JP2009059316A - Measuring device, measurement program, and measuring method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To solve a problem wherein difference in processing time between object processes of communication processing cannot be measured before performing the communication processing. <P>SOLUTION: This measurement program makes a computer output synchronous waiting time after measurement processing is performed as a process and is synchronous with the other first process and perform the measurement processing for performing the first communication with the other first process. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to a measurement apparatus, a measurement program, and a measurement method.

  Patent Document 1 discloses a technique for measuring a procedure processing time on a processor. Patent Document 2 discloses a technique for measuring a time during which a process is in a waiting state in a semaphore.

JP-A-10-63550 Japanese Patent Laid-Open No. 10-63516

  In each of the above-mentioned patent documents, measurement is performed by causing a computer to execute a measurement process that is executed as a process and synchronizes with another process, and then outputs a waiting time for the synchronization and performs communication with the other process. There is no program. Therefore, the technique of the patent document has the problem that the synchronization waiting time due to the difference in the processing time of the communication processing target process cannot be measured before the communication processing is executed, either individually or in combination. is there.

  An object of the present invention is to provide a measurement apparatus, a measurement program, and a measurement method that solve the above-described problems.

  The measurement program according to an embodiment of the present invention is executed as a process in a computer and synchronizes with the first other process, and then outputs the synchronization waiting time and the first other process. The measurement process for performing the first communication is executed.

  In the measurement method according to an embodiment of the present invention, after the computer synchronizes with the first other process, the computer outputs the synchronization waiting time and performs the first communication with the first other process. As the measurement process to be carried out.

  The present invention can measure the synchronization waiting time due to the difference in processing time generated between the communication processing target processes before executing the communication processing.

  FIG. 1 shows a measuring apparatus 10 according to a first embodiment of the present invention. The measuring apparatus 10 is composed of, for example, a parallel processing computer including a plurality of processors 11. The number of processors 11 may be one.

  The measuring apparatus 10 can execute a program in parallel. A control unit of a program executed in parallel in the measuring apparatus 10 is called a process 12. That is, the measuring apparatus 10 executes a plurality of processes 12 in parallel. Each process 12 is executed by one of the processors 11. The process 12 is controlled by a control program (not shown). Here, the control program is an OS (Operating System) such as UNIX (registered trademark) or Linux.

  The measuring device 10 also includes a clock 13. The clock 13 indicates the current time. The unit of time is preferably about a millimeter row, but is not limited to this.

  The measuring device 10 is connected to a storage device 15. The storage device 15 includes a profile storage area 20, a distributed parallel application 30, and a communication library 40. The distributed parallel application 30 is a program executed by the measurement apparatus 10. The communication library 40 is a set of communication program parts.

  The communication library 40 includes a measurement program 41 and a communication program 42 as program parts. The measuring device 10 is connected to an input / output device 16.

  FIG. 2 shows the configuration of the distributed parallel application 30. The distributed parallel application 30 is usually composed of a plurality of user procedures 31. The user procedure 31 may have a nesting relationship. The user procedure 31 includes a business process program 32 and a measurement program 41.

  The business processing program 32 is a program part created to achieve the business purpose of the distributed parallel application 30. For example, if the distributed parallel application 30 calculates the area of a certain figure, it is a program that performs processing for area calculation.

  On the other hand, the measurement program 41 is a program that is called from the process 12 that executes the user procedure 31 and performs communication for the business processing program 32 and measurement that is the object of the present invention. The measurement program 41 is extracted from the communication library 40 and incorporated in the user procedure 31 when the distributed parallel application 30 is created. For example, the linker incorporates the measurement program 41 when editing the link of the distributed parallel application 30. The dynamic link management program may be incorporated when the measurement program 41 is activated from the user procedure 31.

  When carrying out the measurement according to the present invention, the measurement program 41 is incorporated in the user procedure 31, but the communication program 42 is incorporated during normal operation. The communication program 42 is called from the process 12 that executes the user procedure 31 and performs communication for the business processing program 32. The communication program 42 does not perform the measurement that is the object of the present invention.

  The distributed parallel application 30 is normally distributed to a plurality of processes 12 inside the measuring apparatus 10 and executed in parallel. FIG. 3 illustrates this situation.

