JP6322968B2 - Information processing apparatus, information processing method, and program - Google Patents

Description

  The present invention relates to a technique for assigning execution of an application to a computing resource suitable for execution of the application.

  In general, data communication between computation nodes is performed via a communication network (hereinafter also simply referred to as “network”), and file I / O (Input / Output: input / output) is performed with respect to a storage disk.

  Depending on the execution specifications of data communication and file I / O, that is, characteristics, the network and storage disk suitable for execution of data communication and file I / O differ. When constructing a system using these elements such as network and storage disks, if a large number of elements suitable for the above characteristics are used, a system suitable for executing an application having various characteristics can be constructed (that is, the application The execution performance can be improved).

  On the other hand, a system using elements suitable for a plurality of characteristics generally has a problem that the construction cost is higher than a system using elements suitable for only some characteristics. Therefore, it is considered to construct a system suitable for a plurality of characteristics while suppressing the construction cost. For example, instead of building all the elements that make up the system with elements that can support multiple characteristics, the elements that make up the system are divided into several groups, and the characteristics that differ between groups It is possible to adopt elements.

  However, in order to construct a system by the above method, it is necessary for the end user to grasp the characteristics of the elements constituting the application and the system and to manually use the groups in the system manually.

  The above work requires knowledge of the computer and is complicated, and is difficult for an end user who is only familiar with the calculation result and is not familiar with the computer. On the other hand, the communication in the network between the computation nodes and the file I / O to the disk storage greatly affect the execution performance of the application. Therefore, it is desired to execute in a system suitable for these executions.

  In relation to the above technique, for example, Patent Document 1 discloses a heterogeneous multiprocessor system including a plurality of processor elements having different instruction sets and configurations. The multiprocessor system includes means for improving the processing of the entire multiprocessor system by preventing a shortage of resources of a specific processor element.

  Japanese Patent Application Laid-Open No. 2004-228561 discloses a scheme that is utilized for scheduling by determining program characteristics based on a source program.

Japanese Patent No. 4936517 JP 2003-271405 A

  As described above, in order to divide the elements constituting the system into several groups and construct the system so as to adopt elements having different characteristics between the groups, knowledge of computers and complicated work are required. Therefore, there is a problem that it is difficult for general end users to implement.

  Further, although the means disclosed in the above-mentioned Patent Document 1 allows tasks to operate efficiently on different types of resources, it focuses only on the processing of the processor, so communication in the network between computation nodes and file storage to disk storage It cannot be applied to I / O.

  Further, Patent Document 2 does not disclose a technique for improving the execution performance of an application without requiring knowledge of a computer and complicated work.

  The present invention has been made in view of the above problems, and mainly provides an information processing apparatus and the like that can reduce the construction cost and can improve the execution performance of the application while eliminating the trouble of the end user. With a purpose.

  The first information processing apparatus according to the present invention, when a specific description that affects an execution characteristic of an application is included in the application, generates a specific description analysis that generates additional information related to an execution frequency of an instruction or a function called by the specific description And a calculation resource allocation unit that determines a calculation resource suitable for execution of the instruction or function based on the additional information and allocates the calculation resource to execution of the application.

  The first information processing method of the present invention generates additional information regarding the execution frequency of an instruction or a function called by the specific description when the application includes a specific description that affects the execution characteristics of the application. Based on the additional information, a calculation resource suitable for execution of the instruction or function is determined, and the calculation resource is allocated to execution of the application.

  This object is also achieved by a computer program that implements the information processing apparatus or information processing method having the above-described configurations by a computer, and a computer-readable storage medium that stores the computer program. .

  According to the present invention, it is possible to reduce the construction cost and improve the execution performance of the application without taking the trouble of the end user.

