JP2518504B2 - Target program generator - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a program executed in a multiprocessor system composed of plural processors.
The present invention relates to a target program generation device that generates a ram .

[0002]

2. Description of the Related Art Since parallel processing generally involves overhead, if the parallelizable part does not have sufficient granularity,
This will increase the execution time of the program.

Therefore, in the conventional object program generating apparatus of the multiprocessor system, the following processing is performed.

When the granularity of the parallelizable portion can be obtained at compile time, the granularity is compared with an empirically determined threshold value. If the granularity is equal to or greater than the threshold value, parallelized code is generated. Generate code.

Further, when the granularity of the parallelizable portion is unknown at the time of compilation, two codes of the parallelized code and the serial code are generated, and the granularity found at the time of execution is compared with the threshold, and the granularity is the threshold. If the above is satisfied, the parallelized code is executed, and if it is smaller, a code for generating the serial code is generated (for example, SX system software S
UPER-UX FORTRAN77 / SX Parallel Processing Function User's Guide, February 1993, 3rd edition, published by NEC Corporation).

[0006]

However, the above-mentioned conventional technique has a problem that the program performance may deteriorate in a multi-job environment because the threshold value is empirically determined.

That is, conventionally, the threshold value is empirically determined,
Since the time slice value of the multiprocessor system is not taken into consideration, the parallelizable code may execute a parallelizable part whose granularity is smaller than the time slice value. Further, in a multi-job environment, the timing of the time slice time of each processor forming the multi-processor system may deviate.

In such a case, the processing of each task corresponding to the parallelized code is completed within one time slice time,
In addition, the task that has completed the process for the first time remains in the synchronization waiting state until all the tasks corresponding to the parallelized code are completed. FIG. 4 is a diagram showing this state by taking a multiprocessor system composed of two processors PA and PB as an example. As can be seen from the figure, the processing of each of the tasks A and B corresponding to the parallelized code is completed within one time slice time, and the task A, which has completed the processing first, waits for the synchronization until the processing of the task B is completed. It becomes a state. This synchronization waiting time is a time shift time of the time slice time between the processor PA and the processor PB.

As described above, according to the conventional technique, the ratio of the synchronization waiting time to the actual processing time is large, such that the synchronization waiting time for a task whose processing is completed within one time slice time becomes a deviation time between processors. Therefore, there is a problem that program performance may deteriorate in a multi-job environment.

It is an object of the present invention to provide an object program generation device capable of improving program performance in a multi-job environment by reducing the ratio of wasted time due to synchronization waiting when parallelized code is executed. It is in.

[0011]

In order to achieve the above object, the present invention aims to generate a program to be executed by a multiprocessor system composed of a plurality of processors.
In the program generation device , parallelized code generation means for generating a parallelized code for a parallelizable part of the source program, serial code generation means for generating a serial code for the parallelizable part of the source program, and at the time of compilation. A code that generates the granularity of the parallelizable part if the granularity of the parallelizable part of the source program can be determined, and a code that determines the granularity of the parallelizable part at execution time if it cannot be determined. Granularity code generating means for generating, a time slice value extraction code generating means for generating a code for extracting a time slice value of the multiprocessor system, a time slice value, a granularity, and the number of processors constituting the multiprocessor system. Based on the parallel code generation Is provided with a selection code generating means for generating a code for selecting whether to perform one stage of the serial code is parallelized code or the serial code generating means to produce generated at runtime.

[0012]

The parallel code generating means and the serial code generating means respectively generate a parallel code and a serial code for the parallelizable part of the source program.

The granularity code generating means generates a code indicating the granularity of the parallelizable portion of the source program when the granularity of the parallelizable portion of the source program can be obtained at the time of compilation, and when the granularity of the parallelizable portion of the source program cannot be obtained. Generate code that requires granularity at run time.

The time slice value extraction code generating means generates a code for extracting the time slice value of the multiprocessor system.

The selection code generating means generates a code for selecting which of the parallel code or the serial code is to be executed at the time of execution based on the time slice value, the granularity and the number of processors constituting the multiprocessor system. .

[0016]

Embodiments of the present invention will now be described in detail with reference to the drawings.

FIG. 1 is a block diagram of an embodiment of the present invention.

A target program generation device (hereinafter referred to as a computer) which inputs the source program 1 and outputs the target program 3
2) which generates parallel code for the parallelizable part of the source program 1, and parallel code generating means 21 for generating a parallel code for the parallelizable part of the source program 1 and a code indicating the value when the granularity of the parallelizable part is obtained at compile time. Granularity code generation means 22 for generating a code for determining the granularity at the time of execution when the granularity of the parallelizable part is not determined at the time of compilation.
And serial code generation means 23 for generating a serial code for the parallelizable part of the source program 1,
A time slice value extraction code generating means 24 for generating a code for extracting a time slice value of a multiprocessor system, and a time slice value as a threshold for granularity, and either one of a parallelized code and a serial code is selected at the time of execution. The selection code generating means 25 for generating a selection code to be executed and the serial code generating means 26 for generating a serial code for the non-parallelizable part of the source program 1 are configured.

FIG. 2 is a flow chart showing a processing example of the compiler 2, and FIG. 3 is a diagram showing an example of code for one parallelizable part generated by the compiler 2. The present embodiment will be described below with reference to the drawings. The operation of will be described.