  First, the OS activates a user procedure 31 (for example, Procedure A) designated as a starting point of the distributed parallel application 30 as a process 12 (for example, Process 0). The activated Procedure A executes the business processing program 32.

  When the procedure A business processing program 32 reaches a processing portion that can be executed in parallel, it generates a process 12 (for example, Process 1 to n) different from Process 0. One or a plurality of processes 12 may be generated. The processor 11 of the measuring apparatus 10 executes the generated Process 0 and the generated Process 1 to n in parallel.

  A program (for example, Procedure.a1 to an) executed by the generated Process1 to n is specified at the time of generation.

  Process 1 to n execute the business processing program 32 of Procedure.a1 to an, respectively, and then call the measurement program 41 to transmit and receive the calculation result. Process 0 also calls the measurement program 41 when the business processing program 32 ends.

  The measurement program 41 called above synchronizes each process 12 of Process0 to n. That is, the process 12 that has finished executing the business process program 32 waits for the process 12 that has not yet executed the business process program 32. When the last process 12 finishes executing the business processing program 32, the measurement program 41 synchronously outputs the waiting time generated in each process 12 of each Process0 to n to the profile storage area 20, and performs necessary communication.

  Here, for example, the distributed parallel application 30 divides a certain figure into a plurality of sections, executes area integration of each section in parallel, and finally adds them together to obtain the area of the figure. .

  Specifically, it is assumed that the distributed parallel application 30 obtains the area of each division by the business processing program 32 of each process 12 of Process0 to n and adds the calculated values by Process0.

  In this case, in the above-described mutual communication, each process 12 of Process1 to n transmits each calculation result to Precess0. Process 0 receives the calculation result from each process 12 of Process 1 to n.

  The distributed parallel application 30 of the above example can be described in various ways. For example, MPI (Message-Passing Interface) may be used. At this time, the above communication is realized as message passing.

  The distributed parallel application 30 that can use the present invention is not limited to this example. Other distributed parallel applications 30 may be used as long as a part of the processing is distributed in parallel by a plurality of processes 12 and the results of the parallel processing are mutually exchanged.

  After transmitting / receiving the result of message passing, the processes 1 to n usually disappear after receiving the disappearance instruction from the process 0. Process 0 executes the next user procedure 31. Alternatively, Procedure A may be repeatedly executed under the control of the OS and other user programs.

  FIG. 4 shows a synchronization table 50 for synchronizing the processes 12 described above. The synchronization table 50 includes a parent user procedure ID 51. The parent user procedure ID 51 is an identifier of the user procedure 31 that generated the process 12 to be synchronized. In the example of FIG. 3, the identifier of Procedure A is the parent user procedure ID 51.

  The synchronization table 50 further includes synchronization process IDs 52, synchronization states 53, and waiting start times 54 corresponding to the number of processes 12 to be synchronized. In the example of FIG. 3, each of Process 0 to n includes a synchronization process ID 52, a synchronization state 53, and a waiting start time 54. The synchronization state 53 indicates whether the synchronization of the process 12 indicated by the synchronization process ID 52 is completed or not completed.

  The synchronization table 50 is stored in a memory (not shown) of the measurement apparatus 10.

  FIG. 5 is a flowchart of the measurement program 41. As described above, the user procedure 31 calls the measurement program 41. When the user procedure 31 is called, the identifier of the user procedure 31 that generated the process 12 to be synchronized and the identifier of the synchronization target process 12 are given as parameters. These pieces of information are stored in the parent user procedure ID 52 and the synchronization process ID 52 of the synchronization table 50, and are compared with the parent user procedure ID 52 and the synchronization process ID 52 of the synchronization table 50.

  For example, when the parent process 12 that executes the user procedure 31 indicated by the parent user procedure ID 51 generates the child process 12, or after that, the parent process 12 and the child process 12 perform the above information. Shared. Then, on each process 12, the user procedure 31 gives the measurement program 41 as a parameter.

  The measurement program 41 checks whether the synchronization table 50 exists on the memory (S51). This is performed by searching the synchronization table 50 having the parent user procedure ID 51 that matches the identifier of the user procedure 31 given by the parameter at the time of calling. If the synchronization table 50 cannot be found (N in S51), the measurement program 41 creates the synchronization table 50 based on the given parameters (S5B). The program is set to indicate that synchronization is incomplete as the initial value of the synchronization state 53.

