JPH0689188A - Automatic parallel processing system - Google Patents

Abstract

PURPOSE:To suppress the overhead to improve the processing efficiency of the whole of a program by processing loops, where guarantee of end values of local data in micro tasks is required, in parallel without exclusive control or synchronous control by the automatic parallel function of a compiler. CONSTITUTION:A loop recognizing means 21 analyzes data dependence relations of loops by a data dependence relation analysis means 22. A parallelism discriminating means 23 recognizes loops which can be executed in parallel. An end value guarantee discriminating means 24 discriminates local data in each task and discriminates whether guarantee of the end value is required after the end of the loop or not. A memory assigning means 25 discriminates data hierarchies and assigns memories to individual hierarchies. A parallel code generating means 26 generates a code for parallel execution and end value guarantee. Thus, shared data between tasks executing the same process is used to guarantee the end values of local data in micro tasks without exclusive control or synchronous control, thereby shortening the execution time of the whole of the program.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an automatic parallel processing system, and more particularly to an automatic parallel processing system for a memory sharing type multiprocessor computer system having a compiler, a linker and a task scheduler.

[0002]

2. Description of the Related Art In the field of science and technology calculation such as fluid analysis, meteorological analysis, plasma analysis, structural analysis, resource exploration, optimization problem, etc., a large array is often processed using some loops. According to the loop structure, such arithmetic processing is
In the method of fetching data one element at a time from the array and sequentially executing it with scalar instructions, it often takes tens of days to occupy and execute a high-performance ultra-large computer, and considering cost performance, etc. It wasn't practical.

Further, even in a method of executing parallel execution by vector instructions by a supercomputer or the like having a vector processor having a vector operation function, in the case of a single processor, there is a limit to the speedup of the processor itself, which is a great leap. No improvement in processing speed can be expected.

For this reason, a multitasking system capable of parallel processing by a multiprocessor has come to be used. Further, the parallel processing includes macro tasking which processes a subroutine or a group of subroutines as a unit, and microtasking which processes a loop, a sentence or a group of sentences as a unit.

When a user uses macrotasking, it is necessary to review top-down whether the algorithm of the program is suitable for parallel processing and analyze the parallelism of processing and the independence of data. Very sophisticated technology is required. Also, when using microtasking, it is necessary to focus on the local part of the program from the bottom up and analyze the parallelism of the processing for each loop iteration and each group of statements, data independence, etc. Yes, automatic parallelization by the compiler is indispensable because it requires very careful attention.

In the automatic parallelization processing, information such as a control structure such as a loop and a definition quotation relation of data is analyzed, a parallel executable part having no data dependency is determined, and a code for the parallel execution is generated. . In addition, even for parts that have data dependencies, exclusive control and synchronization control can be performed,
By modifying the source, parallel execution is possible. However, in general, since the microtask has a small unit of parallel processing, the overhead has a large influence on the processing time of the entire program. Therefore, it is necessary to suppress the overhead by performing parallelization without using exclusive control or synchronous control as much as possible.

In the conventional automatic parallelization processing method, when local data in a microtask is quoted at the end of a loop to be parallelized, the final round of the loop is executed only by multitasking. It is necessary to guarantee the final price by performing serial control and setting the final value of the local data in the multitask to the storage location of the data that is quoted after the loop ends, but it is sufficient for the overhead. I couldn't speed it up.

[0008]

In the above-described conventional automatic parallelization processing system by microtasking, it is necessary to increase the parallel processing unit and suppress exclusive control or synchronous control to improve the processing efficiency of parallelization. Greatly affect the.

An object of the present invention is to improve the processing efficiency of the entire program by parallelizing a loop that requires guarantee of the final value of local data in a microtask without using exclusive control or synchronous control and suppressing overhead. It is to provide an automatic parallelization processing method.

