JP2012252591A - Process allocation system, process allocation method, and process allocation program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an automatic optimal arrangement of a process which shortens the execution time of a parallel program using an MPI library in a large-scale PC cluster system of a three-dimensional torus structure.SOLUTION: In a large-scale PC cluster system in which network switches connected with a plurality of nodes are mutually connected in a three-dimensional torus structure, a job management node 11 which performs process allocation for execution of a parallel program for an MPI program holds position coordinates of nodes in the three-dimensional torus structure. On the basis of the position coordinates of the nodes, the number of network switches in paths in the three-dimensional torus structure between calculation nodes 12 connected with the network switches and I/O nodes 13, which are access targets during file input/output by the calculation nodes, is calculated; a set of I/O node 13 and calculation node 12 with the minimum calculated value is specified; and the MPI rank of the calculation node 12 of the specified set is determined.

Description

  The present invention relates to a process allocation system, a process allocation method, and a process allocation program. Specifically, the execution time of a parallel program using an MPI library is reduced on a large-scale PC cluster system having a three-dimensional torus structure. The present invention relates to a technology that enables automatic optimal arrangement of processes.

  In recent large-scale PC cluster systems, computation nodes are connected by a network topology having a three-dimensional torus structure, and a parallel program is executed by using a large number of computation nodes simultaneously. The computing resource for executing the parallel program is determined by the job management node. The job management node secures a calculation node having a calculation capacity and a memory capacity necessary for executing the parallel program with the number of nodes specified by the user. The parallel program performs calculations using the CPU resources of the calculation nodes thus secured, and performs the data input / output and file input / output processes for the other nodes to advance the overall calculation.

  In general, most parallel programs in the scientific and technical computing field are implemented using an MPI (Message Passing Interface) library. When executing a parallel program, messages (data) are transmitted and received between a plurality of processes in accordance with the MPI protocol, and the calculation as a whole proceeds.

  With regard to these technologies, when a large-scale MPI program is submitted to a PC cluster system in which calculation nodes are connected by a network topology with a three-dimensional torus structure, resources are secured so that the distance between the calculation nodes is minimized. A scheduling technique to be performed (see Patent Document 1) has been proposed.

JP 2006-146864 A

  As described above, the parallel program executed in the large-scale PC cluster system advances the entire calculation while transmitting and receiving messages (data) between a plurality of processes. A plurality of calculation nodes that execute the process are determined by the job management program according to the above-described conventional technique. On the other hand, the time required to execute the parallel program can be obtained by the sum of the calculation time and the communication time. Among these, the calculation time is optimized so that the calculation time is substantially the same between the processes. Here, the communication time varies depending on the distance between the computation nodes that execute the message transmission / reception process.

  For this reason, when aiming to reduce the total time required to execute the parallel program, a technique for optimally allocating processes in addition to assigning calculation nodes is required in order to reduce the communication time as much as possible. However, in the prior art, shortening of the communication time is not considered in executing the parallel program, and the user has manually collected the communication log and tried to arrange the calculation nodes to which the process is assigned.

  SUMMARY OF THE INVENTION An object of the present invention is to provide a technique that enables automatic optimal arrangement of processes for reducing the execution time of a parallel program using an MPI library on a large-scale PC cluster system having a three-dimensional torus structure. .

  The process allocation system of the present invention that solves the above problems is for executing a parallel program for an MPI program in a large-scale PC cluster system in which network switches connected to a plurality of nodes are connected to each other in a three-dimensional torus structure. An information processing system that performs the process assignment of: a storage unit that holds the position coordinates of each node in the three-dimensional torus structure, and based on the position coordinates of each node in the three-dimensional torus structure in the storage unit, A three-dimensional torus connected to each network switch between each computation node that receives a process assignment and performs computation upon execution of the parallel program, and each I / O node that is an access target at the time of file input / output by the computation node Calculate the number of network switches in the path of the structure, I / O node calculation value is minimized and to identify the set of computing nodes, characterized in that it comprises an arithmetic unit for executing processing for determining the MPI rank of the specified set of compute nodes.

  In addition, the process allocation method of the present invention provides a process allocation for executing a parallel program for an MPI program in a large-scale PC cluster system in which network switches connected to a plurality of nodes are interconnected in a three-dimensional torus structure. In the storage unit, a computer that holds the position coordinates of each node in the three-dimensional torus structure is connected to each network switch based on the position coordinates of each node in the three-dimensional torus structure in the storage unit. A network switch on a path in a three-dimensional torus structure between each computation node that receives a process allocation and performs computation in parallel program execution and each I / O node that is an access target when the file is input / output by the computation node The number is calculated and the calculated value is minimized. Identifying a set of I / O node and the compute nodes, and executes the process of determining the MPI rank of the specified set of compute nodes.

  In addition, the process allocation program of the present invention provides a process allocation for executing a parallel program targeting an MPI program in a large-scale PC cluster system in which network switches connected to a plurality of nodes are connected to each other in a three-dimensional torus structure. In the storage unit, the computer holding the position coordinates of each node in the three-dimensional torus structure is connected to each network switch based on the position coordinates of each node in the three-dimensional torus structure in the storage unit. A network switch on a path in a three-dimensional torus structure between each computation node that receives a process allocation and performs computation in parallel program execution and each I / O node that is an access target when the file is input / output by the computation node The number is calculated and the calculated value is Become identifies a set of I / O node and the compute node to execute the process of determining the MPI rank of the specified set of compute nodes, it is characterized.

  According to the present invention, it is possible to automatically and optimally arrange processes for reducing the execution time of a parallel program using an MPI library on a large-scale PC cluster system having a three-dimensional torus structure.

It is a figure which shows the structural example of the large-scale PC cluster system containing the process allocation system (job management node) of this embodiment. It is a figure which shows the structural example of the calculation node of this embodiment. It is a figure which shows the structural example of the job management node of this embodiment. It is a figure which shows the one-dimensional torus structure in this embodiment. It is a figure which shows the structural example of the network switch and calculation node in this embodiment. It is a figure which shows the two-dimensional torus structure of this embodiment. It is a figure which shows a part of three-dimensional torus structure of this embodiment. It is a figure which shows the large-scale PC cluster system by the three-dimensional torus of this embodiment. It is a figure which shows the parallel processing by the collective communication function of the MPI program of this embodiment. It is a figure which shows the space which performs a parallel program on the large-scale PC cluster system by the three-dimensional torus of this embodiment. It is a figure which shows the magnitude | size of the large-scale PC cluster system by the three-dimensional torus of this embodiment. It is a figure which shows the calculation object which comprises the large-scale PC cluster system by the three-dimensional torus of this embodiment. It is a figure which shows the I / O object which comprises the large-scale PC cluster system by the three-dimensional torus of this embodiment. It is a figure which shows the naming convention of the network switch and node of this embodiment. It is a figure which shows the distance between objects in the torus of this embodiment. It is a figure which shows the distance between nodes of the large-scale PC cluster system by the three-dimensional torus of this embodiment. It is a flowchart which shows process sequence example 1 of the process allocation method of this embodiment. It is a flowchart which shows process sequence example 2 of the process allocation method of this embodiment. It is a figure which shows the model which calculates the minimum distance of the I / O node of this embodiment, and a calculation node. It is a flowchart which shows process sequence example 3 of the process allocation method of this embodiment. It is a figure which shows the work area | region used by the minimum distance calculation in MPI collective communication function execution of this embodiment. It is a flowchart which shows process sequence example 4 of the process allocation method of this embodiment. It is a figure which shows the function specification (API) of the recursive call function which performs the minimum distance calculation in MPI collective communication function execution of this embodiment. It is a flowchart which shows process sequence example 5 of the process allocation method of this embodiment. It is a flowchart which shows process sequence example 6 of the process allocation method of this embodiment.

