JPH1027167A - Method for distributing load of parallel computer - Google Patents

Abstract

PROBLEM TO BE SOLVED: To shorten the lapse of time in the whole parallel program by allowing constitutional programs themselves to execute load distribution considering load to be applied to respective computers at the time of allocating respective constitutional programs in the parallel program constituted of plural constitutional programs to respective computers in the parallel computer constituted of plural computers. SOLUTION: The history sampling programs 4 (4-1 to 4-4) of respective computers 1 (1-1 to 1-4) sample the resource using rate data of respective constitutional programs 2 (2-1 to 2-4) and store the sampled data in the resource using history 12. A schedular 6 finds out an average resource using rate from the history 12, stores the found value in average resource using data 13 and allocates constitutional programs successively from the constitutional program having the largest using rate of a noticed resource to the computers 1 successively from the computer having the smallest using rate of the noticed resource while referring to the using rates of respective computers 1 stored in current load data 15 as to the noticed resource set up in user setting information 14.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a parallel computer composed of a plurality of computers, and more particularly to a case where each component program of a parallel program composed of a plurality of component programs is assigned to each computer and executed in parallel. And a load balancing method.

[0002]

2. Description of the Related Art In recent years, distributed memory type parallel computers, which are composed of a plurality of computers and each having a dedicated storage device and a processor connected to each other via a high-speed network, have begun to spread. Each constituent program of the parallel program composed of a plurality of constituent programs is assigned to each computer and executed in parallel. When a configuration program communicates with another configuration program via a network, an MPI (Message Passing)
Interface), PVM (Parallel)
Message communication between programs is performed via a parallel communication library such as Virtual Machine.
These parallel communication libraries provide an application interface that allows a configuration program to easily communicate with other configuration programs.

When assigning each constituent program to a computer, the scheduling computer collects load information of each computer and assigns each constituent program in ascending order of the load.

[0004] Note that related to this kind of technology are Japanese Patent Application Laid-Open Nos. 7-234847 and 5-735.
No. 15 publication.

[0005]

According to the above prior art, the load applied to the computer by each constituent program itself has not been considered. For this reason, a configuration program with a long elapsed time during execution among a plurality of configuration programs may be assigned to a computer with a heavy load, and as a result, the elapsed time of such a configuration program is extended and becomes a bottleneck, There is a problem that the elapsed time of the entire parallel program increases. Particularly when the parallel program is a program that is repeatedly executed many times, it is desired to execute the program with a stable elapsed time.

An object of the present invention is to reduce the elapsed time of the entire parallel program by performing load distribution in consideration of the load applied to the computer by the constituent program itself.

[0007]

SUMMARY OF THE INVENTION According to the present invention, a resource use rate at the time of executing each constituent program on each computer is obtained,
A parallel computer that obtains the resource usage rate of each computer as a whole without executing the configuration program, and allocates the configuration programs to the computers in order from the configuration program with the highest resource usage rate to the computer with the lowest resource usage rate. Is characterized.

According to the present invention, among resources such as CPU and I / O, attention is paid to resources that are likely to become bottlenecks from the characteristics of the constituent programs, and the constituent programs are arranged in order from the constituent program having the highest utilization rate of the noted resources. Are assigned to the computers in such a way that the use ratio of the resource of interest is small, so that the elapsed time of the constituent programs that are likely to be a bottleneck can be prevented from increasing, and the elapsed time of the entire parallel program can be reduced. .

[0009]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 is a configuration diagram of a distributed memory type parallel computer. Each of the computers 1-1 and 1 that constitute the parallel computer
2, 1-3, 1-4, and 1-5 are connected to a common communication path 20, and each have an arithmetic processing unit having the same processing capability and a dedicated storage device. The parallel program 11 is 1
Each of the constituent programs is an application program of a user which is assigned to any one of the computers 1 and executed in parallel with the other constituent programs. The configuration program 2 is a configuration program that is allocated to the computers 1-1 to 1-4, loaded into a dedicated storage device, and executed.
The communication program 3 is a program that causes the computer 1 to transmit and receive information to and from another computer 1, and is preferably a communication control program including a parallel communication library in terms of providing an easily usable application interface. The history collection program 4 is a program for collecting history data of the resource usage rate and the elapsed time during execution of the configuration program 2. The load data collection unit 5 is hardware and a program that collects resource utilization data of the entire computer 1. The configuration of the computers 1-2, 1-3 and 1-4 is the same as that of the computer 1-1, and when the configuration program 2, communication program 3, history collection program 4, and load data collection unit 5 of each computer are distinguished. , Computer identifiers 1 to 5 are added.

