JP5526914B2 - Analysis device, analysis method, and analysis program - Google Patents

Description

  The present invention relates to an analysis apparatus, an analysis method, and an analysis program.

  In recent years, a distributed memory multiprocessor, which is a system in which a plurality of processors are connected using an interconnect device (Interconnect), has been developed. The distributed memory multiprocessor assigns a process to each processor and communicates with each other using a communication library such as MPI (The Message Passing Interface Standard) to perform the entire distributed parallel processing.

  In such distributed parallel processing, performance is improved by appropriately allocating a load to each process. For this reason, an application program (distributed parallel processing program) operating on a distributed memory multiprocessor parallel computer has load balance uniformity as an item of performance characteristic index. It is expected that the uniformity of the load balance will be observed easily and precisely, and the performance characteristics of the program will be analyzed accurately.

  Each process of the distributed parallel processing program performs calculation processing and then performs communication processing with each other. In communication processing, communication waiting occurs when calculation processing of the communication partner is not completed. Therefore, in order to measure the uniformity of the load balance, it is effective to grasp the state of occurrence of waiting for communication.

  As a program performance analysis method, a trace log method and a sampling method are known. The trace log method outputs the occurrence of an event to a log together with accompanying information such as time each time various events occur during program execution. This is a technique for analyzing various performances by analyzing logs after running the program. The trace log method is advantageous in that the time and number of occurrences of various events can be accurately recorded. However, the amount of logs to be output is enormous and the program execution over a long time cannot be analyzed. In addition, the log output itself may affect the performance characteristics.

  In the sampling method, when the program is executed, the running state of the program is confirmed and recorded at regular time intervals. Then, after running the program, various performance analyzes are performed by statistically analyzing the records. The sampling method can suppress the amount of recording as compared with the trace log method, and can analyze program execution over a long period of time. In addition, since there is little disturbance, there is little influence on the performance characteristics.

JP-A-5-250339 JP-A-6-59944 JP 2004-341750 A JP-A-6-83608 JP 2007-207173 A

  However, in the sampling method, it is impossible to accurately know the time and number of occurrences of various events. Therefore, in the performance analysis of the program by the conventional sampling method, it is difficult to identify the cause of the deterioration of the communication waiting state.

  Thus, in one aspect, the present invention aims to provide an improved sampling scheme.

  In one proposal, the analysis apparatus, the analysis method, and the analysis program acquire sampling data obtained by sampling the waiting time of the own process at the time of communication with other processes for each of a plurality of processes that are distributed and processed in parallel. Based on the sampling data, for each of the plurality of processes, the total waiting time of the own process between the other process and the waiting time of the other process during the communication between the other process and the own process. The total time is obtained, and the result is analyzed to evaluate the distribution of processing to a plurality of processes.

  According to the present invention, an improved sampling method can be provided.

FIG. 1 is a configuration diagram of a distributed memory multiprocessor parallel computer according to an embodiment. FIG. 2 is an explanatory diagram of communication between processes. FIG. 3 is an explanatory diagram of the distributed process. FIG. 4 is an explanatory diagram of a comparative example of sampling data. FIG. 5 is an explanatory diagram of a processing state corresponding to the sampling data D0. FIG. 6 is a flowchart for explaining the processing operation in the processors 11 to 14. FIG. 7 is a flowchart for explaining the processing operation of the analysis device 30. FIG. 8 is an explanatory diagram of a specific example of the total data D2. FIG. 9 is an explanatory diagram for the analysis by the analysis unit 33. FIG. 10 is a flowchart for explaining the processing operation in the processors 11 to 14 when the evaluation for each function is performed. FIG. 11 is a flowchart illustrating details of sampling (S600). FIG. 12 is a flowchart for explaining the processing operation of the analysis apparatus 30 when evaluating a function. FIG. 13 is an explanatory diagram of specific examples of the waiting situation matrix, the function matrix, and the aggregation matrix. FIG. 14 is an explanatory diagram for the evaluation from the aggregation matrix and the creation of a correction guideline.