  Embodiments of the present invention will be described below in detail with reference to the drawings. FIG. 1 is a network configuration diagram including a job management node 11 which is a process allocation system according to this embodiment. A job management node 11 shown in FIG. 1 is a computer that enables automatic optimal arrangement of processes to shorten the execution time of a parallel program using an MPI library on a large-scale PC cluster system having a three-dimensional torus structure.

  A large-scale PC cluster system 1 to which a large number of computing nodes 12 and the like are connected has a configuration in which a login node 10, the job management node 11, the computing node 12, and an I / O node 13 are connected to a network 14. Yes. Each node is one computer and communicates with other nodes through the network 14. A user of the PC cluster system 1 connects to the login node 10 using terminal emulator software or the like, and performs program editing, compilation, data editing, and the like. Further, in such a PC cluster system 1, when executing a parallel program that simultaneously uses a plurality of calculation nodes 12 corresponding to the MPI program, the user gives an instruction to submit a job through an operation on the login node 10.

  FIG. 2 is a diagram illustrating a configuration example of the calculation node according to the present embodiment, and FIG. 3 is a diagram illustrating a configuration example of the job management node according to the present embodiment. When the above job is submitted, the login node 10 notifies the job management program 304 operating on the job management node 11. A scheduling function 306, which is one function of the job management program 304, determines a plurality of calculation node candidates that execute the parallel program, and stores the information in the job queue 307.

  Further, the job control function 305, which is another function of the job management program 304, always manages the state of the calculation node 12 of the job in the execution state, and a plurality of calculation node candidates determined by the scheduling function 306 can be used. At the point of confirmation, a parallel program process is generated for the computation node candidate, and execution is started all at once.

  The parallel program process executed on the computation node 12 is performed through the network 14 while performing file access for the I / O node 13 and message transmission / reception with the parallel program process operating on another computation node 12. Proceed with overall execution.

  When the job control function 305 of the job management program 304 detects the end of all processes in the parallel program being executed, all the computation nodes 12 that have executed the parallel program have changed from the program executing state to the program not executing state. It becomes. The calculation node 12 in a state where the program is not executed can be used as a calculation resource for executing a new program by the job control function 305. Information on the job for which the program execution has been completed is notified to the user (user terminal, etc.) who has performed the operation at the login node 10 by the job control function 305 of the job management program 304 via e-mail or the like. The Thus, the user who has received the notification by e-mail or the like can check the output, execution status, error information, etc. of the program that submitted the job by operating the login node 10.

  Here, the hardware configuration of the job management node 11 and the job management node 11 will be described again with reference to FIGS.

  The calculation node 12 is a CPU unit that executes various programs such as a storage unit 201 configured by a memory, a hard disk drive, and the like, and a parallel program 204 stored in the storage unit 201, and performs overall control of the computer itself and various calculations. 202, and a network interface unit 203 that exchanges data with other apparatuses via the network 14. The storage unit 201, the CPU unit 202, and the network interface unit 203 are connected by a BUS.

  In such a computation node 12, when the parallel program 204 is executed, the corresponding program is loaded into the storage unit 201, and a predetermined operating system (not shown) performs execution control thereof. The parallel program 204 includes a calculation function 205, a file input / output function 206, a data input / output function 207, and the like, and uses the CPU unit 202 and the storage unit 201 of the calculation node 12. The file input / output function 206 and the data input / output function 207 of the parallel program 204 require communication with the outside of the computation node, and perform communication using the network interface unit 203.

  The job management node 11 executes a storage unit 301 configured by a memory, a hard disk drive, and the like, a job management program 304 held in the storage unit 301, and the like, and performs overall control of the computer itself, various determinations, calculation, and control. A CPU section 302 that performs processing and a network interface section 303 that connects to other devices via the network 14 to exchange data.

  In such a job management node 11, a job management program 304 is loaded in the storage unit 301, and the job management program 304 is always operating using the CPU unit 302 and the storage unit 301. Information on job submission from the login node 10 is acquired by the job management program 304 through the network interface unit 303. The job management program 304 uses the job scheduling function 306 to determine calculation node candidates for executing the parallel program. The job control function 305 of the job management program 304 performs process generation of a parallel program to be executed by the calculation node 12 and acquisition of the status of the program executed by the calculation node 12 through the network interface unit 303. The job scheduling function 306 refers to the latest execution information of all the calculation nodes 12 acquired from the job control function 305, and determines calculation node candidates for executing the parallel program. The scheduling function 306 registers information on job scheduling in the job queue 307. The scheduling function 306 generates a parallel program process for the calculation node 12 and starts execution when the job control function 305 detects that the calculation node candidate to be executed is available. Let

  FIG. 4 is a diagram showing a one-dimensional torus structure. The torus structure is a structure having a characteristic of returning to the original reference point when traveling in one direction from the reference point. In the structure shown in FIG. 4, the object 401 is used as a reference point, and the process returns to the object 401 when proceeding clockwise. Similarly, the structure of FIG. 4 is a torus structure since the original object 401 is also returned in the case of proceeding counterclockwise. When the torus structure shown in FIG. 4 is a network structure that constitutes a computer system, the object 401 includes a network switch therein, and the network switches are connected to each other by a network cable. In a network having such a structure, when a failure of one network switch or a disconnection of a network cable occurs, communication of the object 401 becomes possible by performing communication using the network switch and the network cable in the reverse direction. The structure has redundancy.

  FIG. 5 is a diagram showing an internal structure of an object (network component) when a computer system is constructed with a network topology based on a torus structure. An object 401 includes a network switch 501 and a plurality of nodes 502. The network switch 501 connects network cables to all six directions of the x-axis direction, the y-axis direction, and the z-axis direction with the network switch 501 as the center, and the network switch of the adjacent object (network component) Communication with 501 is performed.