  If the synchronization table 50 can be found (Y in S51), the program checks whether the synchronization of all the other processes 12 has been completed (S52). This is performed by checking whether or not the synchronization state 53 of the processes 12 other than its own process 12 indicates synchronization completion.

  If synchronization is not completed in at least one process 12 (N in S52), that is, if there is a process whose synchronization status 53 indicates that synchronization has not been completed for another process 12, the program A waiting start time 54 is set. This is done by reading the clock 13 and storing it at the waiting start time 54.

  Thereafter, the program waits for another process 12 to synchronize (S54). When synchronization is completed in all the other processes 12, the waiting is canceled, and the program performs necessary communication (S55) and ends. That is, the program returns to the calling user procedure 31. The necessary communication is, for example, transmission / reception of calculation results between the processes 12 by message passing.

  When the synchronization of all the other processes 12 is completed (Y in S52), the program calculates the synchronization waiting time (measured value) of each of all the processes 12 (S56). This is performed by obtaining a difference between the current value of the clock 13 and the waiting start time 54 of each process 12. Note that the synchronization waiting time of the process 12 is 0.

  The program outputs the synchronization waiting time of each process 12 thus determined to the profile storage area 20 (S57).

  The synchronization waiting time may be output as a new record each time synchronization is completed. When the user procedure 31 (for example, Procedure A) is repeatedly executed, the synchronization waiting time may be accumulated in each process 12 (for example, Process 0 to n) every time synchronization is completed. Accumulating repeated execution results smooths the effects of accidental factors on synchronization latency.

  The program scans the synchronization table 50 and cancels the waiting of all other processes 12 that are waiting for synchronization (S58).

  Thereafter, the program performs reinitialization of the synchronization table 50 in preparation for re-execution of the same user procedure 31 (S59). That is, the program resets all the synchronization states 53 so as to indicate that synchronization is not completed. Thereafter, the program performs necessary communication (S5A) and ends. The necessary communication is, for example, transmission / reception of calculation results between the processes 12 by message passing.

  FIG. 6 is a flowchart of the communication process of the measurement program 41. This communication process is a process performed in both steps S55 and S5A in FIG. The input parameters for this processing are the identifier of the process 12 to be transmitted / received, the transmission message, or the reception message area. These parameters are given when the user procedure 31 calls the measurement program 41.

  When the input parameter indicates transmission (Y in S61), the program transmits a transmission message to the transmission target process 12 (Precess0) (S62).

  When the input parameter indicates reception (N in S61), the program receives a message from all the processes 12 (Process1 to n) to be received (S63).

  FIG. 6 is also a flowchart of the communication program 42. When the user procedure 31 does not perform the measurement according to the present invention, the communication program 42 is used instead of the measurement program 41.

  FIG. 7 illustrates the measurement values output to the profile storage area 20 by the measurement program 41. The measurement value indicates the synchronization waiting time for each synchronization target process 12 for each user procedure 31 (Procedure A) that generated the process 12 to be synchronized. As described above, the synchronization waiting time may be a measured value for each synchronization, or may be a cumulative value when synchronization is repeated. If the processing time of the business processing program 32 is long, the synchronization waiting time is short. Conversely, if the processing time of the business processing program 32 is short, the synchronization waiting time is long (FIG. 3). Therefore, the processing time difference of the business processing program 32 executed in each process 12 can be understood from the synchronization waiting time.

  Display software (not shown) reads the measurement values from the profile storage area 20 and displays them on the input / output device 16, for example.

  FIG. 8 shows a characteristic part of this embodiment. FIG. 6 shows a measurement program 41 executed as the process 12 in the measurement apparatus 10 which is a computer. After synchronizing with the other process 12, the program causes the waiting time for the synchronization to be output and the measurement process for performing communication with the other process 12 to be executed.

  The effects of this embodiment are as follows.

  The first effect is that the processing time difference caused by the business processing program 32 can be measured with high accuracy. This difference in processing time is caused by a difference in the amount of calculation of the business processing program 32 distributedly executed in each process 12 and a difference in the degree of competition of shared resources used at the time of execution. As a result of this measurement, performance tuning for more efficient parallel execution becomes possible. Specifically, it is possible to increase / decrease the number of shared resources and change the assignment of each business processing program 32 to each process 12 appropriately.