Referring to step S1 of FIG. 2, the parallelization code generation means 21 takes out a parallelizable part in the source program 1. Referring to step S2, the parallelization code generation means 21 processes the parallelizable part. Generate parallel code. This parallelization code is shown as parallelization code 34 in FIG.

Next, referring to step S3, the granularity code generation means 22 generates a granularity code.

When the granularity code generation means 22 can obtain the granularity of the parallelizable part at the time of compilation based on the parallelization code generated by the parallelization code generation means 21, the value is used as the granularity code. The granularity value can be obtained by simulating the instruction execution time based on parameters such as the number of operations of the generated code (intermediate text or object code), the number of load / store of operands, and the order of instructions.

On the other hand, if the granularity of the parallelizable part cannot be obtained at compile time, for example, if the parallelizable part includes a loop and the number of iterations of the loop cannot be determined until execution time, the parallel Generate code that computes the granularity of the convertible part at runtime. Also, I
Even if a part that is conditionally executed by an F statement or the like is included, the granularity cannot be obtained unless it is executed.
The granularity code generation means 22 has an appropriate assumption regarding the true rate of the IF statement, or the assumption regarding the true rate is incorporated in the source program 1, and the granularity is calculated at compile time according to the assumption, and the calculated granularity value. It is also possible to generate a code indicating. The code generated by the granular code generation means 22 is shown as code 31 in FIG.

Next, referring to step S4, the serial code generating means 23 generates a serial code for the parallelizable part. This serial code is shown as serial code 35 in FIG.

Next, referring to step S5, the time slice value extraction code generating means 24 generates a code for extracting the time slice value assigned to this program. The code generated by the time slice value extraction code generating means 24 is shown as the code 32 in FIG. Here, if the time slice value does not change, it is performed only once at the beginning of program execution or at the entrance of a procedure, and each part refers to the value to optimize the number of calls to the operating system. It is possible.

With reference to the final step S6, the selection code generating means 25 divides the time slice value T and the granularity G by the number N of processors in the multiprocessor system, G.
/ N, and if T <= G / N, parallelization code 3
4 is executed, and if T> G / N, serial code 35
If then else en
Generate a code that is a combination of difs. This code is shown as code 33 in FIG.

At the time of waiting for synchronization during parallel code execution
In the case of the example of FIG. 4, between the processor PA and the processor P
The time difference of the timing of the time slice value with B
It This is 1/2 of the time slice value T on average.
I can think. Therefore, it is the same for G / N processing.
The waiting time is T / 2, and the total processing time is G / N + T /
2, the ratio W of the synchronization waiting time to the total processing time is W = T
/ 2 ÷ (G / N + T / 2). From this relational expression, T
> G / N W> 1 / 3≈0.333, T ≦ G / N
W ≦ 1 / 3≈0.333. Therefore, if T ≦ G / N
Only if the parallelizable part is
1 / N of the execution time at the time of executing
The parallel code is executed only when the chair value is T or more.
If so, the maximum ratio of synchronization waiting time is 33.3%.
So that the performance of the program will be able to
It becomes possible to improve.

The serial code generating means 26 takes out a non-parallelizable part from the source program 1 and generates a serial code corresponding to the non-parallelizable part.

By repeating the above processing,
The target program 3 in which the source program 1 is parallelized can be generated.

[0030]

As described above, according to the present invention, the time slice value is taken into consideration to determine whether to execute the parallelized code or the serial code, and in a multi-job environment. Moreover, the ratio of the synchronization waiting time to the actual processing time can be reduced as compared with the conventional technique, so that there is an effect that the performance of the parallelized program in a multi-job environment can be improved.

[Brief description of drawings]

FIG. 1 is a block diagram of an embodiment of the present invention.

FIG. 2 is a flowchart showing a processing example of a compiler 2.

FIG. 3 is a diagram showing an example of a code for a parallelizable part.

FIG. 4 is a diagram for explaining a problem of the conventional technique.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 ... Source program 2 ... Compiler 3 ... Object program 21 ... Parallelization code generation means 22 ... Granular code generation means 23 ... Serial code generation means 24 ... Time slice value extraction code generation means 25 ... Selection code generation means 26 ... Serial code generation means

Claims (2)

(57) [Claims] 1. A program for executing in a multiprocessor system including a plurality of processors is generated.
In the object program generation device , parallelized code generation means for generating parallelized code for the parallelizable part of the source program, serial code generation means for generating serial code for the parallelizable part of the source program, and compilation At times, if the granularity of the parallelizable part of the source program can be calculated, a code indicating the granularity of the parallelizable part is generated, and if it cannot be calculated, the granularity of the parallelizable part is calculated at runtime. Granularity code generation means for generating a code, time slice value extraction code generation means for generating a code for extracting a time slice value of the multiprocessor system, time slice value, granularity, and the number of processors constituting the multiprocessor system Based on the parallelization Characterized by comprising a selection code generating means for generating a code for selecting whether the parallelized code or the serial code generating means de generating means to generate to perform one of the serial code generated at run time Purpose
Program generator . 2. The selection code generation means selects the parallelization code when a value obtained by dividing the granularity by the number of the processors is equal to or more than the time slice value, and otherwise selects the serial code. The target program generation device according to claim 1, wherein
Place .