Performance evaluation of macrodataflow computation on shared memory multiprocessors

Abstract

T h e coarse grain parallel processing on shared memory multiprocessor systems has been implemented using multi-tasking. However, this scheme has drawbacks such as difficulty in t h e extraction of parallelism among coarse grain tasks by ordinary users and large dynamic scheduling overhead caused by operating system calfs or run-time library calls. On t h e other hand, in t h e proposed Fortran macrodataflow computation scheme, t h e compiler automatically generates coarse grain tasks called macrotasks, exploits parallelism among macrotasks, and generates a dynamic scheduling routine t h a t schedules macrotasks t o processors at run-time with small overhead. This paper presents performance evaluaThe present authors have proposed the macrodataflow computation scheme[5,6] as a parallel processing scheme among coarse grain tasks, such as basic blocks, loops and subroutines. In macrodataflow computation, (1) the compiler generates coarse grain tasks, or macrotasks, (2) exploits parallelism among macrotasks by calculating the earliest executable conditions, (3) generates a special scheduling routine that schedule macrotasks t o processors at runtime. This paper presents performance evaluation of macrodataflow computation on shared memory multiprocessor systems. Section 2 briefly describes the macrodataflow computation scheme. Section 3 describes performance evaluation of macrodataflow computation on shared memory multiprocessor systems. tion of macrodataflow computation on shared memory multiprocessor systems. T h e results on four pro-2 MACRODATAFLOW COMPUTATION SCHEME cessors of a K S R l show t h a t macrodataflow computation reduces execution time t o 1/2.81 of sequential execution time while a n ordinary multi-threading retion time.This section briefly describes the macrodataflow comscheme is a coarse grain parallel processing scheme among subroutines, loops and basic blocks. duces execution time t o 1/2.19 of sequential execuputation scheme. The macrodataflow computation 2.1 Generation of Macrotasks 1 INTRODUCTIONParallel processing of Fortran programs on shared memory multiprocessor systems has been implemented using multi-tasking and micro-tasking[l, 21. In multitasking, a user specifies parallelism of a program by compiler directives, and an OS or a runtime library generates coarse grain tasks and schedules them to processors at runtime. However, multi-tasking has drawbacks such that the extraction of parallelism among coarse grain tasks by users is difficult, and dynamic scheduling by an OS or a runtime library has large scheduling overhead. Micro-tasking uses loop concurrentization techniques[3, 41 used in parallelizing compilers for multiprocessor systems available on the market. There still exist, however, sequential loops that cannot be concurrentized efficiently because of complex loop carried dependencies and conditional branches. A Fortran program is decomposed into coarse grain tasks called macrotasks. The compiler generates three types of macrotasks, namely, a Block of PseudoAssignment statements (...