  The network switch 501 includes a network cable 503 that connects the x-axis directions, a network cable 504 that connects the y-axis directions, a network cable 505 that connects the z-axis directions, and a network that connects to the node 502 included in the object 401. Cable 506 is used for connection.

  The network switch 501 performs switching for the packets acquired from these paths, and sends the packets in the node direction inside the object or in the direction of the external object (network component). When communication is performed between nodes 502 connected to the same network switch 501, data transfer between the nodes 502 can be completed by switching only one network switch 501. When communicating with a node 502 connected to a different network switch 501, data is sent from the transmission source node 502 to the network switch 501 to which the node 502 belongs, and from there to a network switch 501 included in another object 401. Data is sent. This communication is repeated for the adjacent network switch 501 until the object 401 including the receiving node 502 is reached. Finally, the network switch 501 including the receiving node 502 changes to the receiving node 502. Sending data completes the data transfer between the two nodes.

  FIG. 6 shows a two-dimensional torus structure. The two-dimensional torus structure is a structure that simultaneously satisfies the characteristics of returning to the original reference point when traveling in one direction from the reference point with respect to two axes. In the structure of FIG. 6, the object 401 is used as a reference point, and the original object 401 is returned to both the right direction and the left direction. Similarly, the object 401 is used as a reference point, and the original object 401 is returned to both the case where the object 401 moves upward and the case where the object 401 moves downward. Since these features are combined, the structure of FIG. 6 satisfies the requirements as a two-dimensional torus structure. The object 401 includes a network switch 501 and a node 502, and a network having a two-dimensional torus structure can be configured by connecting network cables in the x-axis direction and the y-axis direction of the network switch 501.

  FIG. 7 shows a three-dimensional torus structure. The three-dimensional torus structure is a structure that simultaneously satisfies the characteristics of returning to the original reference point when traveling in one direction from the reference point with respect to the three axes. In the structure of FIG. 7, the object 401 is used as a reference point, and the original object 401 is returned to each other when traveling in the x-axis direction, traveling in the y-axis direction, and traveling in the z-axis direction. Since all of these characteristics are satisfied, the structure of FIG. 7 satisfies the requirements as a three-dimensional torus structure. The object 401 includes a network switch 501 and a node 502. By connecting network cables in the x-axis direction, y-axis direction, and z-axis direction of the network switch 501, a network having a three-dimensional torus structure can be configured. it can.

  FIG. 8 is a diagram illustrating a configuration example of a large-scale PC cluster system having a three-dimensional torus structure. The large-scale PC cluster system 1 connects objects (network components) with a network topology having a three-dimensional torus structure. The object 401 is classified into two types, and an object 803 constituted by one network switch 501 and a plurality of calculation nodes 12, one network switch 501 and a plurality of calculation nodes 12, and one or a plurality of I The object 804 includes the / O node 13, the login node 10, and the job management node 11. When a parallel program is executed as a job in the large-scale PC cluster system 1, the user connects to the login node 10 included in any of the objects 804 using terminal emulator software or the like, and submits the job by a predetermined operation.

  The calculation node candidates for executing the parallel program are determined by the scheduling function 306 of the job management program 304 that always operates in the job management node 11 included in any of the objects 804, so that the calculation node candidates are as dense as possible. The scheduling is performed so that the computation node candidates that execute the parallel program are aggregated.

  As a result of performing such scheduling, each computation node 12 included in the plurality of objects 401 in the job execution space 802, which is a part of the three-dimensional space of the three-dimensional torus structure, executes the parallel program.

  When a program executed on the computation node 12 performs file input / output, if the I / O node 13 is connected to the same network switch 501 and the file in it is an input / output target, communication between the network switches File input / output is possible without the need for Otherwise, file input / output is performed for an I / O node connected to another object 804, and communication is repeated several times between network switches to perform file input / output.

  FIG. 9 is a diagram illustrating an example of a data flow when a collective communication function such as the MPI_BCAST function is executed in the MPI program. The MPI_BCAST function provides a function of giving data owned by only one process to all other processes. The MPI_BCAST function is a synchronous function, and when the MPI_BCAST function is executed in all processes, after synchronization is performed internally, communication for data transfer is performed in parallel, and all the data owned by one process is all remaining. Can be given to the process.

  When the MPI program is executed in N processes and the MPI_BCAST function is called from all processes, N-1 data transfer is performed from one process that owns data to the remaining (N-1) processes that do not own data. Do not do the method. By the method shown in FIG. 9, communication is performed at maximum log 2 N times (rounded up after the decimal point) per process of the MPI program to transfer data to all processes.

  FIG. 9 shows the data flow when the MPI_BCAST function is executed when the MPI program is executed in parallel in 8 processes. In this case, since N = 8, data transfer to all processes is completed at the maximum number of log 2 N times per process, that is, a maximum of 3 transfer times. In FIG. 9, since the parallel number is “8”, MPI ranks 0 to 7 are given to each of the 8 processes.

  Data is initially held by the process of MPI rank “0”, and this is given to 7 processes from MPI ranks 1 to 7 by the MPI_BCAST function. In the first transfer, data is transmitted and received between the MPI rank 0 process and the MPI rank 4 process. Next, in the second transfer process, the MPI rank 0 process transmits and receives data to and from the MPI rank 2 process. The MPI rank 4 process has already received data. At this time, the data received from the MPI rank 0 is immediately transmitted and received between the MPI rank 4 process and the MPI rank 6 process. When the transfer process up to the second time is completed, there are four MPI ranks 0, 2, 4, and 6 that own the data. Next, at the time of performing the third transfer process, the MPI rank 0 process performs transmission / reception with the MPI rank 1 process. Similarly, a process that has already received data performs transmission / reception with a process that has not yet been received, an MPI rank 2 process is an MPI rank 3 process, an MPI rank 4 process is an MPI rank 5 process, and an MPI rank process. The rank 6 process transmits to and receives from the MPI rank 7 process. As a result, when all the transmission / reception so far is completed, all processes have the data. Data transfer to all processes is completed in a short time by repeating data transmission / reception appropriately for processes that have not received data while other MPI rank processes perform data transmission / reception three times. can do. In this method, transfer to all processes is completed in log2 N times. For example, when the MPI_BCAST function is executed in the MPI program of 65536 processes, the transfer time of 16 times (log2 65536 = 16) is all required. 65536 processes can have the same data.

  FIG. 10 is a diagram showing the distance between the calculation node 12 and the I / O node 13 in the PC cluster system 1 having a three-dimensional torus structure. The PC cluster system 1 having a three-dimensional torus structure is composed of an object including only a calculation node 12 (hereinafter referred to as a calculation object) and an object including an I / O node 13 (hereinafter referred to as an I / O object). A job execution space 802 is configured by I / O objects.

