p-HPF支持多范例并行计算的并行编译技术

Compiling Methods Supporting Multi Paradigm Parallel Computing in p HPF

多范例并行是大规模并行应用系统的本质特征 ,实现 p- HPF对多范例并行计算的编译支持不仅可以弥补数据并行范例本身的一些缺点 ,而且可以提高并行应用系统的效率 .文中在论述 cluster环境下 Global,L ocal,Serial三种典型并行计算模型的基础上 ,给出了实现 p- HPF对三种模型的典型代表 F77,F77+MPI,Sca L A PACK调用的并行编译技术 ,包括参数重分布技术、存储转换技术、全局与局部信息交换技术以及局部数组参数的上下界处理技术等 .

This paper first discusses three types of parallel computing models in cluster environment, namely GLOBAL, LOCAL and SERIAL. Then we present key compiling techniques of p-HPF calling F77, F77+MPI, ScaLAPCK which are typical programs of the three computing model. The main contents are as follows: (1) For redistributing arguments between GLOBAL and LOCAL model programs and gathering/scattering arguments between LOCAL and SERIAL subroutines, the communication optimization methods are advanced. (2) The store modality differences of arrays in p-HPF, F77 and C languages are analyzed and the transformation methods between formal and concrete arguments in different paradigm language subroutines are discussed. (3) Two exchanging manners of data distribution information between p-HPF and other local procedures are proposed. The first manner is passing handle of DAD (Distributed Array Description) to subroutines directly, and extrinsic procedures employ the handle and some intrinsic functions to manage global and local information of arguments. The other one is reconstructing DAD information in p-HPF compiler according to requirements of extrinsic subroutines before calling them. (4) The mapping method of lower bounds and upper bound of array arguments between p-HPF program and local F77 procedures is detailed. Finally, we give some examples to verify the correctness and validity of those techniques and conclude that they are effective methods for developing large-scale multi-paradigm parallel applications in cluster environment using high-level parallel language p-HPF.