It is a figure which shows the outline | summary of a structure of the information processing system which concerns on the 1st Embodiment of this invention. It is a block diagram which shows the structure of the information processing system which concerns on the 1st Embodiment of this invention. It is a figure which shows the example in which data exchange is simultaneously performed between all the processes in the information processing system which concerns on the 1st Embodiment of this invention. It is a figure which shows the example in which data exchange is performed in the shape of a ring between processes in the information processing system which concerns on the 1st Embodiment of this invention. It is a flowchart which shows operation | movement at the time of comprising all the nodes by the network topology with a high bisection bandwidth in an information processing system. It is a flowchart which shows operation | movement at the time of comprising all the nodes by the network topology with a low bisection bandwidth in an information processing system. It is a flowchart which shows the operation | movement which distributes the system which performs an application in the information processing system which concerns on the 1st Embodiment of this invention. It is a figure which shows the example of description of an application program. It is a figure which shows the example of description of an application program. It is a figure which shows the example of description of an application program. It is a flowchart which shows operation | movement at the time of comprising all the nodes by the SAS (Serial Attached SCSI) disk in the information processing system. It is a flowchart which shows operation | movement at the time of comprising all the nodes by the SATA (Serial ATA) disk in the information processing system. 6 is a flowchart showing an operation of distributing an application between a system using a SAS disk and a system using a SATA disk in the information processing system according to the first embodiment of the present invention. It is a block diagram which shows the structure of the information processing system which concerns on the 2nd Embodiment of this invention. It is a figure which shows a part of example of an application program. It is a figure which shows the instruction list which is the result of disassembling the application execution binary. It is a figure which shows the outline | summary of a structure of the information processing system which concerns on the 3rd Embodiment of this invention. It is a figure which illustrates the hardware constitutions of the information processing apparatus which can implement | achieve the information processing system which concerns on each embodiment of this invention.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

First Embodiment FIG. 1 is a diagram showing an outline of a configuration of an information processing system 100 according to a first embodiment of the present invention. As shown in FIG. 1, the information processing system 100 includes a job scheduler 110, computing system _{120 1,} 120 2, ..., and 120 _n. The application execution request is sent from the end user to the job scheduler. At this time, the end user also requests the number of CPUs (Central Processing Units) cores in each computing system and the memory capacity necessary for executing the application. In the present embodiment, an application represents a software program (computer program) that can realize a specific function executed in the information processing system 100.

  The information processing system 100 belongs to the field of HPC (High Performance Computing).

  In a system belonging to the general HPC field, software called a job scheduler is used in order to share the system among a plurality of users and maximize the amount of execution of an application.

  The job scheduler executes the requested application in the secured computing system when the computing system having the number of CPU cores and the memory capacity requested by the end user can be secured. As described above, the selection of the computing system for executing the application is automatically performed by the job scheduler.

  Here, in the HPC field, some types (languages) are often used as programming languages used to describe application programs capable of realizing applications. Some types of APIs (Application Programming Interfaces) for a library for performing distributed parallel processing in which processing is divided into a plurality of computing systems and processed in parallel are also frequently used. The specific description (details will be described later) in the programming language and the API of the library are characterized in that the system suitable for execution differs depending on the type used for them.

  In addition, an application in the HPC field generally has a feature that a part of processing is repeatedly executed. The repeated execution is described by a loop structure of a programming language.

  In this way, by examining the description in a program having a different environment suitable for execution and the loop structure where the description exists, it is possible to predict the characteristics suitable for execution of the application with high accuracy. Further, by inputting the predicted result to the job scheduler, the application can be automatically executed by a system suitable for executing the application without requiring manual operation by the end user.

  Details of the information processing system 100 according to the present embodiment will be described below.

  FIG. 2 is a block diagram showing the configuration of the information processing system 100 according to the first embodiment of the present invention. As illustrated in FIG. 2, the information processing system 100 includes a compiler 210, a job scheduler 220, and systems 230 and 240.

  The compiler 210 includes a compiler normal function unit 211, a specific description analysis unit 212, and an additional information generation unit 213. The job scheduler 220 includes an application execution receiving unit 221, an iteration count threshold setting unit 222, a loop structure depth threshold setting unit 223, an additional information analysis unit 224, and a calculation resource allocation unit 225.

  The compiler 210 outputs an application execution request 310. The application execution request 310 includes application execution binary 311, iteration number additional information 312, and loop structure depth additional information 313. The application execution request 310 is passed when an end user requests the job scheduler 220 to execute an application.

  When the compiler 210 receives the application program 200, the compiler normal function unit 211 performs a normal compilation function. The compiler normal function unit 211 generates an application execution binary 311 based on the application program 200. Subsequently, the compiler normal function unit 211 sends the application program 200 to the specific description analysis unit 212.

  The specific description analysis unit 212 examines a location where the specific description is described in the application program 200. The specific description is a description in at least a part of the program code that affects the execution characteristics of the application program among the descriptions constituting the application program. When the specific description is found, the specific description analysis unit 212 checks the depth of the loop structure where the specific description is described and the number of loop iterations.

  The specific description analysis unit 212 sends the result of the examination to the additional information generation unit 213. The additional information generation unit 213 generates the number of loop iterations as the iteration number additional information 312 and the loop structure depth as the loop structure depth additional information 313. If the loop repetition count is not statically given in the application program 200, the additional information generation unit 213 uses a unique value (for example, “−1”) that can identify the loop repetition count additional information. 312 is generated.