The resource use history 12 is history data stored in a storage device connected to the communication path 20 and shared by all computers, and collected by the history collection program 4. The average resource usage data 13 stores data on the average resource usage rate and the elapsed time of each component program 2. The user setting information 14 is a parameter used at the time of scheduling for allocating each constituent program 2 to the computer 1. The current load data 15 is resource utilization data of each computer 1. The predicted load data 16 is load data predicted when the resource usage of the configuration program 2 allocated to the current load of the computer 1 is added. The parallel program 11, the average resource usage data 13, the user setting information 14, the current load data 15, and the predicted load data 16
It is stored on a dedicated storage device of the computer 1-5.

The scheduler 6 is a program stored and executed in the dedicated storage device of the computer 1-5. The scheduler 6 reads the resource use history 12, calculates the average resource use data of each constituent program 2, and calculates the average resource use data. Stored in the data 13, read the user setting information 14, determine the assignment priority of each constituent program based on the parameters,
The current load data of each computer 1 is collected via the load data collection unit 5 and stored in the current load data 15.
The load of each computer 1 when the configuration program 2 is added is calculated from the current load data 15 and the predicted load data 1
6 and the configuration program 2 to be assigned to each computer 1 is determined. The communication program 7 includes the computer 1
-5 is a program for transmitting and receiving information to and from another computer 1. Note that the computer 1-5 may be a computer dedicated to scheduling, or another computer 1 after scheduling.
May be a computer that executes a configuration program in parallel with.

FIG. 2 is a diagram showing a data format of the resource use history 12. The resource use history 12 records a resource use rate and an elapsed time when each component program 2 is executed at least once. The item number is a number indicating the number of executions. The elapsed time is a time from the start of execution of the configuration program 2 to the end thereof. In the case of transaction processing, the elapsed time is the time from when a transaction is input to the configuration program 2 until when the resulting data is output, and corresponds to a turnaround time. CPU
When the configuration program 2 is executed independently,
This is the ratio of the usage time of the arithmetic processing device to the elapsed time of the configuration program 2. The I / O usage rate is a ratio of the usage time of the input / output device such as the external storage device and the transmission of the communication path 20 to the elapsed time of the configuration program 2.

First, the configuration programs 2-1, 2-2, 2
-3 and 2-4 are computers 1-1, 1-2, 1-
3 and 1-4 and executed in parallel. The configuration program 2 proceeds as necessary while communicating with another configuration program 2 via the communication program 3 and the communication path 20. Each history collection program 4
Collects data on the CPU usage rate, the I / O usage rate, and the elapsed time while the configuration program 2 is being executed, and transmits the resource usage history 12 via the communication program 3 and the communication path 20.
To be stored. In this way, the history collection program 4
Each time the configuration program 2 is executed, the data of the resource usage rate and the elapsed time is collected and stored in the resource usage history 12. Even in the case of the same program, the resource use rate and the elapsed time may differ depending on the input data. In such a case, the configuration program 2 is executed a plurality of times and each history data is recorded in the resource use history 12. Is desirable. The computer 1 may run another application program (not shown) other than the configuration program 2 at the same time as the configuration program 2. However, while the history collection program 4 is collecting the history data of the configuration program 2, the other programs are not run. Is desirable. It is also assumed that the resource usage of other programs is stationary and does not change significantly over time. The load data collection unit 5 collects resource utilization data of the entire computer 1 at least during execution of another program.

FIG. 3 is a diagram showing a data format of the average resource use data 13. The average resource use data 13 is an average resource use rate and an average elapsed time of each component program 2 calculated based on the data of the resource use history 12. The priority indicates the order of selection when allocating each constituent program 2 to the computer 1.

The user setting information 14 includes the configuration program 2
Is an example of a parameter used at the time of allocation scheduling, and is composed of the following parameters. (A) Priority resource usage rate: CPU usage rate (b) CPU usage rate limit value: 90% (c) I / O usage rate limit value: 90% (d) Minimum elapsed time: 10 seconds Priority resource usage rate Specifies which of the CPU usage rate and the I / O usage rate is to be preferentially allocated to the computer 1, and specifies the resource usage rate that can be a bottleneck in the elapsed time. In general, the CPU utilization is specified for a CPU-bound parallel program, and the I / O is specified for an I / O-bound parallel program.
Specify the usage rate. The CPU usage limit specifies the upper limit of the CPU usage of the computer 1, and the I / O usage limit is
Specifies the upper limit of the I / O usage rate as viewed from Calculator 1. The limit value of the resource usage rate gives a resource usage rate that becomes an actual limit when the elapsed time increases sharply due to a slight increase in the resource usage rate when the computer 1 exceeds the limit value. is there. The minimum value of the elapsed time is a parameter that specifies the lower limit of the elapsed time that is set to lower the priority of the configuration program 2 whose elapsed time is small.