  Hereinafter, an analysis device, an analysis method, and an analysis program disclosed in the present application will be described in detail based on the drawings. The present invention is not limited to the following specific examples.

[System configuration]
FIG. 1 is a configuration diagram of a distributed memory multiprocessor parallel computer according to an embodiment. In the example shown in FIG. 1, the processors 11 to 14 are connected to an interconnect device (interconnect) 21, and the processors 11 to 14 can communicate with each other via the interconnect device 21.

  Each of the processors 11 to 14 executes one or a plurality of processes. In the example shown in FIG. 1, the processor 11 executes the application process Pa1, and the processor 12 executes the application process Pa2. Similarly, the processor 13 executes the application process Pa3, and the processor 14 executes the application process Pa4.

  Application processes Pa1 to Pa4 are processes in which application program processes are distributed and allocated, and are processed in parallel by the processors 11 to 14. The processors 11 to 14 perform calculation processing on the processes allocated to themselves, and then communicate via the interconnect device 21 to synchronize the processing.

  FIG. 2 is an explanatory diagram of communication between processes. In the example shown in FIG. 2, the transmission process MPI_Send for the process Pa2 occurs in the process Pa1. The process Pa2 receives the communication from the process Pa1 by the reception process MPI_Recv. Thereafter, similarly, in the process Pa2, a transmission process MPI_Send for the process Pa1 occurs, and the process Pa1 receives by the reception process MPI_Recv. Further, the process Pa1 and the process Pa2 perform synchronization by MPI_Barrier.

  FIG. 3 is an explanatory diagram of the distributed process. In the example shown in FIG. 3, the process Pa1 calculates the process assigned to itself over 80 ms, and then synchronizes with the communication process of 20 ms. On the other hand, the process Pa2 ends the calculation process assigned to itself in 40 ms, and a communication wait state for waiting for the end of the calculation process of the process Pa1 occurs for 40 ms.

  Similarly, the process Pa3 ends the calculation process assigned to itself in 20 ms, and a communication waiting state for waiting for the end of the calculation process of the process Pa1 occurs for 60 ms. In addition, the process Pa4 ends the calculation process assigned to itself in 60 ms, and a communication waiting state for waiting for the end of the calculation process of the process Pa1 occurs for 20 ms.

  In the example shown in FIG. 3, since the calculation process of the process Pa1 is heavy, the processes Pa2 to Pa4 are in a waiting state. For this reason, if the process assigned to the process Pa1 is assigned to another process, the load balance can be improved.

  FIG. 4 is an explanatory diagram of a comparative example of sampling data. As shown in FIG. 4, the sampling data D0 acquired by sampling from the distributed processes Pa1 to Pa4 acquires the calculation processing cost, communication waiting cost, and communication processing cost of each process.

  From this sampling data D0, it is possible to know how much time each process has spent for calculation, waiting for communication, and communication processing. In the example shown in FIG. 4, since the calculation processing cost of the process Pa3 is higher than that of other processes, it can be considered that the process Pa3 disturbs the communication waiting state of each process. However, since the process Pa3 also has a communication waiting cost and is waiting for the completion of calculation of another process, there is a possibility that the process is delayed by another process.

  FIG. 5 is an explanatory diagram of a processing state corresponding to the sampling data D0. In the example shown in FIG. 5, one square corresponds to the cost 10 of the sampling data D0. As shown in FIG. 5, until the synchronization timing t <b> 1, the process Pa <b> 4 executes a calculation process with a cost of 10 and waits for a communication with a cost of 50. Similarly, until the synchronization timing t1, the process Pa1 executes the calculation process of the cost 50 and waits for the communication of the cost 10. In addition, until the synchronization timing t1, the process Pa2 executes the calculation process of the cost 40, waits for the communication of the cost 20, and the process Pa3 executes the calculation process of the cost 60.

  During the period from the synchronization timing t1 to the synchronization timing t2, the processes Pa1 to Pa3 execute the calculation process of the cost 10 and wait for the communication of the cost 30. And the process Pa4 is performing the calculation process of the cost 40 from the synchronous timing t1 to the synchronous timing t2.

