JPH03113564A - Job parallel executing device - Google Patents

Abstract

PURPOSE:To use the processors with high efficiency by providing a means which fetches and carries out the jobs remaining up to the barrier point of an opened processor and at the same time grants the execution of other processors at the barrier points. CONSTITUTION:A remaining part execution means 6 detects a barrier synchronizing point set right after a processor is opened via another processor assigned to a job under execution and takes out the context of one of processors saved by a context store means 5 to carry out the job part remaining up to a barrier synchronizing point. A substitute execution means 7 carries out the jobs in place of a processor at and after the barrier synchronizing point. Then an execution join means 8 assigns again a processor to the job which is carried out by another processor at the barrier synchronizing point of another processor. Thus it is possible to join the execution at the immediately following barrier synchronizing point when a deprived processor is assigned again to the relevant job. As a result, the processors can be flexibly assigned.

Description

[Detailed description of the invention] [Industrial application field] INDUSTRIAL APPLICATION This invention is utilized for a multiprocessor system.

In particular, it relates to efficient parallel execution means in a multi-job environment.

〔overview〕

The present invention provides a means for parallel execution in a multi-job environment to take over and execute the remaining jobs up to the barrier point of the released processor, and to allow other processors to enter execution at the barrier point. This allows for efficient use of the processor.

[Conventional technology]

In a multiprocessor system, the turnaround time for program execution can be shortened by executing several parts of one program in parallel on multiple processors.

Conventionally, the first method for performing parallel execution of this type was to use a fairly large processing unit in a program as a task, and to allocate each task to an assigned processor for execution. .

A second method is to divide each loop in a program into multiple processors and execute them in parallel.

[Problem that the invention seeks to solve]

However, in the case of the first method, it is necessary to increase the size of the task in order to avoid overhead such as task activation. For example, it is often impossible to make each iteration of a loop a separate task because the overhead would be too large. Even if the task division is optimal for the expected number of processors, if the number of processors assigned to a job is not constant, such as in a multi-job environment, the tasks assigned to the processors can easily become unbalanced. The disadvantage is that it is not always possible to create large independent tasks.

Furthermore, in the case of the second method, in order to correctly set the local area of each processor, a fixed number of processors must be assigned to one job. This has the disadvantage that the processor allocation must be fixed in a multi-job environment, resulting in a significant loss of flexibility.

The present invention aims to eliminate such drawbacks and to provide a job parallel execution device that can flexibly allocate processors.

[Failure to solve the problem]

The present invention provides a job parallel execution device that is equipped with a plurality of processors and is connected to a system in which a job is executed in an environment where the number of processors assigned to the execution of one job is undefined. a storage means for saving the context of one processor assigned to the processor when the processor is released from this job;
Another processor assigned to the job being executed detects the barrier synchronization point immediately after the one processor is released, retrieves the context of the one processor saved in the storage means, and retrieves the context of the one processor and processes the barrier synchronization point immediately after the one processor is released. a first execution means for executing the remaining job part until the barrier synchronization point; a second execution means for executing on behalf of the job to be executed by the one processor after the barrier synchronization point; and a barrier synchronization point for the processor. and execution entry means for reassigning the one processor to a job being executed by the processor of the above nature.

[Effect]

When executing programs in parallel in a multi-job environment, a processor allocation means dynamically allocates or deprives a processor to a job. Further, separate repetitions of independent or partially independent loops in the program of a job are distributed and executed by the processor allocation means to the number of processors allocated to the job. Parts that are commonly executed by all processors other than the parts executed in parallel are executed by each processor. Barrier synchronization means synchronizes all processors currently assigned to the job by processor assignment means. If a processor that has already been allocated during job execution is taken over by another job by the processor allocation means,
Save the stolen processor's context. A processor assigned to this job that was not taken away detects that there is a processor taken away from this job at the barrier synchronization point immediately after, retrieves the saved context of the taken processor, and returns it to the barrier synchronization point. carry out the execution. In addition, other processors that have not been stolen take over the execution of the processor common part of the stolen processor.

When the stolen processor is reassigned to the job, execution begins at the next barrier sync point.

〔Example〕

An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing the configuration of this embodiment.