  The job execution space 802 includes a calculation object 1001 and a calculation object 1002. Here, the distance between the objects is not the distance between the two objects in the physical space but is defined as a difference in the three-dimensional space coordinates of the network switch 501 connecting the two objects 401. After making such a definition, the distance between the I / O object 1003 and the calculation object 1001 and the distance between the I / O object 1003 and the calculation object 1002 are compared. Assume that the coordinates x and y of the calculation object 1001 and the calculation object 1002 are the same, and only z is different. Since the entire PC cluster system 1 has a three-dimensional torus structure, in FIG. 10, the bottom xy plane of the PC cluster system 1 where the I / O object 1003 exists is adjacent to the top xy plane in the z direction. There is a relationship. Therefore, a calculation object closer to the I / O object 1003 in the job execution space 802 is a calculation object 1001 adjacent in the z direction. The job execution space 802 also includes an I / O object 1004. The I / O object 1004 also includes a calculation node. The calculation object 1001, the calculation object 1002, and the I / O object 1004 included in the job execution space 802 are the I / O objects including the calculation node closest to the I / O node. The object 1004 sets the distance to zero.

  FIG. 11 is a diagram showing the size of the PC cluster system 1 having a three-dimensional torus structure. Since the torus structure has the property of returning to the same reference point when traveling in one direction from the reference point, it is possible to travel infinitely in the same direction. Here, it is assumed that the number of objects 401 is X in the x direction until reaching the reference point again. Similarly, the number of the objects 401 that proceed in the y direction and the z direction and reach the reference point is My and Mz. An origin 401 of a three-dimensional space serving as a common reference in the x, y, and z directions is provided, and an object 401 at a coordinate position of (0, 0, 0) is defined. Since there are Mx objects in the x direction, My objects in the y direction, and Mz objects in the z direction, all the objects 401 constituting the PC cluster system 1 having a three-dimensional torus structure can be represented by three-dimensional coordinates. it can. All objects 401 exist at any coordinate from (0,0,0) to (Mx-1, My-1, Mz-1), and the total number of objects is Mx × My × Mz.

  FIG. 12 is a diagram showing the relationship between the three-dimensional coordinate position of the calculation object and the internal structure of the object. The calculation object includes a network switch 501 and a plurality of calculation nodes 12. The coordinate position of the calculation object is the same as the coordinate position of the network switch 501, and is a position obtained by adding or subtracting 1 to the x, y, z direction of the adjacent network switch 501. Because of the three-dimensional torus structure, the coordinate positions that the x coordinate can take are from 0 to Mx-1. Similarly, the y coordinate is from 0 to My-1, and the z coordinate is also from 0 to Mz-1.

  When the x coordinate of a network switch 501 is 0 or Mx-1, the coordinate of the adjacent network switch 501 is a coordinate position obtained by subtracting 1 when -1 or Mx is not a coordinate value and the x coordinate is 0. When Mx-1 and the x coordinate is Mx-1, the coordinate position obtained by adding 1 is 0. The same applies when the y-coordinate value is 0 or My-1 or the z-coordinate value is 0 or Mz-1.

  A node number is also set for the calculation node 12 inside the calculation object. In FIG. 12, the calculation object has 6 calculation nodes 12 and node numbers 01 to 06 are set. This node number is set to be a unique value among objects. When a node is specified from the PC cluster system 1 having a three-dimensional torus structure, it can be obtained by a combination of the three-dimensional coordinates of the object 401 and the node number.

  FIG. 13 is a diagram showing the relationship between the three-dimensional coordinate position of the I / O object and the internal structure of the object. An I / O object is almost the same as a calculation object, but includes an I / O node 13 in addition to a plurality of calculation nodes 12. The coordinate position of the I / O object is the same as the calculation object, and the node number is the same. Regarding the calculation node 12 and the I / O node 13 inside the I / O object, the example of FIG. 13 has six calculation nodes 12 and one I / O node 13. The node numbers of the calculation nodes 12 are 01 to 06 and the node numbers of the I / O node 13 are 07 so that these node numbers have unique values in the I / O object.

  FIG. 14 is a table showing the correspondence between the names of the network switch 501 and the node 502 constituting the object, the three-dimensional space coordinates, and the node numbers. An object 401 constituting the PC cluster system 1 having a three-dimensional torus structure has three-dimensional coordinates, and the calculation node 12 and the I / O node 13 have node numbers that are unique within the same object.

  An object has one network switch 501. By giving the three-dimensional coordinates of the object to the network switch name, it is possible to have a unique network switch name in the three-dimensional torus space. In this example, the name of the network switch is “S + coordinate position”, the first character S is a switch, and other characters represent the coordinate position of the network switch 501. In the information indicating the coordinate position, each coordinate value of x, y, z is represented by one character, and the total is three characters. When the coordinate value is less than 10, it can be expressed by one character from 0 to 9 as a decimal value. When the coordinate value is less than 16, it is expressed as a hexadecimal value by one character from 0 to F. When the coordinate value is 16 or more, the problem is solved by using an alphabetic character after F as an n-ary number (16 = G, 17 = H, 18 = I, and so on).

  The node name uses a node number in addition to the coordinate position of the object 401. The node name is “N + coordinate position + _ + node number”. The first character N indicates a node, and other characters represent object coordinates and node numbers. The node name includes a coordinate position, which is the same as the coordinates of the network switch 501 to which the node 502 is connected. Therefore, the network switch 501 to which the node 502 is connected is obvious from the node name. When the names of the coordinate positions of the node names are the same, it indicates that they are connected to the same network switch 501 and the distance between the nodes is “0”. On the other hand, if the part of the coordinate position in the node name is different, it indicates that it is connected to a different network switch 501.

  FIG. 15 shows the distance between objects. As described above, the torus structure has a characteristic of returning to the original reference point when traveling in the same direction from the reference point. The one-dimensional torus structure illustrated in FIG. 15 includes Mx objects 401 and uses the object 1501 as a reference point. When considering the distance from the reference point to the object 1502 that has advanced n clockwise, there are two types of distances: a clockwise distance and a counterclockwise distance.

  The distance between the object 1501 and the object 1502 is “n” when proceeding clockwise and “Mx−n” when proceeding counterclockwise. Here, when n is smaller than Mx / 2, the distance traveled clockwise is small, and when larger than Mx / 2, the distance traveled counterclockwise is small. The distance between the objects takes two values depending on the path that travels in this way, but a smaller value that is less than or equal to Mx / 2 is defined as the distance between the objects. The object 401 includes a network switch 501 and a plurality of nodes 502, and the distance between the objects is equal to the distance between the two network switches and the distance between the two nodes. The distance here means the number passing through the network switch 501 in the PC cluster system 1 having a three-dimensional torus structure.

