Optimal Compilation of HPF Remappings

Abstract

Applications with varying array access patterns require to dynamically change array mappings on distributed-memory parallel machines. Hpf (High Performance Fortran) provides such remappings, on data that can be replicated, explicitly through the realign and redistribute directives and implicitly at procedure calls and returns. However such features are left out of the hpf subset or of the currently discussed hpf kernel for e ciency reasons. This paper presents a new compilation technique to handle hpf remappings for message-passing parallel architectures. The rst phase is global and removes all useless remappings that appear naturally in procedures. The code generated by the second phase takes advantage of replications to shorten the remapping time. It is proved optimal: A minimal number of messages, containing only the required data, is sent over the network.The technique is fully implemented in hpfc, our prototype hpf compiler. Experiments were performed on a Dec Alpha farm.

Optimizations for Efficient Array Redistribution on Distributed Memory Multicomputers

“…Further, the approach used for multidimensional array redistribution involves a series of one-dimensional redistributions, which can be costly. Reference [18] proposes a general theoretical framework for data redistribution; this framework is currently being implemented.…”

Distribution assignment placement: effective optimization of redistribution costs

“…A different approach has been proposed by Coelho and Ancourt. In [7], they describe compilation techniques which optimize remappings by removing useless ones and taking advantage of replications to shorten individual remappings. Optimality in their sense means that for a given remapping, a minimal number of messages is sent over the network.…”

Efficient Index Generation for Compiling Two-Level Mappings in Data-Parallel Programs

“…That is, all the global indices of array elements have to subtract 9 from their original indices in the special case of a twolevel data-processor mapping to get their real indices in the general case. Thus, the compressed local array of processor p 0 becomes A p0 =(&9, &8, &2, 4,5,11,12,18,24,25). In addition, there are three pseudo active elements on processor p 0 (i.e., T (1, 4, 22)).…”

“…Actually, the compressed local array of processor p 0 for the general two-level dataprocessor mapping shown in Fig. 12 is A p0 = (4,5,11,12,18,24,25). However, if the number of compressed local array elements on processor p is known in advance, the compressed local array of processor p can be generated more efficiently.…”

“…With regard to the block-cyclic distribution, the enumerations of the local memory access sequences and the generation of communication sets for compiling array statements have been extensively discussed recently in the literature [7,9,18,19,23,25,28,29]. On the other hand, researchers who have studied the generation of communication sets for compiling array redistributions have seldom taken arbitrary affine alignment into consideration [5,10,11,20,21,24]. However, affine alignment wastes a lot of memory space if the alignment stride is non unit.…”