In this embodiment, as shown in FIG. Processor common part execution means 1, barrier synchronization means 2, loop parallel execution means 3, alternative execution means 7, and processor allocation means 4
, a context storage means 5 , a remaining part execution means 6 , and an execution participant means 8 . In other words, this embodiment is connected to a system in which a job is executed in an environment that includes a plurality of processors and the number of processors assigned to the execution of one job is undefined, and the number of processors assigned to the job being executed is connected to the system. When one processor is released from this job, the context saving means 5, which is a saving means for saving the context of this processor, and the other processors assigned to the job being executed are Remaining part execution means 6, which is a first execution means, detects the immediate barrier synchronization point, retrieves the context of the one processor saved in the storage means, and executes the remaining job part up to the barrier synchronization point; , an alternative execution means 7 which is a second execution means that executes a job to be executed by the one processor on behalf of the one processor after the barrier synchronization point; and execution participant means 8 for reassigning jobs to the jobs being executed by the processors.

Next, the operation of this embodiment will be explained with reference to the drawings. FIG. 2 is a processing flow diagram for FIG. 1.

In step S10, the processor common part execution means 1 executes processing commonly performed by all processors, such as setting local variables of each processor. When a processor is taken away during execution by the processor allocation means 4 in step Sll, the context storage means 5 saves the context being executed on the taken processor in step 312. Thereafter, in step S13, the barrier synchronization means 2 synchronizes all the processors. At this time, if it is detected in step S14 that the processor has been usurped during the execution of the previous processor common part, in step S1
At step 5, the remaining part execution means 6 retrieves the saved context of the usurped processor, and the remaining processors execute the remaining part up to the barrier execution point. Step S16
If a processor is stolen before the previous processor common part and the processor is executed with fewer processors, and if a value is set in a local variable of a processor that will be used later in the previous processor common part, , in step S17, the alternative execution means 7 executes the processor common part of the stolen processor. Thereby, when the processor allocation means 4 returns the stolen processor, it becomes possible to participate in the execution of the returned processor from the next barrier execution point. However, if the operating environment is such that processors that have been stolen are rarely returned, this alternative execution is wasteful and is not performed.

When the processor allocation means 4 returns the processor to the original job due to the completion of the job that stole the processor, the execution participant means 8 places the processor in a waiting state,
When the return of the processor is detected in step 318 after the other processors executing the job have performed the next barrier synchronization, the execution entry means 8 is caused to enter execution in step S19. In step S20, when the data dependencies of each repetition in the job program are independent or partially independent, the loop parallel execution means 3 assigns the independent repetition parts to the respective processors and executes them in parallel. In this part, the execution time is shortened almost in proportion to the number of processors assigned to the job, and
It can be executed correctly no matter how many processors are assigned. In step 321, when the currently executing processor is taken away by the processor allocating means 4, in step S22, the context saving means 5 saves the context that was being executed on the taken processor.

Thereafter, in step S23, the barrier synchronization means 2 synchronizes all the processors. At this time, step S
If it is detected in step S24 that a processor has been usurped during the previous parallel partial execution of the loop, in step S25 the remaining partial execution means 6 retrieves the saved context of the stolen processor and uses the remaining undeprived context. The processor executes the remaining portion up to the barrier execution point. When the processor allocation means 4 returns the processor to the original job due to the end of the job that stole the processor,
After the execution entry means 8 puts the processor in a waiting state and the other processors executing the job perform the next barrier synchronization, when it is detected in step 326 that the processor has returned, the execution entry means 8 puts the processor in a waiting state and in step S27 the execution entry means 8 puts the processor in a waiting state. Participate in the implementation of 8.

〔Effect of the invention〕

As explained above, the present invention efficiently uses the processors assigned to a job in a multi-job environment even when the number of processors assigned to a job changes dynamically by being taken away or returned at any time. There is an effect that can be done.

[Brief explanation of drawings]

FIG. 1 is a block configuration diagram showing the configuration of an embodiment of the present invention. FIG. 2 is a processing flow diagram showing the operation of the embodiment of the present invention. ■... Processor common part execution means, 2... Barrier synchronization means, 3... Loop parallel execution means, 4... Processor allocation means, 5... Context storage means, 6.
... Remaining partial execution means, 7... Alternative execution means, death...
Execution entry means.

Claims (1)

[Scope of Claims] 1. In a job parallel execution device equipped with a plurality of processors and coupled to a system in which a job is executed in an environment where the number of processors allocated to the execution of one job is undefined, A storage means that saves the context of one processor assigned to a running job when the processor is released from this job, and a storage means that saves the context of this processor when one processor assigned to the running job is released; a first execution means that detects a barrier synchronization point immediately after the barrier synchronization point, retrieves the context of the one processor saved in the storage means, and executes the job portion remaining up to the barrier synchronization point; a second execution means for executing a job to be executed by the one processor on behalf of the processor; and execution entry that reassigns the one processor to the job being executed by the other processor at the barrier synchronization point of the other processor. A job parallel execution device characterized by comprising: means.