  That is, in the example shown in FIG. 5, although the calculation cost of the process Pa3 is high and the overall communication waiting state is deteriorated, it is from the synchronization timing t1 to the synchronization timing t2 that most adversely affects the communication waiting state. Process Pa4. As described above, when there is a process in which the load varies, that is, there is a fluctuation in the amount of calculation processing, and there is a process that waits for another process, data of the calculation processing cost, the communication waiting cost, and the communication processing cost that the sampling data D0 has The communication waiting status cannot be accurately evaluated.

  Therefore, the analysis device 30 shown in FIG. 1 acquires sampling data obtained by sampling the waiting time of the own process at the time of communication with other processes for each of a plurality of processes that are distributed and processed in parallel. Based on the acquired sampling data, the analysis device 30 determines, for each of the plurality of processes, the total waiting time of the own process between the other processes and the communication between the other processes and the own process. Aggregation for obtaining the waiting time of other processes, that is, the total waiting time is performed, and the result of the aggregation is analyzed to evaluate the distribution state of processing for a plurality of processes.

  Therefore, as illustrated in FIG. 1, the analysis device 30 is connected to the interconnect device 21 and communicates with the processors 11 to 14 via the interconnect device 21. In FIG. 1, the configuration in which the analysis device 30 is implemented as one device is shown, but for example, any of the processors 11 to 14 may execute a program that realizes the function of the analysis device 30. A processor for analysis may be connected separately.

  The processor 11 executes the sampling thread Ps1 in addition to the application process Pa1. Similarly, the processor 12 executes the sampling thread Ps2 in addition to the application process Pa2. The processor 13 executes the sampling thread Ps3 in addition to the application process Pa3, and the processor 14 executes the sampling thread Ps4 in addition to the application process Pa4.

  The sampling thread Ps1 confirms and records the running status of the application process Pa1 at regular time intervals. At this time, the sampling thread Ps1 creates sampling data D1a by accumulating the time cost from when the application process Pa1 enters the communication waiting state to when the communication waiting state is released for each communication partner.

  In the example shown in FIG. 1, the sampling data D1a indicates that the communication waiting cost for the destination process Pa2 is 0, the communication waiting cost for the destination process Pa3 is 10, and the communication waiting cost for the destination process Pa4 is 30. Show.

  Similarly, the sampling thread Ps2 confirms the running status of the application process Pa2 at regular time intervals, and calculates the time cost from entering the communication wait state until the communication wait state is released for each communication partner. To create sampling data D1b.

  In addition, the sampling thread Ps3 checks the running status of the application process Pa3 at regular time intervals, and calculates the time cost from entering the communication waiting state until the communication waiting state is canceled for each communication partner. Integration is performed to create sampling data D1c.

  The sampling thread Ps4 checks the running status of the application process Pa4 at regular time intervals, and calculates the time cost from entering the communication waiting state until the communication waiting state is released for each communication partner. Integration is performed to create sampling data D1d.

  The analysis device 30 includes a data acquisition unit 31, a data totaling unit 32, and an analysis unit 33. The data acquisition unit 31 acquires sampling data D1a, D1b, D1c, and D1d from the sampling threads Ps1 to Ps4.

  The data totaling unit 32 creates total data D2 by totaling the sampling data D1a, D1b, D1c, D1d. From each of the aggregated data D2, for each of the processes Pa1 to Pa4, it is possible to obtain the standby time for waiting for communication with other processes and the standby time for waiting for other processes.

  The analysis unit 33 is a processing unit that analyzes the aggregated data D2, evaluates a distribution state of processes for the processes Pa1 to Pa4, and outputs an evaluation result.

[Processing operation]
FIG. 6 is a flowchart for explaining the processing operation in the processors 11 to 14. In the following description, the processor 11 is described as an example, but the same applies to the processors 12 to 14.