[0010]

The automatic parallelization processing system of the present invention is a multitasking system in which a source program is translated to generate a target program for parallel execution, and the target program and a runtime routine are combined. A source program to be translated by a compiler in an automatic parallelization processing method of a memory-sharing multiprocessor computer system including a linker that generates an executable program of a configuration and a task scheduler that manages tasks for parallel execution After inputting, the loop recognition means for recognizing control structures such as loops in the source program, and the data dependence analysis means for recognizing the data reference relationship in the recognized loop and recognizing the data dependence relationship. Of executable parts in parallel by the control structure and data dependency of the created loop To ensure defined citation data at the time of the preliminary parallel execution, the parallel determination means for recognizing the exclusive control and synchronous control necessary, by the parallel determination means,
For a loop that is determined to be executable in parallel, final value guarantee determination means that identifies local data in each task and determines whether or not a final price guarantee is required at the end of the loop, declaration information of the source program, parallel The data hierarchy of shared data between tasks, shared data between tasks and shared data between tasks and unique data within a task is determined from the data dependency information at run time, end price guarantee information and the maximum number of processors that can be used in the system, and allocated to memory. A parallel code generation for generating a code for parallel execution and a final value guarantee including a mode for exclusive control and synchronous control for the obtained parallel executable part by the memory allocation means for performing the above and the parallelism determination means. Means for executing the execution of the target program including parallel code generated by the compiler in a task scheduler. Depending on the information that is sometimes requested, task generation means for generating a task for parallel execution and a processor that can be used at the time when parallel execution is requested are secured, the secured processor is assigned to the task for parallel execution, and the number of secured processors And a task allocation means for notifying the program of an identification number capable of identifying each processor, a task control means for scheduling parallel execution and exclusive control and synchronization control between tasks, and releasing the task and processor at the end of parallel execution. A task release means is provided, and by using the number of processors that can be executed in parallel and the processor identification number information notified from the task scheduler, a data hierarchy of shared data between same process execution tasks is provided to perform exclusive control and synchronization. In a task that can be executed in parallel without controlling And it is configured by automatically performing closing guarantee local data by the compiler.

[0011]

The automatic parallelization processing method of the present invention uses the number of processors capable of parallel execution notified from the task scheduler and the processor identification number information,
The data scope is the same as the shared data between tasks,
Shared between tasks that are executed sequentially in the same process,
A data hierarchy called shared data between same process execution tasks that guarantees the ordering of definition quoting relationships is provided, and the compiler automatically generates parallel code that enables parallel execution without performing exclusive control or synchronization control, and memory allocation. I am doing it.

[0012]

Embodiments of the present invention will now be described with reference to the drawings. FIG. 1 is a block diagram showing an embodiment of an automatic parallelization processing system of the present invention. 2 is a diagram showing the concept of shared data between same process execution tasks.
On the other hand, FIG. 3 shows the FOR as an example of the source program 1.
It is a figure which shows the program by TRAN language. Also,
FIG. 4 is a diagram showing an example of a program in which the program of FIG. 3 is parallelized by a conventional technique. On the other hand, FIG. 5 is a diagram showing an example of a state of a data hierarchy and memory allocation in which the program of FIG. 3 is parallelized by a conventional technique. In addition, FIG.
FIG. 4 is a diagram showing an example of an execution status of each task in which the program of FIG. 3 is parallelized by a conventional technique. Next, FIG. 7 is a diagram showing an example of a program in which the program of FIG. 3 is parallelized by the automatic parallelization processing method of the present embodiment. Also, FIG.
FIG. 4 is a diagram showing an example of a state of a data hierarchy and memory allocation in which the program of FIG. 3 is parallelized by the automatic parallelization processing method of this embodiment. Further, FIG. 9 is a diagram showing an example of the execution status of each task in which the program of FIG. 3 is parallelized by the automatic parallelization processing method of the present embodiment.

As shown in FIG. 1, this embodiment comprises a compiler 2, a linker 5 and a task scheduler 7. The source program 1 is translated by the compiler 2 to generate a target program 3 for parallel execution. Then, the target program 3 and the run-time routine 4 are linked by a linker 5 to generate an executable program 6 having a multitask structure, and the executable program 6 is linked to the task scheduler 7
Are executed in parallel under the control of.

The concept of shared data between same process execution tasks, which is a feature of the present invention, will be described with reference to FIG. It is premised that a unit that can be executed in parallel (for example, a unit of repeated processing of a loop structure, etc.) has no data dependency between the units, and that unit is used as one task for parallel processing. Intertask shared data 27 is used as a data hierarchy commonly accessed between the parallel processing units.
There is a concept.

In the example of FIG. 2, the FORTRAN language CO is used.
The arrays A, B, and C declared in the MMON statement are referenced in the DO loop, and when the DO loop is executed in parallel, they are treated as inter-task shared data. Array A, B,
There is no data dependency in C, and this DO loop is 2
When parallel execution is performed by one processor, each parallel processing unit is sequentially assigned to the processor as a task.