  FIG. 16 is a diagram showing a method for calculating the distance between nodes of the PC cluster system 1 using a three-dimensional torus structure. The PC cluster system 1 using a three-dimensional torus has a torus structure in each of the x, y, and z directions by the object 401. In the example of the PC cluster system in FIG. 16, it is composed of 10 objects in the x direction (Mx = 10), 10 objects in the y direction (My = 10), and 10 objects in the z direction (Mz = 10). Suppose you are.

  If the object 401 at the coordinate position (0, 0, 0) has a node 1601 and this node number is 07, the node name can be defined as N000_07 by the definition illustrated in FIG. If the object 401 at the coordinate position (1, 1, 1) has a node 1602 and the node number is 01, the node name is similarly defined as N111_01. When the object 401 at the coordinate position (9, 0, 1) has a node 1603 and the node number is 06, the node name is defined as N901_06.

  Here, the distance between the node 1601 and the node 1602 is considered. The coordinate position of the node 1601 is (0, 0, 0) as shown in the node name, and the coordinate position of the node 1602 is (1, 1, 1). When considering the distance in the x direction, 1-0 = 1 in the right direction. Since Mx = 10 in the left direction, 10−1 = 9. Since the smaller value is the distance, the distance between the node 1601 and the node 1602 in the x direction is 1. Similarly, the distance in the y direction is 1, and the distance in the z direction is 1. Here, the distance between the nodes is a value obtained by adding all the distances in the x direction, the y direction, and the z direction, and the distance between the nodes 1601 and 1602 is 3 obtained by adding the distances.

  Similarly, the distance between the object 1601 and the object 1603 is obtained. The coordinate position of the node 1601 is (0, 0, 0), and the coordinate position of the node 1603 is (9, 0, 1). When considering the distance in the x direction, 9−0 = 9 in the right direction. Since Mx = 10 in the left direction, 10−9 = 1. In order to employ a small value, the distance in the x direction between the node 1601 and the node 1603 is 1. Similarly, the distance in the y direction is 0, and the distance in the z direction is 1. Thus, the distance between the node 1601 and the node 1603 is 2. Finally, the distance between the node 1602 and the node 1603 is obtained. The coordinate position of the node 1602 is (1, 1, 1), and the coordinate position of the node 1603 is (9, 0, 1). When considering the distance in the x direction, 9-1 = 8 in the right direction. Since Mx = 10 in the left direction, 10−9 + 1 = 2. In order to employ a small value, the distance between the node 1602 and the node 1603 in the x direction is 2. Similarly, the distance in the y direction is 1, and the distance in the z direction is 0. Thus, the distance between the node 1602 and the node 1603 is 3.

  The actual procedure of the process allocation method in this embodiment will be described below with reference to the drawings. Various operations corresponding to the process allocation method described below are realized mainly by the job management program 304 executed by the job management node 11. And such a program is comprised from the code | cord | chord for performing the various operation | movement demonstrated below.

  FIG. 17 is a diagram illustrating a processing flow example of the scheduling function 306 and the job control function 305 of the job management program 304 in the job management node 11 together with the user operation procedure when executing the parallel program.

  The user logs in to the login node 10 using terminal emulator software or the like through a network outside the large-scale PC cluster system 1 (S1701), and performs an interactive operation on the login node 10. Further, the user creates input data of the parallel program on the log-in node 10 on the file system (S1702), and then performs an operation for submitting the job to execute the parallel program execution instruction (S1703). The execution instruction (S 1703) on the login node 10 is immediately notified via the network 14 to the job management program 304 that always operates in the job management node 11.

  On the other hand, the scheduling function 306 of the job management program 304 operating on the job management node 11 obtains queue information defining the calculation node configuration to be executed through a file or database (S1707), and the program specified by the user. Compare the requirements for execution with the queue configuration. Job scheduling (S1708) is performed using the result and the usage status information of the computing node 12 provided by the job control function 305, and information for parallel program execution is managed in units of jobs. The job management program 304 sets a unique identification number called a job ID for each job.

  The scheduling function 306 of the job management node 11 registers a job generated by the job scheduling (S 1708) for a job queue based on FIFO (First In First Out) (S 1709). The registered job can be acquired from the job queue 307 using the job ID as a key.

  In addition, the scheduling function 306 returns the job ID of the job as a response message to the user (the operation terminal) who performed the interactive operation at the login node 10 and submitted the job.

  The job control function 305 that has detected a program that can execute a program without executing another job program among all the calculation node candidates that match the execution requirement is executed in relation to the calculation node candidate that executes the program. Initialization (S1710) is performed, and the status of the job is changed from “Waiting for execution” to “Starting execution” for the information registered in the job queue 307 by the job registration (S1709).

  Thereafter, the job control function 305 generates a job execution program process for the computation node candidate, and starts execution of the program by job execution (S1711). At this time, the status of the job is changed from “execution start” to “under execution” in the job queue 307. The user of the login node 10 can check the execution status of the job by the job ID. In this case, for example, the job control function 305 receives information on the status of the job having the corresponding job ID from the job queue 307. Is read and displayed. Further, the job control function 305 that has detected the end of the program being executed in the computing node 12 performs the job execution end (S1712) process. In the process of job execution end (S1712), the program execution result information is transmitted to the user account of the login node 10 by e-mail or the like, and from the “execution in progress” to the “execution end” in the job queue 307. Change the job status to "". The user of the login node 10 (at a predetermined terminal) receives the e-mail transmitted from the job management program 304, and browses the contents and the reference of the result file output by the program on the login node 10. The result is confirmed (S1705).

  FIG. 18 is a diagram showing a flow of processing for adding an MPI rank determination function to the scheduling function 306 for the purpose of minimizing the communication time in the MPI program. The scheduling function 306 of the job management program 304 executed in the job management node 11 acquires information on the job queue 307 from a file, a database, or the like as definition information of the calculation node configuration to be executed (S1801). Next, the scheduling function 306 acquires the execution status of the calculation nodes 12 constituting the job queue 307 using the execution status acquisition function of the calculation nodes 12 included in the job control function 305 (S1802).

  The scheduling function 306 then performs job scheduling based on information such as the execution status of the computing node 12 and the restriction conditions of the job queue 307, the priority of jobs that have already been scheduled, and requirements specified for parallel program execution. (S1803) is performed, and information (job) to be registered in the job queue 307 is generated.

  In the conventional scheduling function 306, the job registration (S1804) completes the scheduling function 306 of the job management program 304. In this method, the calculation node determined by job scheduling (S1803) before job registration (S1804) is used. Therefore, the general MPI program determines the MPI rank (S1805) for shortening the communication time.