FIG. 4 is a diagram showing a data format of the current load data 15. The current load data 15 sets an average resource usage rate as the magnitude of the load for each computer 1.

FIG. 5 is a diagram showing a data format of the predicted load data 16. The predicted load data 16 sets the configuration program 2 allocated to each computer 1 and the predicted resource usage as the magnitude of the load expected when the configuration program 2 is added.

FIG. 6 is a flowchart showing the flow of processing of the scheduler 6. The scheduler 6 stores the history data when the configuration program 2 has been executed at least once in the resource usage history 12, and starts the process before the configuration program 2 is loaded on each computer 1. Each computer 1 may execute a program other than the configuration program 2. The scheduler 6 reads the data of the resource use history 12 via the communication program 7 and the communication path 20 (step 61), calculates an average resource use rate and an average elapsed time for each constituent program 2 (step 62), and It is stored in the resource use data 13 (step 63). Next, the user setting information 14 is read (step 64), and the assignment priority is assigned to each constituent program 2 based on the priority resource usage rate (step 65). For example, if the CPU usage rate is the priority resource usage rate, the priority order of 1, 2,... Next, referring to the average resource usage data 13, if there is a configuration program 2 whose average elapsed time is equal to or less than the minimum value set in the user setting information 14, it is extracted (step 6).
6), the average CPU excluding the extracted configuration program 2
1, 2, ... in order from the configuration program 2 with the highest usage rate
The priority order is re-assigned, and for the excluded configuration programs 2, the priority order is re-assigned so as to give successive precedence orders in descending order of the elapsed time (step 67). By this processing, the configuration programs 2 whose elapsed time is equal to or less than the minimum value are given priority in order from the lowest, and the remaining configuration programs 2 are given priority of 1, 2,... Will be granted. Configuration program 2 with extremely small elapsed time
Is a small load on the computer 1 irrespective of whether or not the priority resource usage rate is large because the load on the computer 1 is small.
Assign to This processing prevents the configuration program 2 with an extremely short elapsed time from being erroneously assigned to the computer 1 with a small load, thereby preventing imbalanced load distribution and increasing the elapsed time of the entire parallel program. Next, the scheduler 6 requests the load data of the computer 1 from the load data collection unit 5 via the communication program 7 and the communication path 20, and stores the obtained resource usage rate of the computer 1 in the current load data 15 (step 68). ). Next, if the assignment of the configuration program 2 has not been completed (step 69)
NO), the highest-priority configuration program 2 that has not been allocated is selected and temporarily allocated to the computer 1 with the lowest priority resource utilization among the computers 1 that have not been allocated, and the following formula is used. Calculates the predicted resource usage rate of the computer 1 (step 70). Predicted resource utilization = current average resource utilization of computer 1 + average resource utilization of selected constituent program 2 The predicted resource utilization is calculated for both the CPU utilization and the I / O utilization. Next, the calculated predicted resource usage rate is compared with the CPU usage rate limit value and the I / O usage rate limit value input from the user setting information 14 (step 71). When neither the CPU usage rate nor the I / O usage rate exceeds the limit value (NO in step 71), the selected configuration program 2 is allocated to the temporarily allocated computer 1 (step 72). Configuration program 2 assigned corresponding to computer 1
, The calculated predicted CPU usage rate and the predicted I / O usage rate. When either the CPU usage rate or the I / O usage rate exceeds the limit value (YES at step 71), a flag is set corresponding to the corresponding computer 1 in the predicted load data 16 to determine whether the allocation has been completed or not. If there is any remaining computer 1 that has not been tried (YES in step 73), the process returns to step 70 to repeat the above processing. When all the configuration programs 2 have been allocated to the computer 1 in this way (step 69 YES), the scheduler 6 reads out each of the configuration programs 2 from the parallel program 11 in accordance with the allocation of the configuration programs 2 on the predicted load data 16 and allocates the allocated computer. 1 and start (step 74). When there is no remaining computer 1 (step 7
3NO), the scheduler 6 displays a message to the effect that the execution of the parallel program is stopped on a display device (not shown) (step 75).