At this time, the data accessed as a process are the first, third, fifth, ..., 99th elements of the arrays A, B, and C, as shown in FIG. It is kept. Similarly, the data accessed as a process is the second, fourth, sixth, ...
.., 100 elements, and the order of access is maintained. Therefore, between the tasks executed in the same process, the access order of the referenced data is guaranteed to be the same order as in the case of sequentially executing the scalar instructions.

Based on this concept, when a data hierarchy called shared data between same process execution tasks 28 is provided, as shown in FIG. 2, the first, third, fifth, ...
99 elements can be assigned to shared data between same process execution tasks of a process. Similarly, array A,
The second, fourth, sixth, ..., 100 elements of B and C can be assigned to the same process execution task shared data of the process.

Shared data between same process execution tasks 28
Has the same data scope as the task-to-task shared data 27, but is shared only between tasks that are sequentially executed on the same process, and therefore has the characteristic that the order of the data definition citation relationship is guaranteed. . Therefore, by introducing a new data hierarchy called shared data between same process execution tasks 28 into the conventional technique, when parallel processing units having a data dependency relationship are executed in parallel, data having a dependency relationship is assigned to each process. Therefore, it is possible to perform parallel execution without performing exclusive control or synchronous control.

As shown in FIG. 3, an example of the source program 1 is a program for operating in a loop 31 with given arrays A, B, and C, storing the result in an array C and a variable X, and returning the result. , When the loop 31 is parallelized, it is necessary to guarantee the final value of the variable T.

When the program shown in FIG. 3 is translated and executed by the conventional technique, as shown in FIGS.
The task scheduler calls the SERVE routine 41 to secure an available processor at timing 61, and calls the PARALLEL_DO routine 42 at timing 62 to execute the loop 43 N-1 times in parallel.

Then, at timing 63, SERIAL
By calling the routine 44, the end waiting of the loop 43 is synchronized, and the final time (Nth time) of the loop 31 is executed as the serial section 45 by the master task (task 1), and the final value of the in-task unique data T42 is determined. Can be assigned.

Thereafter, at timing 64, END_S
By calling the ERIAL routine 46, the standby state of the slave tasks (task 2, task M) waiting for the completion of execution of the serial section 45 is released. Finally, at timing 65, the RELEASE routine 47 is called to free the processor.

In this embodiment, the timings 63 and 6 shown in FIG.
In order to suppress the synchronization control of No. 4 and improve the execution efficiency of the entire program, the following processing is performed.

The compiler 2 comprises a loop recognizing means 21, a data dependency analyzing means 22, a parallelism determining means 23, a final price guarantee determining means 24, a memory allocating means 25 and a parallel code generating means 26.

The loop recognition means 21 analyzes the control structure of the input source program 1 and recognizes the loop 31. The data dependency relationship analyzing unit 22 analyzes the definition quotation relationship of the data in the loop 31 and recognizes that there is no dependency relationship that spans the repetition of the loop. Parallelism determination means 2
3 determines that the entire loop 31 can be executed in parallel.

Further, the final price guarantee judging means 24 uses the loop 3
Since it is quoted at the end of 1, it is determined that it is necessary to guarantee the closing price, and as shown in FIG. 8, as the shared data 83 between the same process execution tasks, only the maximum number of processors available in the system 7 and PDO84 (for identifying how many times the loop is executed) and PT85 (for data that substitutes the variable T because the tasks executed in the same process are sequentially executed) are prepared. As shown, loop 31 is transformed into loop 74 and loop 7
After the end of the parallel execution of 4, the process that executed the final round of the loop 74 is determined, and the text string 75 that substitutes the element of PT85 corresponding to the process into the variable X as the final value of the variable T is generated.

On the other hand, the memory allocating means 25 converts the respective data appearing in the program into the inter-task shared data 87 and the same process execution task shared data 8 as shown in FIG.
3. Classify and assign to the data hierarchy of the in-task unique data 88. The parallel code generation unit 26 generates a code for parallel execution, including the parallel executable portion recognized by the parallelism determination unit 23 and the text portion modified and generated by the final price guarantee determination unit 24.

As described above, when the program shown in FIG. 3 is translated and executed by the automatic parallelization processing method of this embodiment, as shown in FIG. 7 and FIG.
The SERVER routine 71 is called, the available processor is secured by the task scheduler, and the secured number of processors is returned to the inter-task shared data PNUM 82. Also, by calling the IPID routine 72,
The task scheduler returns the processor identification number to which the own task is assigned to the same process execution task shared data PID86.

