the subroutine SUB2. 1-way presently direction first communication schedule _71, is registered in the 1-way superior direction second communication schedule _{7 2.}

The shift communication execution unit 6-4 includes a first communication down direction first communication schedule 7 ₁ , a first element up direction second communication schedule 7 ₂ , and a second element down direction third communication schedule 7 _{3 in} the shift communication schedule table 7. , Shift communication for the array data B is executed.

When the subroutine SUB1 of FIG. 4 is started for the second time or later and the loop is executed in parallel, the mapping analysis unit 6-1 of the shift communication unit 6 started immediately before the loop performs the divisional arrangement method of the array data B to be shifted. Is analyzed. The communication schedule reuse unit 6-2 searches the shift communication schedule table 7 for the same divided arrangement method as the arrangement method of the array data B specified by the mapping analysis unit 6-1, and the corresponding shift communication schedule table 7 the detection that the 2-dimensional under the eyes direction one element of the third communication schedule 7 _3, on the second dimension direction 2 element content fourth communication schedule 7 ₄ shift communication schedule required is already registered To do.

Shifting communication execution unit 6-4 refers to the fourth communication schedule 7 ₄ 2D under the eye direction of the third communication schedule 7 _3, second dimension upward shift communication schedule table 7, the required shift communication Do.

Furthermore, when the subroutine SUB2 of FIG. 4 is started for the second time and the loop is executed in parallel, the mapping analysis unit 6-1 of the shift communication unit 6 started immediately before the loop splits the array data C to be shifted. Analyze the placement method. The communication schedule reuse unit 6-2 searches the shift communication schedule table 7 for the same divided arrangement method as the arrangement method of the array data C specified by the mapping analysis unit 6-1, and the corresponding shift communication schedule table 7 The first communication schedule 7 _{1 for} one element in the first dimension down direction, the second communication schedule 7 _{2 for} one element in the first dimension up direction, and the third communication schedule 7 ₃ for one element in the second dimension down direction are already present. Detect that it is registered.

Shifting communication execution unit 6-4 refers to the one-dimensional undereye first communication schedule 7 _{first direction,} first dimension second communication schedule 7 on the direction of the _2, 2-dimensional under eye third communication schedule 7 ₃ directions Perform the necessary shift communication.

  Thereafter, when the subroutine SUB1 and subroutine SUB2 in FIG. 4 are repeatedly activated, the necessary shift communication schedule has already been generated and registered in the shift communication schedule table 7. Therefore, similarly, shift communication can be performed without performing schedule generation.

  Thus, even if the target array, the target dimension and direction of the required shift communication change each time, or if the divided allocation method is the same even if the procedure is different, the communication is included in one shift communication schedule table 7. By generating the schedule independently in each dimension and in each direction, the cost of communication schedule generation can be reduced, and shift communication can be performed at high speed.

FIG. 1 is a block diagram illustrating a configuration of a distributed memory multiprocessor system according to this embodiment. FIG. 2 is a block diagram illustrating the configuration of the shift communication execution unit in the present embodiment. FIG. 3 is a table illustrating the configuration of the shift communication schedule table. FIG. 4 shows a program example in the HPF language. FIG. 5 illustrates the configuration of an array that is divided and arranged in the storage device of the parallel processing node.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 ... Distributed memory type multiprocessor system 2 ... Parallel operation processing node 2-1 to 2-n ... Operation processing node 3 ... Computing device group 3-1 to 3-n ... Computing device 4 ... Storage device group 4-1 to 4 -N ... Storage device 5 ... Inter-processing node communication unit 6 ... Shift communication unit 6-1 ... Mapping analysis unit 6-2 ... Communication schedule reuse unit 6-3 ... Communication schedule generation unit 6-4 ... Shift communication execution unit 7 ... Shift communication schedule table 7 _{1 to} 7 _n ... 1st to m-th communication schedule

Claims (3)

A shift communication method for a distributed memory multiprocessor system, comprising:
(A) a step in which the mapping analysis means analyzes a method for dividing and arranging the arrays for performing shift communication;
(B) a step of searching for whether or not there is a shift communication schedule table for the same divided arrangement, based on an analysis result obtained by the analysis;
(C) The communication schedule reuse means is a result of searching the shift communication schedule table,
If there is no shift communication schedule table for the same divided arrangement, generate a new shift communication schedule table corresponding to the divided arrangement method,
When the shift communication schedule table for the same division configuration are present, with respect to the direction of the required dimensions required, the steps of each dimension, each direction independent of the shift communication schedule is already find whether it is registered,
(D) The shift communication schedule generation means newly generates a shift communication schedule that is independent of each dimension and each direction, corresponding to the necessary divided arrangement method, dimension, and direction that are not registered in the shift communication schedule table, and registering to the shift communication schedule table,
(E) A shift communication method comprising: a step in which shift communication execution means refers to the shift communication schedule registered in the shift communication schedule table and executes shift communication. Mapping analysis means for analyzing a split arrangement method of an array for performing shift communication;
Based on the analysis result obtained by the analysis, it is searched whether there is a shift communication schedule table for the same divided arrangement . (1) If there is a shift communication schedule table for the same divided arrangement, the required dimension Search for whether or not a shift communication schedule independent of each dimension and each direction has already been generated for the required direction, and if there is a shift communication schedule that has not been generated, request the generation of a new communication schedule , (2) If there is no shift communication schedule table for the same divided arrangement, communication schedule reuse means for generating a new shift communication schedule table corresponding to the divided arrangement method ;
When the shift communication schedule corresponding to the shift communication schedule, dimension, and direction is not registered, the shift communication schedule corresponding to the required division arrangement method, dimension, and direction is independent of each dimension and each direction. Shift communication schedule generating means for newly generating and registering in the shift communication schedule table ;
The shift with reference to the registered shifted communication schedule in the communication schedule table, distributed memory multiprocessor system comprising a shifting communication execution unit that performs a shift communication. A program capable of executing shift communication of a distributed memory type multiprocessor system,
Computer
Mapping analysis means for analyzing a split arrangement method of an array for performing shift communication;
Based on the analysis result obtained by the analysis, it is searched whether there is a shift communication schedule table for the same divided arrangement . (1) If there is a shift communication schedule table for the same divided arrangement, the required dimension Search for whether or not a shift communication schedule independent of each dimension and each direction has already been generated for the required direction, and if there is a shift communication schedule that has not been generated, request the generation of a new communication schedule , (2) If there is no shift communication schedule table for the same divided arrangement, communication schedule reuse means for generating a new shift communication schedule table corresponding to the divided arrangement method ;
When the shift communication schedule corresponding to the shift communication schedule, dimension, and direction is not registered, the shift communication schedule corresponding to the required division arrangement method, dimension, and direction is independent of each dimension and each direction. Shift communication schedule generating means for newly generating and registering in the shift communication schedule table ;
The shift by referring to the communication schedule table registered shifted communication schedule, program for functioning as a shift communication execution unit that performs a shift communication.

Images