  The job scheduler 220 performs the same function as the normal job scheduler in the calculation resource allocation unit 225. That is, the computing resource allocation unit 225 secures a computing system that executes an application according to the application execution request 310 and instructs the execution of the application in the secured computing system.

  In addition, the job scheduler 220 holds a threshold value used to determine a tendency that a specific description in a program is executed in the iteration count threshold setting unit 222 and the loop structure depth threshold setting unit 223. The threshold value includes a threshold value used for comparison with the loop structure repetition count in the application program 200 and a threshold value used for comparison with the loop structure depth. The threshold used for comparison with the loop structure iteration count is held in the iteration count threshold setting unit 222. The threshold used for comparison with the depth of the loop structure is held in the loop structure depth threshold setting unit 223. The threshold may be stored in the job scheduler 220 in advance by the system administrator.

  The job scheduler 220 receives the application execution request 310 at the application execution receiving unit 221. The application execution receiving unit 221 passes the repetition count additional information 312 and the loop structure depth additional information 313 included in the application execution request 310 to the additional information analysis unit 224. In addition, the application execution receiving unit 221 passes the application execution request 310 to the calculation resource allocation unit 225.

  The additional information analysis unit 224 first compares the received iteration count additional information 312 with the value set in the iteration count threshold setting unit 222. If the received additional iteration number information 312 is larger than the value set in the iteration count threshold setting unit 222, the additional information analysis unit 224 applies to the characteristic that “specific description is often executed”. to decide.

  Here, if the unique value for identifying that the iteration count of the loop is not statically given in the application program 200 is set in the iteration count additional information 312, the additional information analysis unit 224 determines that the comparison Without comparison, a comparison is made regarding the loop structure depth. That is, the additional information analysis unit 224 compares the received loop structure depth additional information 313 with the value set in the loop structure depth threshold setting unit 223.

  When the loop structure depth additional information 313 is larger than the value set in the loop structure depth threshold setting unit 223, the additional information analysis unit 224 indicates that the application program 200 applies to the characteristic that “specific description is often executed”. to decide. The additional information analysis unit 224 passes the determination result to the calculation resource allocation unit 225.

  The calculation resource allocation unit 225 determines a system for executing the application based on the application execution request 310 passed from the application execution reception unit 221 and the determination result passed from the additional information analysis unit 224. The computing resource allocation unit 225 allocates a system 230 suitable for execution of the specific description for application execution having a characteristic that the specific description is frequently executed. On the other hand, the calculation resource allocation unit 225 allocates a system 240 that is not suitable for execution of a specific description for the execution of an application that does not have the characteristic that the specific description is frequently executed.

  As described above, the information processing system 100 determines a system suitable for executing an application and distributes the execution of the application to the system.

  Next, details of the above configuration will be described using a specific example.

  Here, attention is focused on a program using MPI (Message Passing Interface). In the field of HPC, as the amount of computation increases, distributed parallel computation that distributes processing to a plurality of computers exists as a commonly used technique. There are several methods for performing distributed parallel computation, and among them, it is now common to perform distributed parallel computation according to a message passing API standard called MPI. That is, targeting an application program that uses MPI in this embodiment occupies a large proportion of applications in the HPC field.

  Therefore, the operation of the information processing system 100 when an application program using MPI is executed in this embodiment will be described below.

  MPI is an API that implements distributed parallel computation by message passing, and the types of communication APIs provided by MPI are broadly classified into “one-to-one communication APIs” and “collective communication APIs”. For example, “MPI_Alltoall ()”, which is a function of “collective communication system API”, is used when data is exchanged between all processes (0 to 3) at the same time as shown in FIG. On the other hand, “MPI_Sendrecv ()”, which is a function of “one-to-one communication system API”, is often used when data is exchanged between processes (0 to 3) as shown in FIG. As for the bi-sectional bandwidth of the network (bi-section bandwidth: divided bandwidth), the execution of the collective communication system API requires a higher bi-section bandwidth compared to the execution of the one-to-one communication system API. . The bi-sectional bandwidth is obtained by dividing the network into two and adding up the bandwidths of all the links between them. Here, for example, a system configured with a network topology having a bi-sectional bandwidth higher than a predetermined threshold is referred to as a “system with a high bi-sectional bandwidth”. Further, a system configured with a network topology having a bisexual bandwidth lower than a predetermined threshold is referred to as a “system with a low bisexual bandwidth”.