  First, the processor 11 executes the process Pa1 assigned from the application program. The process Pa1 generates a sampling thread Ps1 at the start of processing (S101). The processor 11 executes calculation processing in the process Pa1 (S102) and also performs sampling by the sampling thread Ps1 (S201).

  After completing the calculation process (S102), the processor 11 waits for the erasure of the sampling thread Ps1 (S103), outputs the sampling data D1a (S104), and ends the process.

  FIG. 7 is a flowchart for explaining the processing operation of the analysis device 30. As shown in FIG. 7, first, the data acquisition unit 31 acquires the sampling data D1a, D1b, D1c, D1d (S301), and the data totaling unit 32 totals the sampling data D1a, D1b, D1c, D1d ( S302). And the analysis part 33 analyzes the total data D2, outputs an evaluation result (S303), and complete | finishes a process.

[Specific examples of data and processing]
FIG. 8 is an explanatory diagram of a specific example of the total data D2. The aggregated data D2 shown in FIG. 8 indicates the communication waiting cost for the receiving process and the transmitting process. Specifically, when the process on the receiving side, that is, the sampled process is the process Pa1, the communication waiting cost for the destination process Pa2 is 0, the communication waiting cost for the destination process Pa3 is 10, and the communication waiting cost for the destination process Pa4 is 30.

  Further, for the process Pa2 on the receiving side, the communication waiting cost for the transmission destination process Pa1 is 10, the communication waiting cost for the transmission destination process Pa3 is 20, and the communication waiting cost for the transmission destination process Pa4 is 30.

  Similarly, for the process Pa3 on the receiving side, the communication waiting cost for the destination process Pa1 is 0, the communication waiting cost for the destination process Pa2 is 0, and the communication waiting cost for the destination process Pa4 is 30.

  For the process Pa4 on the receiving side, the communication waiting cost for the destination process Pa1 is 40, the communication waiting cost for the destination process Pa2 is 30, and the communication waiting cost for the destination process Pa3 is 50.

  The total of the rows of the aggregated data D2 is the total of the waiting time that the process waited for communication with the other process, and the total of the columns of the aggregated data D2 is the sum of the waiting time that the other process waited for Become.

  That is, in the example shown in FIG. 8, the total waiting time of process Pa1 is 40, the total waiting time of process Pa2 is 40, the total waiting time of process Pa3 is 30, and the total waiting time of process Pa4 is 120. is there. The total waiting time of the process Pa1 is 50, the total waiting time of the process Pa2 is 30, the total waiting time of the process Pa3 is 80, and the total waiting time of the process Pa4 is 90.

  The total value of the rows, that is, the total value of the cost in the receiving side direction (lateral direction) is the total of the total time that the own process waited for communication, and a large value indicates that the calculation processing was light. In the example of FIG. 8, it is shown that the waiting time of the process Pa4 is 120, and the load of the process Pa4 is lower than that of other processes.

  The total value of the column, that is, the total cost in the transmission direction (vertical direction) is the total time that the own process waited for other processes to wait for communication, and a large value indicates that processing was heavy. ing. In the example of FIG. 8, the total waiting time of the process Pa4 is 90, which indicates that the process Pa4 is in a higher load than other processes. The process having the next highest load is process Pa3.

  Comprehensively determining the waiting time and the waiting time, the process Pa4 has a large blur between the case where the load is low and the case where the load is high compared to other processes, and there is a large room for improvement in the communication waiting state. It can be seen that Pa3 has room for improvement in the waiting state for communication because the load is high.

  FIG. 9 is an explanatory diagram for the analysis by the analysis unit 33. As shown in the evaluation table D3 of FIG. 9, when both the cost of the standby time and the cost of the standby time are low (L_L), the calculation processing of the process is appropriate and the communication waiting time is small. Also, it does not make the other party wait. Therefore, no improvement is necessary.

  When the cost of the waiting time is low and the cost of the waiting time is high (L_H), the process has a lot of calculation processing and the communication waiting time is small. Also, it often makes the other party wait. For this reason, improvement is necessary and it is desirable to reduce work.