  The process of determining the MPI rank to be added to the scheduling function 306 (S1805) consists of four processes. First, the process of determining the calculation node 12 having the MPI rank of “0” (S1806) is performed. In a general MPI program, a calculation node 12 that arranges a process having an MPI rank “0” that often plays a special role different from other nodes is determined.

  In the next process of the MPI rank determination process (S1805), the table for holding information for minimizing the communication path by executing the MPI collective communication function is initialized (S1807). Processing for determining a calculation node other than “0” is performed (S1808). These processes determine the MPI process arrangement for shortening the communication time of the MPI collective communication function. The MPI rank for shortening the communication time of the entire MPI program is determined by adding the processing so far. In the last process added to the process of determining the MPI rank (S1805), a process of outputting the determined MPI rank to the outside such as a file (S1809) is performed. Also, in the job registration process (S1804), after the information registered in the job queue 307 is associated with the information of the file to which the MPI rank is output, the information is registered in the job queue 307. For jobs registered in the job queue 307, changes in the state of the computation node 12 (node failure, node addition, etc.), changes in job execution status (forced termination, job freeze, priority change), etc. Therefore, rescheduling may be performed to change the computation node 12 once determined. These processes to be added according to this method are performed again from S1806 to S1809 with respect to newly scheduled calculation node candidates when rescheduling that changes the calculation node 12 occurs. Then, MPI rank information that matches the job registered in the job queue 307 is generated.

  FIG. 19 is a configuration diagram of the PC cluster system 1 having a three-dimensional torus structure for deriving the concept of processing for determining a calculation node 12 having an MPI rank of “0”. In many cases, a process with an MPI rank of “0” performs special processing compared to other processes. For example, the process of accessing an input / output file is performed by a process with an MPI rank of “0” representing all processes, or input data read by a process with an MPI rank of “0” is read by an MPI such as an MPI_BCAST function. For example, a collective communication function can be used to distribute to all processes.

  The short distance between the computation node 12 that executes file input / output and the I / O node 13 is advantageous in terms of shortening the input / output time and efficient use of network resources. In the PC cluster system 1 using the three-dimensional torus in FIG. 19, it is assumed that the I / O node 1902 is at the node number 07 at the coordinate position (0, 0, 0). On the other hand, the execution of the parallel program uses the computation node 12 in any object in the job execution space 802. Which of the calculation nodes 12 in the job execution space 802 should be set to MPI rank “0” is determined by calculating the distance between objects and specifying the node having the smallest distance. The nodes thus identified can perform file input / output with the minimum amount of communication most efficiently.

  Here, the equation for calculating the distance between the objects is Equation 1903, and the coordinate position included in the I / O node name when following the rule of the node name illustrated in FIG. 14 is (x1, y1, z1). The coordinate position included in the calculation node name is (x2, y2, z2). The job management node 11 calculates the distance between objects by applying these two coordinates to the formula 1903. The distances are obtained from combinations of all the calculation nodes 12 and the I / O nodes 13, and the calculation node 12 having the shortest distance is determined as a calculation node that executes the process of the MPI rank “0”.

  The job execution space 802 in FIG. 19 is composed of only calculation objects, but when an I / O object is included, the calculation node 12 included in the I / O object is the calculation node with the shortest distance (I / O). In this case, the computation node 12 includes the MPI rank “0” in the I / O object. The I / O object includes a plurality of calculation nodes 12, and the node number of the calculation node 12 having the MPI rank “0” is the calculation node 12 having the smallest node number of the calculation nodes 12 in the object.

  FIG. 20 is a flowchart showing the minimum distance calculation in the MPI collective communication function execution according to the present embodiment, and is a flowchart for obtaining the calculation node 12 for executing the process of setting the MPI rank “0” by the job management node 11. . This flow shows the detailed processing contents of the node determination (S1806) of the MPI rank “0” in FIG.

  Here, the job management node 11 initializes the minimum distance between the nodes, the node number and the three-dimensional coordinate position regarding the calculation node 12 and the I / O node 13 (S2002). Subsequently, the job management node 11 acquires the coordinate information of the I / O node 13 constituting the PC cluster system 1 from, for example, a definition file or database in the storage unit 301 (S2003). Similarly, the job management node 11 acquires the coordinate information of the calculation node 12 from the definition file, the database, or the like (S2004). Through these processes, the job management node 11 prepares a list in which all I / O node names are stored and a list in which all calculation node names are stored.

  Next, the job management node 11 obtains the calculation node 12 having the MPI rank “0” by the double loop structure process using the I / O node name list and the calculation node name list. The iterative process S2005 is an iterative process using a list of I / O node names, and the job management node 11 disassembles the I / O node name in the loop and obtains the coordinates and node number of the I / O node 13 (S2006). . Next, in step S2007, the job management node 11 repeats the calculation node name list, decomposes the calculation node name in the loop, and acquires the coordinates and node number of the calculation node 12 (S2008). .

  Subsequently, the job management node 11 calculates the distance between the nodes from the I / O node coordinates and the calculation node coordinates acquired in the processes of S2006 and S2008 (S2009). The job management node 11 compares the distance calculated in S2009 with the minimum distance between the nodes obtained so far in the determination process S2010, and if the distance calculated in S2009 is smaller than the existing minimum distance (S2010). : Yes), the minimum distance between the nodes, the calculation node coordinates and the node number, and the I / O node coordinates and the node number are updated with the information of the corresponding node (S2011). The job management node 11 performs the repetition end S2012 for the calculation node 12 and the repetition end S2013 for the I / O node 13, and performs the above processing for all combinations of the I / O node 13 and the calculation node 12. Finally, the I / O node coordinate and node number with the shortest distance and the calculation node coordinate and node number are determined, and the calculation node 12 obtained here is set as the calculation node having the MPI rank “0”.

  FIG. 21 is a diagram showing table information of a list of calculation nodes handled in the process of determining the MPI rank. The calculation node list table 2101 has the same number of elements, that is, records, as the number of processes executing the parallel program. Each process has three pieces of information. One is the name of a computation node (nodename) that executes the process. The column of the calculation node name includes two pieces of information including the object including the calculation node, that is, the three-dimensional coordinate position of the network switch 501 and the node number. The column of group ID (groupid) is used for work in executing the MPI rank determination process, and the initial values are all “0”. The MPI rank column (mpirank) sets the determined MPI rank, and the initial values are all “−1”. A negative MPI rank value indicates that the MPI rank has not yet been determined, and a value of 0 or more indicates that the MPI rank has been determined.