  For example, consider a system that executes an application using 1000 nodes. As shown in FIG. 5, when all 1000 nodes are configured with a network topology having a bi-sectional bandwidth higher than a predetermined threshold, the collective communication system API provided in the MPI can It is executed without degradation.

  However, in general, a system having a high bisectional bandwidth is characterized by a higher construction cost than a system having a low bisectional bandwidth.

  On the other hand, as shown in FIG. 6, if all 1000 nodes are configured with a network topology having a low bi-sectional bandwidth, the construction cost of the system can be reduced, but the execution performance of the collective communication API provided by MPI is low. The problem of deteriorating arises.

  Therefore, in the present embodiment, as shown in FIG. 7, an operation in which the information processing system 100 distributes a system that executes an application program using MPI will be described. In other words, the job scheduler determines the characteristics of the application program using MPI, and distributes the execution of the program to a system having a high bisection bandwidth and a system having a low bisection bandwidth based on the determination result.

  FIG. 8 shows a description (program code) of an application program (source program) that calls a function “MPI_Alltoall ()” that is a collective communication system API provided by MPI at a position (a) having a loop structure depth “3”. It is a figure which shows some examples. When the application program shown in FIG. 8 is supplied to the compiler 210 shown in FIG. 2, the compiler normal function unit 211 of the compiler 210 generates an application execution binary 311 by performing compilation.

  The compiler normal function unit 211 also sends the application program to the specific description analysis unit 212. The specific description analysis unit 212 analyzes the presence or absence of the collective communication system API in the application program. As shown in FIG. 8, since the application program calls “MPI_Alltoall ()”, the specific description analysis unit 212 detects it.

  Subsequently, the specific description analysis unit 212 analyzes the loop structure at the place where “MPI_Alltoall ()” is called. At the position (a) shown in FIG. 8, the depth of the loop structure can be identified as “3”. On the other hand, the number of loop iterations is indicated as “N1”, “N2”, “N3”, and is not given statically. Therefore, the specific description analysis unit 212 sets “−1” indicating that the loop iteration count is not statically given as an analysis result. The specific description analysis unit 212 sends the analysis result to the additional information generation unit 213. Here, “−1” is generated as the additional number of iterations information 312 and “3” is generated as the loop structure depth additional information 313.

  As described above, the application execution binary 311, the iteration count additional information 312, and the loop structure depth additional information 313 generated by the compiler 210 are passed to the job scheduler 220 as the application execution request 310. Here, “1000” is set in the iteration count threshold setting unit 222 of the job scheduler 220 and “2” is set in the loop structure depth threshold setting unit 223 in advance by the system administrator.

  The job scheduler 220 receives the application execution request 310 at the application execution reception unit 221. The application execution receiving unit 221 sends an application execution request 310 to the calculation resource allocation unit 225. In addition, the application execution reception unit 221 sends the repetition count additional information 312 and the loop structure depth additional information 313 included in the application execution request 310 to the additional information analysis unit 224.

  The additional information analysis unit 224 sets the received iteration count additional information 312 and loop structure depth additional information 313 to the value set in the iteration count threshold setting unit 222 and the loop structure depth threshold setting unit 223, respectively. Compare with the value obtained. Then, the additional information analysis unit 224 determines the characteristics of the application program based on the comparison result.

  Here, since “−1” indicating that the loop iteration count is not statically given is set in the iteration count additional information 312, the additional information analysis unit 224 adopts the comparison result regarding the loop iteration count. Without adopting the comparison results regarding the loop structure depth. That is, when the value “2” set in the loop structure depth threshold setting unit 223 is compared with the value “3” of the loop structure depth additional information 313, the value of the loop structure additional information 313 is larger. Therefore, the additional information analysis unit 224 determines that “the collective communication system API provided by MPI is often executed”.

  The additional information analysis unit 224 sends the determination result to the calculation resource allocation unit 225. The calculation resource allocation unit 225 determines that the system (system 230 shown in FIG. 2) with the high bi-sectional bandwidth shown in FIG. Assign to the execution of the application program. The system 230 executes the application program. With the above operation, the information processing system 100 according to the present embodiment distributes systems that execute application programs using MPI.

  Next, an example of executing an application program different from the above will be described.

  FIG. 9 is a diagram illustrating an example of another description of an application program using MPI. In the application program shown in FIG. 9, the number of loop iterations is statically given at the location (b) where the MPI collective communication API is described. That is, the number of loop iterations is given as “1000”. Furthermore, since the loop at the location is described twice, the total number of repetitions is “1000000”.