  When the cost of the standby time is high and the cost of the standby time is low (H_L), the process has a small amount of calculation processing and a large waiting time for communication. And it rarely makes the other party wait. In other words, since there are many times when no calculation processing is performed, improvement is necessary and it is desirable to increase work.

  When the cost of the waiting time and the cost of the waiting time are both high (H_H), the process often waits for communication and often makes the other party wait. Therefore, improvement is necessary, and it is desirable to reduce the fluctuation of the work amount.

  As described above, the analysis apparatus 30 can accurately evaluate the load balance, that is, the uniformity of the processing distribution state, in units of processes from the process standby time and the standby time, and output an improvement guideline.

[Function evaluation]
The analysis device 30 can also evaluate the function by integrating the waiting time for each user function at the time of sampling. FIG. 10 is a flowchart for explaining the processing operation in the processors 11 to 14 when the evaluation for each function is performed. In the following description, the processor 11 is described as an example, but the same applies to the processors 12 to 14.

  When the communication waiting process is started, the processor 11 records the time Tst when the communication is started (S401), and performs the receiving process (S402). Then, when reception is completed and communication waiting is canceled, time Tend is recorded (S403), and a waiting time that is the difference between time Tend and time Tst is added to the waiting status matrix for each transmission source (S404).

  Further, when executing the process Pa1 assigned by the application program, the processor 11 generates a sampling thread Ps1 from the process Pa1 (S501). The processor 11 executes a calculation process in the process Pa1 (S502) and performs sampling by the sampling thread Ps1 (S600).

  After the calculation process (S502) of the process Pa1 ends, the processor 11 waits for the sampling thread Ps1 to be deleted (S503), outputs the sampling data D1a (S504), and ends the process.

  FIG. 11 is a flowchart illustrating details of sampling (S600). As shown in FIG. 11, the processor 11 refers to the wait status matrix every time sampling is performed from the process Pa1 (S601) (S602). Then, the sampling is added to the function-by-function matrix of the user function (S603), and the waiting situation matrix is cleared to zero (S604). For example, call-graph profiling may be used for sampling by function.

  FIG. 12 is a flowchart for explaining the processing operation of the analysis apparatus 30 when evaluating a function. As shown in FIG. 12, first, the data acquisition unit 31 acquires sampling data (S701), and the data totaling unit 32 totals (S702). The analysis unit 33 analyzes the matrix for each function for each user function, finds the total time required for reception and the total time for reception, creates a total matrix (S703), and outputs an analysis report ( S704).

[Specific examples of data and processing]
FIG. 13 is an explanatory diagram of specific examples of the waiting situation matrix, the function matrix, and the aggregation matrix. The wait status matrix D4 illustrated in FIG. 13 is data indicating the wait status of the process Pa4. The matrix D4 waits for communication at cost 40 in communication with the process Pa1 on the transmission side, waits for communication at cost 30 in communication with the process Pa2 on the transmission side, and costs 50 in communication with process Pa3 on the transmission side. Indicates that communication has been waited.

  The function-by-function matrix D5 illustrated in FIG. 13 is data indicating a waiting state for each function of the process Pa4. For each function c1, the function matrix D5 waits for communication at cost 300 by communication with the process Pa1 on the transmission side, waits for communication at cost 10 by communication with the process Pa2 on the transmission side, and It shows that the communication waiting of the cost 100 was performed by this communication.

  Similarly, the function-by-function matrix D5 is a communication of the function c2 at a cost of 100 for communication with the process Pa1 on the transmission side, a cost of 300 for communication with the process Pa2 on the transmission side, and a cost of 200 for communication with the process Pa3 on the transmission side. Indicates waiting.

  Further, the function-specific matrix D5 indicates that the function c3 is waiting for communication with the cost 10 for communication with the process Pa1 on the transmission side, the cost 30 for communication with the process Pa2 on the transmission side, and the cost 300 for communication with the process Pa3 on the transmission side. It has been shown that.