  The calculation node list table 2101 in the upper part of FIG. 21 is a table in an initial state, in which all group IDs are set to “0” and all MPI ranks are set to “−1”. On the other hand, when the process described with reference to FIG. 20 is performed and the node name having the MPI rank “0” is determined, the job management node 11 sets the MPI rank column value to “0” for the determined calculation node 12. Set to.

  FIG. 22 is a flowchart for initializing the calculation node list table 2101 and shows the detailed processing contents of the initialization (S1807) of the rank table of FIG. First, the job management node 11 acquires a list of calculation node candidates determined by the scheduling function 306, and sets it to “Nodelist” in a predetermined storage area (S2201). Next, the job management node 11 secures the calculation node list table 2101 in the memory, and initializes the counter i set in the predetermined storage area to zero (S2202). Next, the job management node 11 sets the node name (nodename) of the calculation node list table 2101 inside the iterative process S2203 to the calculation node name obtained by the index i from the list of calculation node lists, and the calculation node list table 2101. The group ID (groupid) is initialized to zero (S2204). Next, in the comparison process S2205, the job management node 11 has the MPI rank determined by the determination of the node of the MPI rank “0” (S1806) and the MPI rank “0” obtained from the calculation node list with the index i. Is compared with the calculation node name.

  If these node names do not match (S2205: no), the job management node 11 sets “−1” representing MPI rank undecided in the MPI rank (mpirank) of the calculation node list table 2101 (S2206). On the other hand, if the node names match (S2205: yes), the job management node 11 sets the MPI rank value “0” in the MPI rank (mpirank) of the calculation node list table 2101 (S2207).

  Thereafter, the job management node 11 updates the counter i (S2208), and performs the iterative process as long as it can access the list of the calculation node list with the index i by the end determination of the iterative process S2203 (S2209). When all the processes are completed, the calculation node names determined by the scheduling function 306 are set to the calculation node names of all elements of the calculation node list table 2101, and “0” is set to the group ID (groupid). . “0” is set only for the MPI rank (mpirank) of the computation node 12 whose MPI rank is 0, and “−1” is set for the MPI rank (mpirank) of the other computation nodes.

  FIG. 23 is a diagram illustrating a function specification of processing in which the job management node 11 determines an MPI rank other than MPI rank “0”. Here, the processing for performing MPI rank determination (S1808) other than rank 0 in FIG. 18 is set as one function, and this function updates the information stored in the table in FIG. 21 while the MPI rank is other than “0”. The calculation node that executes the process is determined. The function name is “Rank Decision” and requires five arguments. This function performs a recursive call and sets the depth of the recursive call to the fifth argument “depth” (the initial depth is 0).

  The calculation node list table shown in FIG. 21 is designated as the first argument of the Rank Decision function. The number of elements (number of processes) in the calculation node list table shown in FIG. 21 is designated as the second argument. The reference rank is designated as the third argument, and only the node having the MPI rank “0” has been determined at the first call, and “0” is designated. The fourth argument is a target group ID, and “0”, which is an initialized value, is designated by the process of FIG. 22 at the first call. The fifth argument specifies the number of recursive calls of the function, and “0” is specified at the first call. The Rank Decision function obtains the MPI rank of the other party that communicates with the reference rank (base node) when executing collective communication with the depth “depth” within the range of the specified group ID. At this time, the group ID is divided into two, and one calls the Rank Decision function with the group ID that communicates with the reference rank at the next depth. The other updates the new group ID, sets the MPI rank of the partner determined this time as a new reference rank, and calls the Rank Decision function with a new group ID that communicates with the new reference rank at the next depth.

  24 and 25 are flowcharts of processing for determining an MPI rank other than the MPI rank “0”, and shows the detailed processing contents of determination other than the rank 0 node (S1808) in FIG. In this case, the job management node 11 uses the Rank Decision function 2301 when determining a calculation node other than the MPI rank “0”. The Rank Decision function 2301 performs processing according to the function specifications shown in FIG. The MPI rank of the computation node 12 whose MPI rank has already been determined is set in “basenode” given by the argument, and the MPI rank “0” determined by the processing of FIG. 20 is set when the Rank Decision function is called for the first time. .

  In this case, the Rank Decision function first acquires the name of the computation node 12 on which the MPI rank process designated by “basenode” is executed (S2401). Next, the Rank Decision function obtains the distance between the previously obtained calculation node name and each calculation node having the same group ID, and creates a list of distances between the nodes. The Rank Decision function stores the node name having the maximum distance obtained here (S2402). The node name with the largest distance is the computing node 12 that needs the most data transfer from the computing node 12 of “basenode” via the network switch 501.

  Next, the Rank Decision function obtains and stores the distances between the farthest node and all the nodes with the same group (S2403). As a result, the Rank Decision function calculates two distances, that is, the distance from “basenode” and the distance from the node farthest from “basenode” for the nodes having the same group ID. Next, the Rank Decision function divides the group into two based on the two distances obtained up to step S2403 (S2404). So far, what belonged to the same group is divided into two. For example, group A is closer to “basenode” and group B is closer to the node farthest from “basenode”. When the number n of nodes is an even number, the number of nodes belonging to the A group and the B group is n / 2. When the number n of nodes is an odd number, the last remaining node is compared with the distance from “basenode” and the distance from the node farthest from “basenode”, and belongs to the smaller group.

  Next, the Rank Decision function sets a new group ID (S2405). A group close to “basenode” updates the group ID with a new group ID that is a value obtained by doubling the group ID. In the first call, the group ID is “0”, so the group ID that is doubled and updated is also “0”. For the group belonging to the node farthest from “basenode”, a value obtained by doubling the group ID and adding 1 is set as a new group ID. Since the group ID is “0” in the first call, the updated group ID is “1”, and “1” is set as the new group ID.

  By performing step S2405 when the group ID is “0”, the node group is divided into a group “0” and a group “1” as new group IDs, and the group is set.

  Subsequently, the Rank Decision function searches for a node closest to “basenode” in the group belonging to the node farthest from “basenode” (S2406). The MPI rank is set in the table 2102 for the node searched in the process of S2406. The MPI rank set here is a unique value depending on “basenode”, “tablecnt”, and “depth” passed as arguments. In the Rank Decision function, after setting the MPI rank, if the number of groups B divided into two is larger than “0” (S2407: yes) by the determination process S2407, the Rank Decision function is called for the group B (S2408). The Rank Decision function that has been recursively called performs the processes from S2401 to S2406 so far, divides the group and determines one MPI rank, and further calls the Rank Decision function as necessary. If the number of groups A divided into two is larger than 0 by the determination process S2409 after the determination process S2407 (S2409: yes), the Rank Decision function is called for the group A (S2410). The Rank Decision function that has been recursively called performs the processes from S2401 to S2406 so far, divides the group and determines one MPI rank, and further calls the Rank Decision function as necessary. In the recursive call, the table and size updated in accordance with the MPI rank determination, the determined MPI rank, the group ID of the group belonging to the node closer to “basenode”, and a value obtained by adding 1 to “depth” as arguments of the Rank Decision function Will pass.