  When the application program shown in FIG. 9 is received, the information processing system 100 operates in the same manner as described above. That is, the compiler 210 causes the additional information generation unit 213 to generate “1000000” as the iteration count additional information 312 and “2” as the loop structure depth additional information 313. As described above, since “1000” is set in the iteration count threshold setting unit 222 of the job scheduler 220, the additional information analysis unit 224 determines that “the collective communication system API included in MPI is often executed”.

  Furthermore, an example of executing an application program different from the above will be described.

  In the field of HPC, application programs are often written in FORTRAN (registered trademark), which is a language created for the purpose of performing scientific and technical calculations. Therefore, here, focusing on an application program described in FORTAN, an operation when such an application program is executed by the information processing system 100 according to the present embodiment will be described.

  FIG. 10 is a diagram illustrating an example of the description of an application program described in FORTAN. File I / O by FORTRAN is implemented by associating a target file with an identifier used in a program by calling an OPEN statement, and then calling the WRITE statement or the READ statement by specifying the identifier.

  At this time, in the present embodiment, the value specified in the ACCESS specifier when the OPEN statement is called is changed according to the regularity of the file I / O. That is, when the regularity of the file I / O is sequential file access, “SEQUENTIAL” is specified in the ACCESS specifier. On the other hand, when the regularity of the file I / O is random access, “DIRECT” is designated in the ACCESS specifier. If the description of the ACCESS specifier is omitted, it may be equivalent to the case where “SEQUENTIAL” is specified.

  Here, as a storage disk (storage device) as a file input / output destination, a SATA (Serial ATA) disk and a SAS (Serial Attached SCSI) disk are often used today. When a SATA disk and a SAS disk are compared, the random access performance is generally superior to the SAS disk, although there is no significant difference in sequential access performance. In addition, the SATA disk has a feature that it can be obtained at a lower cost than the SAS disk.

  For example, consider a system that executes an application using 1000 nodes. As shown in FIG. 11, when all 1000 nodes are configured by SAS disks, the file I / O of the application program described by FORTRAN is executed without any performance degradation, regardless of which calculation node is used to execute the application. The However, in this case, there is a problem that the construction cost of the system becomes high.

  On the other hand, when all 1000 nodes are configured with SATA disks as shown in FIG. 12, the construction cost of the system becomes low, but on the other hand, there is a problem that the execution performance of the random access file I / O is lowered.

  Therefore, in the following, as shown in FIG. 13, the characteristics of the file I / O are determined based on the description of the OPEN statement in the application program described in FORTAN, and the program is programmed in any system based on the determination result. The operation for distributing the executions of will be described.

  FIG. 10 shows that an WRITE statement is executed at a position where the depth of the loop structure is “3” for a file in which an OPEN statement is executed by a specifier “ACCESS = DIRECT” in an application program described by FOTRAN. The part to do is shown.

  When this application program is input to the compiler 210 shown in FIG. 2, the compiler normal function unit 211 generates an application execution binary 311. Next, the compiler normal function unit 211 sends the application program to the specific description analysis unit 212. The specific description analysis unit 212 checks whether there is an OPEN statement including a specifier of “ACCESS = DIRECT”.

  In the application program shown in FIG. 10, since the OPEN statement including the specifier “ACCESS = DIRECT” is called, the specific description analysis unit 212 detects the portion in which it is described. Next, the specific description analysis unit 212 analyzes the loop structure of the portion where the WRITE statement or the READ statement is executed in the file in which the OPEN statement is executed by “ACCESS = DIRECT”.

  In the program shown in FIG. 10, the depth of the loop structure at the location where the WRITE statement is executed is “3”. On the other hand, the number of loop iterations is described as “N1”, “N2”, “N3” and is not given statically. Therefore, the specific description analysis unit 212 outputs the analysis result by setting the depth of the loop structure as “3” and the number of loop repetitions as “−1” indicating that it is not given statically. The specific description analysis unit 212 sends the analysis result to the additional information generation unit 213. Here, “−1” is generated as the additional number of repetitions information 312, and “3” is generated as the loop structure depth additional information 313.

  The compiler 210 passes the generated application execution binary 311, the iteration count additional information 312, and the loop structure depth additional information 313 to the job scheduler 220. In this case, “1000” is set in the iteration count threshold setting unit 222 of the job scheduler 220 and “2” is set in the loop structure depth threshold setting unit 223 in advance by the system administrator.