  Further, the function matrix D5 indicates that the function c4 is waiting for communication with the cost 400 for communication with the process Pa1 on the transmission side, the cost 300 for communication with the process Pa2 on the transmission side, and the cost 500 for communication with the process Pa3 on the transmission side. It has been shown that.

  The analysis device 30 acquires the matrix for each function from each process, and totals the functions for each function to create a total matrix. The aggregation matrix D6 illustrated in FIG. 13 exemplifies the aggregation result for the function c3.

  In the aggregation matrix D6, for the process Pa1 on the receiving side, the communication waiting cost for the transmission destination process Pa2 is 0, the communication waiting cost for the transmission destination process Pa3 is 100, and the communication waiting cost for the transmission destination process Pa4 is 300.

  For the receiving process Pa2, the communication waiting cost for the transmission destination process Pa1 is 100, the communication waiting cost for the transmission destination process Pa3 is 200, and the communication waiting cost for the transmission destination process Pa4 is 300.

  Similarly, for the process Pa3 on the receiving side, the communication waiting cost for the destination process Pa1 is 0, the communication waiting cost for the destination process Pa2 is 0, and the communication waiting cost for the destination process Pa4 is 300.

  For the receiving process Pa4, the communication waiting cost for the destination process Pa1 is 10, the communication waiting cost for the destination process Pa2 is 30, and the communication waiting cost for the destination process Pa3 is 300.

  Therefore, for the function c3, the total waiting time of the process Pa1, that is, the total row is 400, the total waiting time of the process Pa2 is 600, the total waiting time of the process Pa3 is 300, and the total waiting time of the process Pa4 is 340. Become. In addition, the total waiting time of the process Pa1, that is, the column total is 110, the total waiting time of the process Pa2 is 30, the total waiting time of the process Pa3 is 600, and the total waiting time of the process Pa4 is 900. It becomes.

  FIG. 14 is an explanatory diagram for the evaluation from the aggregation matrix and the creation of a correction guideline. A process evaluation D7 for each function can be obtained from the standby time and the standby time of the aggregation matrix D6. The process evaluation D7 is L_L when the total waiting time of the processes is lower than the average for each function and the total waiting time is lower than the average, the total waiting time is lower than the average, and the total waiting time is L_H when higher than average, total waiting time is higher than average, H_L when total waiting time is lower than average, total waiting time is higher than average, total waiting time is higher than average If H is also high, the value of H_H is taken.

  In the example of the function c3, the average value of the total waiting time is 410 in (400 + 600 + 300 + 340) / 4. The process Pa1 has a total waiting time of 400 and the total waiting time is 110, so L_L, and the process Pa2 has a total waiting time of 600 and the total waiting time is 30, so it becomes H_L. Further, since the total waiting time is 300 and the total waiting time is 600 for the process Pa3, L_H, and the total processing time for the process Pa4 is 340 and the total waiting time is 900. Become.

  Similarly, the process evaluation D7 takes the value of H_H in the processes Pa1 to Pa4 for the function c1. The process evaluation D7 takes a value of H_L for the process Pa1, L_L for the process Pa2, H_L for the process Pa3, and L_L for the process Pa4 for the function c2. The process evaluation D7 takes a value of H_H for the process Pa1 and the process Pa2 and L_L for the process Pa3 and the process Pa4 for the function c4.

  As the entire process, the process Pa1 and the process Pa2 take the value H_L, and the process Pa3 and the process Pa4 take the value H_H.

  From this process evaluation D7, the analysis apparatus 30 creates a process correction guideline D8. No correction is required for L_L, a decrease in work is desirable for L_H, an increase in work is desirable for H_L, and a decrease in work blur is desirable for H_H.

  For this reason, the process correction guideline D8 presents a reduction in shake relative to the processes Pa1 to Pa4 for the function c1, and presents an increase in work for the processes Pa1 and Pa3 for the function c2. Further, the process correction guideline D8 presents an increase in work for the process Pa2 and a decrease in work for the processes Pa3 and Pa4 for the function c3, and a decrease in blurring for the processes Pa1 and Pa2 for the function c4.