  Although the best mode for carrying out the present invention has been specifically described above, the present invention is not limited to this, and various modifications can be made without departing from the scope of the invention.

  According to the present embodiment, it is possible to automatically and optimally arrange processes for reducing the execution time of a parallel program using an MPI library on a large-scale PC cluster system having a three-dimensional torus structure.

  At least the following will be clarified by the description of the present specification. That is, in the process allocation system, the network topology based on the three-dimensional torus structure has an annular structure in each of the x, y, and z axis directions, and communication between two nodes via one or more network switches. The position coordinate information of each node in the storage unit is three-dimensional according to the relative position coordinate with respect to the origin coordinate in each network switch when the predetermined network switch is set to the three-dimensional origin coordinate. Coordinates are defined, and the calculation unit calculates the number of network switches on the path when communication is performed between nodes, and the total difference of x, y, z coordinates between the three-dimensional coordinates of each node. It may be calculated and obtained.

  In the process allocation system, the calculation unit calculates the number of network switches on the path to the node of the MPI rank “0” for each calculation node, and according to the number of processes designated when the MPI program job is submitted. The number of calculation nodes may be selected in ascending order of the calculated number of network switches.

  In the process allocation system, when the MPI collective communication function is executed, the arithmetic unit selects a calculation node group selected according to the number of processes as a network switch on a path with a node of the MPI rank “0”. Depending on the number of the groups, the group having the smaller number of network switches is divided into two groups: a group A having a node with MPI rank “0” as a base node and a group B having the larger number of network switches. Dividing processing and among the calculation nodes included in the group B, the one with the smallest number of network switches is first transferred from the node of the MPI rank “0” in the group B when the MPI collective communication function is executed. Processing to identify as a base node to receive and those of the groups A and B Re respect, as long as the remaining computing nodes other than the base node may be and a process of repeatedly executing the specific processing process and the base node of the grouping.

  In the process allocation system, the computing unit, when implementing MPI, for the computation node group selected according to the number of processes, at least information on the computation node included in the computation node group and the base node of each group. A file may be output as MPI rank information.

1 Large-scale PC cluster system 10 Login node 11 Job management node (process allocation system)
12 computing node 13 I / O node 14 network 201 computing node storage unit 202 computing node CPU unit 203 computing node network interface unit 204 parallel program 301 job management node storage unit 302 job management node CPU unit (calculation unit)
303 Network interface unit 304 of the job management node 304 Job management program 401 Object 501 Network switch 802 Node space 803 secured by the job management program 804 Object composed only of computation nodes 804 Object composed of computation nodes and I / O nodes

Claims (7)

In a large-scale PC cluster system in which network switches connected to a plurality of nodes are connected to each other in a three-dimensional torus structure, the information processing system performs process allocation for executing parallel programs targeting MPI programs,
A storage unit that holds the position coordinates of each node in the three-dimensional torus structure;
Based on the position coordinates of each node in the three-dimensional torus structure in the storage unit, each calculation node connected to each network switch for performing process assignment upon execution of a parallel program, and when a file is input / output by the calculation node Calculating the number of network switches in the path of the three-dimensional torus structure between each I / O node that is the access target, and specifying the set of I / O node and calculation node that minimizes the calculated value, An arithmetic unit that executes a process for determining the calculation node of the MPI rank as “0” using the specified set of calculation nodes as a reference for parallel program execution;
A process allocation system comprising: The network topology based on the three-dimensional torus structure has a ring structure in each of the x, y, and z axis directions, and communication via one or more network switches is executed in communication between two nodes.
The information of the position coordinates of each node in the storage unit is defined as a three-dimensional coordinate by a relative position coordinate with the origin coordinate in each network switch when a predetermined network switch is a three-dimensional origin coordinate,
The arithmetic unit obtains the number of network switches on a path when communication is performed between nodes by calculating a total of respective differences of x, y, and z coordinates between the three-dimensional coordinates of each node. The process allocation system according to claim 1.   The calculation unit calculates the number of network switches on the path to the node having the MPI rank of “0” for each calculation node, and calculates the number of calculation nodes according to the number of processes specified when the job of the MPI program is input. 3. The process allocation system according to claim 1, wherein the selected network switches are selected in ascending order. The computing unit is
When executing the MPI collective communication function, the number of network switches depends on the number of network switches on the path with the MPI rank “0” node serving as a reference for the calculation node group selected according to the number of processes. A grouping process that divides the group A into two groups, the group A having the MPI rank “0” as a base node and the group B having the larger number of network switches, which is the smaller group,
Among the calculation nodes included in the group B, a node having the smallest number of network switches is set as a base node that first receives data transfer from the node of the MPI rank “0” when executing the MPI collective communication function in the group B. Process to identify,
For each of the groups A and B, as long as computation nodes other than the base node remain, the grouping process and the process of recursively executing the specific process of the base node are executed.
The process allocation system according to claim 3.   When the MPI is implemented, the calculation unit includes at least MPI rank information of the MPI rank “0” included in the calculation node group and the base node information of each group included in the calculation node group selected according to the number of processes. 5. The process allocation system according to claim 4, wherein a file is output as rank information. In a large-scale PC cluster system in which network switches connected to a plurality of nodes are connected to each other in a three-dimensional torus structure, in the storage unit, the third order is assigned in order to perform process allocation for parallel program execution targeting MPI programs. A computer that holds the position coordinates of each node in the original torus structure,
Based on the position coordinates of each node in the three-dimensional torus structure in the storage unit, each calculation node connected to each network switch for performing process assignment upon execution of a parallel program, and when a file is input / output by the calculation node Calculating the number of network switches in the path of the three-dimensional torus structure between each I / O node that is the access target, and specifying the set of I / O node and calculation node that minimizes the calculated value, A process allocation method, comprising: executing a process of determining an MPI rank of the specified set of computation nodes. In a large-scale PC cluster system in which network switches connected to a plurality of nodes are connected to each other in a three-dimensional torus structure, in the storage unit, the third order is assigned in order to perform process allocation for parallel program execution targeting MPI programs. In the computer that holds the position coordinates of each node in the original torus structure,
Based on the position coordinates of each node in the three-dimensional torus structure in the storage unit, each calculation node connected to each network switch for performing process assignment upon execution of a parallel program, and when a file is input / output by the calculation node Calculating the number of network switches in the path of the three-dimensional torus structure between each I / O node that is the access target, and specifying the set of I / O node and calculation node that minimizes the calculated value, A process of determining the MPI rank of the specified set of computation nodes;
A process allocation program characterized by that.