  Upon receiving the application execution request 310, the application execution reception unit 221 passes the application execution request 310 to the calculation resource allocation unit 225. In addition, the application execution reception unit 221 sends the repetition count additional information 312 and the loop structure depth additional information 313 included in the application execution request 310 to the additional information analysis unit 224. The additional information analysis unit 224 first compares the value set in the iteration count threshold setting unit 222 with the value of the iteration count additional information 312. Here, since the value of the iteration count additional information 312 is “−1” indicating that the loop iteration count is not statically given, the additional information analysis unit 224 does not employ the comparison result regarding the iteration count. . Subsequently, the additional information analysis unit 224 compares the value “2” set in the loop structure depth threshold setting unit 223 with the value “3” of the loop structure depth additional information 313. As a result, the value of the loop structure additional information 313 is larger. Therefore, the additional information analysis unit 224 determines that “there are many I / Os to the file that has executed the OPEN statement by random access”.

  The additional information analysis unit 224 passes the determination result to the calculation resource allocation unit 225. Based on the determination result, the calculation resource allocation unit 225 allocates a system having a SAS disk shown in FIG. 13 to the application program. A system having a SAS disk executes an assigned application program.

  As described above, according to the first embodiment, the compiler 210 uses the specific description analysis unit 212 to check whether there is a specific description that affects the execution characteristics of the application program. Check the depth of the loop structure where the description is written and the number of loop iterations. If at least one of the depth of the loop structure and the number of loop repetitions is greater than the threshold value, the additional information analysis unit 224 in the job scheduler 220 applies to the characteristic that “specific description is often executed”. to decide. The computing resource allocation unit 225 allocates a system suitable for execution to an application program that is determined to have a lot of specific description execution.

  By adopting the above configuration, according to the first embodiment, it is possible to assign a system having a high bi-sectional bandwidth to the execution of an application program that frequently executes the collective communication API provided in the MPI. In addition, a system having a SAS disk can be allocated to the execution of an application program that is described in FORTRAN and that has a lot of I / O to a file that has been executed with an OPEN statement by random access.

  In other words, calculation resources suitable for application execution can be allocated based on the results of analyzing the characteristics of the programs that make up the application by using a compiler, so that it is possible to reduce the system construction cost and save effort for the end user. The effect that the execution performance of the application can be maximized is obtained.

Second Embodiment FIG. 14 is a block diagram showing a configuration of an information processing system 400 according to a second embodiment of the present invention. As illustrated in FIG. 14, the information processing system 400 includes a job scheduler 420, a system 450, and a system 460. The job scheduler 420 receives the application execution request 410.

  In the information processing system 400 according to the present embodiment, an application program is compiled by a general compiler, and an end user requests the job scheduler 420 to execute the application.

  As shown in FIG. 14, the job scheduler 420 includes an application execution receiving unit 421, a loop structure depth threshold setting unit 422, a specific description analyzing unit 423, an additional information generating unit 424, a loop structure depth additional information 425, and additional information analysis. Unit 426 and calculation resource allocation unit 427.

  The job scheduler 420 receives an application execution request 410 from an end user in the application execution receiving unit 421. The application execution request 410 specifies an application execution binary 411 compiled by a general compiler.

  The application execution reception unit 421 passes the application execution binary 411 specified in the application execution request 410 to the external disassembler 430. The disassembler 430 can be executed by an information processing apparatus that can communicate with the information processing system 400, and executes disassembly for the application execution binary 411.

  As a result of execution by disassembly 430, an instruction list 440 for the application is generated. The instruction list 440 is sent to the specific description analysis unit 423 of the job scheduler 420. Based on the instruction list 440, the specific description analysis unit 423 examines the location of the specific description that affects the execution characteristics of the application, and if the location is found, examines the depth of the loop structure at the location.

  The specific description analysis unit 423 sends the result of the above examination to the additional information generation unit 424. The additional information generation unit 424 generates a result related to the depth of the loop structure as the loop structure depth additional information 425 and sends the generated loop structure depth additional information 425 to the additional information analysis unit 426.

  The additional information analysis unit 426 compares the loop structure depth additional information 425 with the value set in the loop structure depth threshold setting unit 422. When the value of the loop structure depth additional information 425 is larger than the value set in the loop structure depth threshold setting unit 422, the additional information analysis unit 426 has a characteristic that the application program has “specific description is often executed”. Judge that it is true. The result of the determination is referred to as characteristic information.

  The additional information analysis unit 426 passes the characteristic information to the calculation resource allocation unit 427. In the loop structure depth threshold setting unit 422, a threshold is set in advance by the system administrator.

  The calculation resource allocation unit 427 determines a system for executing the application based on the application execution request 410 passed from the application execution receiving unit 421 and the characteristic information passed from the additional information analysis unit 426. In this case, the computing resource allocation unit 427 allocates a system 450 suitable for executing the specific description to the application program. The system 450 executes the assigned application program.