  Then, the process correction guideline D8 presents an increase in work for the entire process Pa1 and process Pa2, and presents a reduction in shake for the entire process Pa3 and process Pa4.

  As described above, the analysis apparatus 30 according to the present embodiment, for each of a plurality of processes subjected to distributed parallel processing from the processors 11 to 14, waits for the own process when communicating with other processes. Get the sampling data sampled. The analysis apparatus 30 that has acquired the sampling data has, for each of the plurality of processes, the total waiting time of the own process with the other process and the other process at the time of communication between the other process and the own process. The total waiting time is obtained, and the result is analyzed to evaluate the distribution state of processing for a plurality of processes. For this reason, the analysis device 30 can accurately evaluate the load balance and contribute to improving the performance of the distributed parallel processing program.

  In addition, a present Example is an example to the last, A structure and operation | movement can be changed suitably and can be implemented. For example, in this embodiment, a system having four processors has been described as an example, but the present invention can be applied to a system having an arbitrary number of processors.

11-14 Processor 21 Interconnect device 30 Analysis device 31 Data acquisition unit 32 Data totaling unit 33 Analysis unit Pa1-Pa4 Process Ps1-Ps4 Sampling thread D1a, D1b, D1c, D1d Sampling data D2 Total data D3 Evaluation table D4 Waiting state matrix D5 Matrix for each function D6 Aggregation matrix D7 Process evaluation D8 Process modification guidelines

Claims (5)

For each of a plurality of processes that are distributed in parallel processing, an acquisition unit that acquires sampling data obtained by sampling the waiting time of the own process at the time of communication with another process;
Based on the sampling data, for each of the plurality of processes, the total waiting time of the own process between the other process and the waiting time of the other process during communication between the other process and the own process. A totaling section that calculates the total time,
An analysis apparatus comprising: an analysis unit that analyzes a result of aggregation by the aggregation unit and evaluates a distribution state of processing for the plurality of processes.   The aggregating unit calculates the sum of the waiting time of the own process with the other process for each function being executed at the time of sampling, and the other process in the communication between the other process and the own process. The analysis apparatus according to claim 1, wherein a total waiting time is obtained, and the analysis unit evaluates a process distribution state for each function. The analysis unit
The total waiting time of the own process with the other process is smaller than that of the other process, and the total waiting time of the other process at the time of communication between the other process and the own process is Evaluate that process allocation should be reduced for processes that are large relative to the process,
The total waiting time of the own process with the other process is larger than that of the other process, and the total waiting time of the other process at the time of communication between the other process and the own process is Assess that the process allocation should be increased for small processes relative to the process,
The total waiting time of the own process with the other process is larger than that of the other process, and the total waiting time of the other process at the time of communication between the other process and the own process is The analysis apparatus according to claim 1, wherein it is evaluated that process variation should be suppressed for a process that is larger than the process. An analysis method executed by an analysis device,
The acquisition unit of the analysis device acquires sampling data obtained by sampling the waiting time of the own process at the time of communication with other processes for each of a plurality of processes to be distributed and processed in parallel;
Based on the sampling data , the totaling unit of the analysis device, for each of the plurality of processes, the total waiting time of the own process between the other process and the communication between the other process and the own process. An aggregation step to find the total waiting time of other processes during
An analysis method comprising: an analysis unit of the analysis device analyzing the aggregation result of the aggregation step and evaluating a distribution state of processing for the plurality of processes. Computer
An acquisition unit that acquires sampling data obtained by sampling the waiting time of the own process at the time of communication with another process for each of a plurality of processes that are distributed and processed in parallel .
Based on the sampling data, for each of the plurality of processes, the total waiting time of the own process between the other process and the waiting time of the other process during communication between the other process and the own process. A totaling section that calculates the total time ,
An evaluation unit that analyzes a result of aggregation by the aggregation unit and evaluates a distribution state of processing for the plurality of processes.
Analysis program to function as

Images