By the processing of the scheduler 6, the configuration program 2 having the resource use data shown in FIG.
Are assigned as shown in FIG. 3 according to the example of the user setting information 14. When the current load data 15 is set as shown in FIG. 4, the configuration program 2 and the prediction The resource usage rate is as shown in FIG. In this example, after the configuration program 2-3 is first allocated to the computer 1-1, the configuration program 2-2 is provisionally allocated to the computer 1-2.
Since the / O usage rate exceeds the limit value, it is allocated to the computer 1-3, and the configuration programs 2-1 and 2-4 are allocated to the computers 1-2 and 1-4 as provisional allocation, respectively.

According to the above-described embodiment, among the constituent programs constituting the parallel program, the resources which are likely to become bottlenecks for extending the execution time of the constituent programs are referred to as the bottlenecks in the resource usage rate. Since the load of the usage rate is allocated to the computer with a small load, the load distribution is performed in such a manner that the elapsed time of the constituent program 2 is not increased, and as a result, the elapsed time or the turnaround time of the entire parallel program 11 can be reduced. Also, at this time, the configuration program 2 having an extremely small elapsed time is allocated to a computer having a relatively large load assuming that the elapsed time of the entire parallel program does not become a bottleneck even if the elapsed time is slightly extended. Such a configuration program 2
Accordingly, the above principle is not disturbed, and as a result, the elapsed time of the entire parallel program 11 is reduced.
Further, in order to eliminate the allocation of the configuration program 2 exceeding the resource usage rate limit of the computer 1, a new bottleneck other than the bottleneck caused by the configuration program 2 itself occurs and the configuration program 2 prevents the elapsed time from increasing, and as a result contributes to shortening the elapsed time of the parallel program 11 as a whole.

If the number of computers constituting the parallel computer is larger than the number of the constituting programs constituting the parallel program, the scheduler 6 selects the computers corresponding to the number of the constituent programs in ascending order of the utilization rate of the resource of interest. Then, the above assignment scheduling may be performed for the selected computer.

In the above embodiment, the CPU usage rate and the I / O usage rate are described as the resource usage rates. And may be separated. Further, a memory (main storage) usage rate may be added.

[0024]

According to the present invention, attention is paid to resources which are likely to become bottlenecks in the constituent programs constituting the parallel program, and the constituent programs which are likely to become bottlenecks in the utilization rate of the resources are converted to a computer having a small load on the resource utilization rate. , The elapsed time or the turnaround time of the entire parallel program is reduced, and in particular, a stable turnaround time of the repeatedly executed parallel program can be realized.

[Brief description of the drawings]

FIG. 1 is a configuration diagram of a distributed memory type parallel computer according to an embodiment.

FIG. 2 is a diagram illustrating a data format of a resource use history 12 according to the embodiment.

FIG. 3 is a diagram illustrating a data format of average resource use data 13 according to the embodiment.

FIG. 4 is a diagram showing a data format of current load data 15 of the embodiment.

FIG. 5 is a diagram illustrating a data format of predicted load data 16 according to the embodiment.

FIG. 6 is a flowchart illustrating a flow of processing of a scheduler 6 according to the embodiment.

[Explanation of symbols]

1: computer, 2: configuration program, 4: history collection program, 6: scheduler, 11: parallel program, 1
2: Resource usage history, 13: Average resource usage data, 14:
User setting information, 15: current load data, 16: predicted load data

Claims (3)

[Claims] A load distribution method scheduled by a computer when each component program of a parallel program composed of a plurality of component programs is allocated to each computer of a parallel computer composed of a plurality of computers and executed in parallel. The resource utilization of the configuration program when each component program is executed on the computer, the resource utilization rate of each computer as a whole without executing the configuration program is acquired, and the utilization rate of the resource of interest is obtained. A load balancing method for a parallel computer, wherein a configuration program is allocated to computers in ascending order of configuration programs from a computer having a lower resource usage rate of a resource of interest. 2. Obtaining an elapsed time when each of the constituent programs is executed by the computer, performing the above assignment by excluding the constituent programs whose elapsed time does not reach a predetermined time, and assigning the elapsed time to the excluded constituent programs. 2. The load balancing method for a parallel computer according to claim 1, wherein the configuration programs are allocated to the computers in ascending order of the configuration programs from the remaining computers in which the utilization rate of the resource of interest is lower. 3. When allocating each constituent program to the computer, if the total resource usage rate obtained by adding the corresponding resource usage rate of the computer to be allocated to each resource usage rate of the configuration program exceeds a predetermined limit value, the allocation is performed. 2. The method according to claim 1, further comprising allocating the resource to a computer having the next highest utilization rate of the resource of interest.