  Next, the operation of the information processing system 400 will be described using a specific example. Here, the call status of the collective communication API of the application is determined from the execution binary of the application using MPI, and the execution of the application is distributed to a system having a high bisection bandwidth and a system having a low bisection bandwidth. The operation will be described.

  FIG. 15 shows a part of a description of an application program that calls “MPI_Alltoall ()”, which is a collective communication system API of MPI, in a triple loop structure (c).

  An application execution binary generated as a result of compiling the source program including FIG. 15 is referred to as an application execution binary 411 shown in FIG. When the application execution request 410 including the application execution binary 411 is sent to the job scheduler 420, the job scheduler 420 receives the application execution request 410 by the application execution reception unit 421.

  The application execution reception unit 421 passes the application execution binary 411 specified in the application execution request 410 to the disassembler 430. The disassembler 430 executes disassembly on the application execution binary 411.

  As a result of the disassembly, an instruction list 440 of the application is generated. FIG. 16 is a diagram showing an instruction list in which a portion corresponding to the portion shown in FIG. 15 is extracted from the result of disassembling the application execution binary.

  The instruction “callq 400850 <MPI_Alltoall @ plt>” 431 indicates a call to “MPI_Alltoall ()”. Each range indicated by the loops 432 to 434 corresponds to each range of the for sentence shown in FIG. The instruction list 440 is passed to the specific description analysis unit 423.

  The specific description analysis unit 423 checks whether the MPI collective communication API is called based on the received instruction list 440. Here, a call to “MPI_Alltoall ()” is found. Further, the specific description analysis unit 423 examines the depth of the loop structure at the location where the MPI collective communication API is called. Here, the specific description analysis unit 423 detects the depth of the loop structure as “3”.

  The specific description analysis unit 423 sends the result of the above examination to the additional information generation unit 424. The additional information generation unit 424 generates “3” as the loop structure depth additional information 425 and sends it to the additional information analysis unit 426.

  Here, “2” is set in advance in the loop structure depth threshold setting unit 422 by the system administrator. The additional information analysis unit 426 compares “2” set in the loop structure depth threshold setting unit 422 with “3” that is the loop structure depth additional information 425. Since the loop structure depth additional information 425 is larger, the additional information analysis unit 426 determines that the characteristic of “the number of calls of the collective communication system API” is applicable.

  The additional information analysis unit 426 passes the checked characteristic information to the calculation resource allocation unit 427. The computing resource allocation unit 427 allocates a system 450 suitable for executing the specific description to the application program. The system 450 executes the assigned application program.

  As described above, according to the present embodiment, the job scheduler 420 in the information processing system 400 passes the application execution binary 411 to the disassembler 430 and receives the instruction list 440 as an execution result of disassembly. In the specific description analysis unit 423, the job scheduler 420 checks whether or not there is a specific description in the instruction list 440. If there is, the job scheduler 420 checks the loop structure depth at that location. When the loop structure depth is larger than the threshold value, the additional information analysis unit 426 determines that the application program corresponding to the instruction list 440 is frequently executed with the specific description. The calculation resource allocation unit 427 allocates a system suitable for the execution to the application program.

  By adopting the above configuration, according to the second embodiment, even when the source program is not given and the characteristic analysis of the program by the compiler is impossible, the same effect as the first embodiment can be obtained. . In other words, computing resources suitable for application execution can be allocated by analyzing application execution binaries, maximizing application execution performance without the need for end users while reducing system construction costs. The effect that it can be obtained.

Third Embodiment FIG. 17 is a diagram showing an outline of a configuration of an information processing apparatus 50 according to a third embodiment of the present invention. As illustrated in FIG. 17, the information processing apparatus 50 includes a specific description analysis unit 51 and a calculation resource allocation unit 52.

  When a specific description that affects the execution characteristics of an application is included in the application, the specific description analysis unit 51 generates additional information regarding the execution frequency of an instruction or function called by the specific description. The calculation resource allocation unit 52 determines a calculation resource suitable for execution of the instruction or function based on the additional information, and allocates the calculation resource to execution of the application.

  By adopting the above configuration, according to the third embodiment, it is possible to reduce the construction cost and to improve the execution performance of the application without taking the trouble of the end user.