Next, at a timing 92, PARALLE
The L_DO routine 73 is called to execute the loop 74 in parallel, and at the timing 93, it is determined whether or not the final round of the loop 74 has been executed in order from the process in which the parallel execution has ended. Substitute the correct closing price for X. After that, at timing 94, RELE
Call the ASE routine 76 to free the processor.

[0030]

As described above, according to the automatic parallelization processing method of the present invention, exclusive control or synchronization is performed in the case of automatically parallelizing a program including a loop in which a final value guarantee of local data in a microtask is required. This is solved by using the data hierarchy of shared data between same process execution tasks without using control, and the overhead of exclusive control and synchronous control is suppressed, which has the effect of reducing the execution time of the entire program. is doing.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of an automatic parallelization processing system of the present invention.

FIG. 2 is a diagram showing a concept of shared data between same process execution tasks.

FIG. 3 shows FORTRA as an example of the source program 1.
It is a figure which shows the program by N language.

FIG. 4 is a diagram showing an example of a program in which the program of FIG. 3 is parallelized by a conventional technique.

5 is a diagram showing an example of a state of a data hierarchy and memory allocation in which the program of FIG. 3 is parallelized by a conventional technique.

FIG. 6 is a diagram showing an example of an execution status of each task in which the program of FIG. 3 is parallelized by a conventional technique.

FIG. 7 is a diagram showing an example of a program in which the program of FIG. 3 is parallelized by the moving parallelization processing method of the present exemplary embodiment.

FIG. 8 is a diagram showing an example of a state of a data hierarchy and memory allocation in which the program of FIG. 3 is parallelized by the automatic parallelization processing method of the present embodiment.

FIG. 9 is a diagram showing an example of the execution status of each task in which the program of FIG. 3 is parallelized by the automatic parallelization processing method of the present embodiment.

[Explanation of symbols]

 1 Source Program 2 Compiler 3 Target Program 4 Runtime Routine 5 Linker 6 Executable Program 7 Task Scheduler 21 Loop Recognition Means 22 Data Dependency Analysis Means 23 Parallelism Judgment Means 24 Final Value Guarantee Means 25 Memory Allocation Means 26 Parallel Code Generation Means 27 Shared data between tasks 28 Shared data between same process execution tasks

Claims (1)

[Claims] 1. A compiler for translating a source program to generate an object program for parallel execution, a linker for combining the object program and a runtime routine to generate an executable program having a multitask structure, and for parallel execution. In the automatic parallelization processing method of the memory sharing type multiprocessor computer system with the task scheduler that manages the tasks of the above, after inputting the source program to be translated into the compiler, control structures such as loops in the source program are input. Parallel execution by the recognized loop recognition means, the data definition analysis means that recognizes the definition citation relationship of the data in the recognized loop and recognizes the data dependency relationship, and the recognized control structure of the loop and the data dependency relationship Judgment of feasible parts and definition of data during parallel execution In order to do so, the parallelism determining means for recognizing the necessary exclusive control and the synchronous control, and the loop determined to be parallel executable by the parallelism determining means identify the local data in each task, A closing price guarantee judging means for judging whether closing price guarantee is necessary at the end of the loop, and from the declaration information of the source program, data dependency information at the time of parallel execution, closing price guarantee information and the maximum number of processors available in the system, the task Memory allocating means for deciding the data hierarchy of inter-process shared data, inter-process shared task shared data, and in-task specific data, and allocating to the memory, and the parallelism determining means for the parallel executable parts obtained. And a parallel code generation means for generating code for parallel execution including the modes for exclusive control and synchronous control and for guaranteeing the closing price. And a task scheduler that generates a task for parallel execution according to the information required when the target program including the parallel code generated by the compiler is executed, and a task scheduler that can be used when parallel execution is requested. Secure processor, allocate the secured processor to the task for parallel execution, and notify the program of the number of secured processors and the identification number that can identify each processor, and scheduling of parallel execution and exclusive control between tasks And a task control means for performing synchronous control, and a task release means for releasing the task and the processor at the end of parallel execution. The number of processors capable of parallel execution and the processor identification number information notified from the task scheduler are used. By doing the same process Automatic parallelization processing characterized by providing a data hierarchy called shared data between row tasks and automatically performing local end-value guarantee of local data within a task that can be executed in parallel without performing exclusive control or synchronous control method.