  The reason why the processing time difference caused by the business processing program 32 can be measured with high accuracy is that the synchronization waiting time is measured by synchronizing the processes 12 in the called measurement program 41 before communication after the execution of the program. It is.

  By this measurement, the communication waiting time generated depending on the communication processing included in the processing time of the process 12 is excluded from the measured value, and the processing time difference of the business processing program 32 can be measured with high accuracy.

  The second effect is that the processing time difference caused by the business processing program 32 can be easily measured. The reason is that the means for measuring synchronization and synchronization waiting time is incorporated in the communication library 40 as a program part. This is because the execution format file for measurement of the distributed parallel application 30 can be created only by relinking, and the sole correction and recompilation are not required.

  The second embodiment of the present invention is as follows. First, in this embodiment, when the OS starts the distributed parallel application 30, the clock 13 is read and stored in a start time storage area (not shown).

  Then, the measurement program 41 executed by the last synchronized process 12 reads the clock 13 before communication, and calculates the elapsed time from the time in the start time storage area. Then, the program additionally outputs the calculated value to the profile storage area 20 (added to S59 in FIG. 5).

  The calculated elapsed time indicates the execution time of the process 12 that takes the longest processing time for the user procedure 31 executed first in the distributed parallel application 30.

  In the present embodiment, it is possible to grasp the variation in the execution time of each process 12 and perform more appropriate performance tuning. The reason is that the execution time of each process 12 can be calculated from the execution time and the synchronization waiting time of each process 12.

  FIG. 9 illustrates measurement values output to the profile storage area 20 by the measurement program 41 of the present embodiment. In this example, the execution times of the business processing programs 32 of Precess0, 1, 2 can be grasped as 4.5 seconds, 1.6 seconds, 13.0 seconds, etc., respectively.

It is a figure which shows the whole structure of a measuring apparatus. It is a figure which shows the structure of a distributed parallel application. It is a figure which shows the mode of execution of a distributed parallel application. It is a figure which shows the structure of a synchronous table. It is a flowchart of a measurement program. It is a flowchart of the communication process of a measurement program. It is a figure which shows a measurement result. It is a figure which shows the structure of a measurement program. It is a figure which shows the measurement result of 2nd Embodiment.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 Measuring apparatus 11 Processor 12 Process 13 Clock 15 Storage device 16 Input / output device 20 Profile storage area 30 Distributed parallel application 31 User procedure 32 Business processing program 40 Communication library 41 Measurement program 42 Communication program 50 Synchronization table 51 Parent user procedure ID
52 Synchronization process ID
53 Synchronization state 54 Waiting start time

Claims (10)

Running as a process on a computer,
A measurement program for executing, after synchronization with the first other process, output of the waiting time for the synchronization and measurement processing for performing first communication with the first other process. In the computer,
2. The measurement program according to claim 1, wherein the measurement processing is performed in synchronization with both the first other process and the second other process that performs second communication with the first process after the synchronization. . The measurement program according to claim 1, wherein the measurement program is a program component called to perform the first communication after a business process from a user procedure executed by the computer. The first communication is performed as message passing for transmitting a calculation result by the business process of the user procedure to the computer that executes the user procedure belonging to the distributed parallel application as the process and the first other process, respectively. The measurement program according to claim 3 to be executed. The measurement program according to claim 4, wherein the computer accumulates the waiting time for each user procedure and each process. A measuring apparatus that executes the measurement program according to claim 1. Computer
A measurement method comprising a measurement step of executing, as a process, an output of a waiting time for the synchronization and a first communication with the first other process after synchronizing with the first other process. The computer is
The measurement method according to claim 7, further comprising the measurement step of synchronizing the first other process and the second other process that performs second communication with the first process after the synchronization. The measurement method according to claim 7, wherein the measurement process is performed by the computer calling a program component for performing the first communication after a business process from a user procedure executed by the computer. The computer that executes the user procedure belonging to the distributed parallel application as the process and the first other process respectively performs the first communication as message passing for transmitting a calculation result by the business process of the user procedure. The measurement method according to claim 9 to be executed.

Images