  Each part of the compiler and job scheduler shown in FIGS. 2 and 14 is realized by the hardware resources illustrated in FIG. That is, the configuration shown in FIG. 18 includes a CPU 10, a RAM (Random Access Memory) 11, a ROM (Read Only Memory) 12, an external connection interface 13, and a storage medium 14. The CPU 10 controls the overall operation of the compiler and job scheduler by reading various software programs (computer programs) stored in the ROM 12 or the storage medium 14 into the RAM 11 and executing them. That is, in each of the above embodiments, the CPU 10 executes a software program that executes each function (each unit) included in the compiler and the job scheduler while appropriately referring to the ROM 12 or the storage medium 14.

  Further, in each of the above-described embodiments, a case where the functions shown in each block in the compiler and job scheduler shown in FIGS. 2 and 14 are realized by a software program as an example executed by the CPU 10 shown in FIG. did. However, some or all of the functions shown in the blocks shown in FIGS. 2 and 14 may be realized as hardware.

  Further, in the present invention described by taking each embodiment as an example, after supplying a computer program capable of realizing the functions described above to the compiler and job scheduler, the CPU 10 reads the computer program into the RAM 11. Achieved by executing.

  The supplied computer program may be stored in a computer-readable storage device such as a readable / writable memory (temporary storage medium) or a hard disk device. In such a case, the present invention can be understood as being configured by a code representing the computer program or a storage medium storing the computer program.

  The present invention can be applied to, for example, the field of high performance computing.

10 CPU
11 RAM
12 ROM
13 External connection interface 14 Storage medium 100, 400 Information processing system 110, 220, 420 Job scheduler 210 Compiler 221, 421 Application execution reception unit 222 Repetition count threshold setting unit 223, 422 Loop structure depth threshold setting unit 224, 426 Additional information Analysis unit 225, 427 Calculation resource allocation unit 230, 240, 450, 460 System 310, 410 Application execution request 311 411 Application execution binary 312 Iteration number additional information 313, 425 Loop structure depth additional information

Claims (6)

If the application contains a specific description that is a description of at least part of the program code including the loop structure that affects the execution characteristics of the application, additional information on the execution frequency of the instruction or function called by the specific description is generated. Specific description analysis means to
Based on the additional information, a calculation resource suitable for execution of the instruction or function is determined, and a calculation resource allocation unit that allocates the calculation resource to execution of the application is provided.
The specific description analyzing means generates additional information including the number of loop iterations and the depth of the loop structure in the loop structure related to an instruction or function called by the specific description,
When at least one of the number of iterations of the loop and the depth of the loop structure is greater than the respective threshold value, the calculation resource allocation unit makes the execution of the application a calculation resource suitable for execution of the instruction or function. assign
Information processing apparatus. The specific description analyzing means, the application executable or execution binary which the executable program is generated is compiled on the basis of the disassembled result, No placement claim 1 Symbol determines the presence or absence of the specific description Information processing device. The specific description analyzing means extracts a description that calls a collective communication function in the application as a specific description,
The computing resource allocation means, if the repeat count is greater than the threshold value of the collective communication function, the execution of the application, bandwidth allocated to computational resources that are configured with a high communication network than the predetermined value according to claim 1, wherein Item 3. The information processing device according to Item 2 . The specific description analyzing means extracts, as the specific description, a description that calls a read instruction or a write instruction to a file with random access in the application,
The computing resource allocation means, the case read command or the repeat count is greater than the threshold of the write instruction, the execution of the application, Claim assigned to configured computing resources by using a storage device which is suitable for random access one or The information processing apparatus according to claim 2 . By computer
If the application contains a specific description that is a description of at least part of the program code including the loop structure that affects the execution characteristics of the application, additional information on the execution frequency of the instruction or function called by the specific description is generated. And
Based on the additional information, determine a computing resource suitable for execution of the instruction or function, and assign the computing resource to execution of the application,
The additional information includes the number of loop iterations and the depth of the loop structure in the loop structure related to the instruction or function called by the specific description,
An information processing method in which execution of the application is allocated when at least one of the number of repetitions of the loop and the depth of the loop structure is greater than each threshold value of the calculation resource . If the application contains a specific description that is a description of at least part of the program code including the loop structure that affects the execution characteristics of the application, additional information on the execution frequency of the instruction or function called by the specific description is generated. Processing to
Based on the additional information, a computing resource suitable for execution of the instruction or function is determined, and a process of assigning the computing resource to execution of the application is executed by a computer,
The additional information includes the number of loop iterations and the depth of the loop structure in the loop structure related to the instruction or function called by the specific description,
The calculation resource is assigned to execute the application when at least one of the number of loop iterations and the depth of the loop structure is greater than a respective threshold value .

Images