The Global Array Programming Model for High Performance Scienti c Computing

Motivatedby t he c haracteristics of current parallel architectures, we h ave d e-veloped an approach t o t he programming of scalable scientiic applications that combines some o f t he best features of message-passing a n d s h ared-memory programming m o d els. Two assumptions permeate our work. The rst is that m o s t high performance parallel computers have, and will continue to h ave, physically distributed memories with non-uniform memory access NUMA timing c har-acteristics. NUMA machines work best with a p plication programs that h ave a high degree of locality i n t heir memory reference patterns. The s e c o n d a s s u mp-tion is that extra programming eeort is, and will continue to be, required to construct such a p plications. Thus, a recurring t heme i n o u r w ork i s t he d evelop-ment o f t echniques and t ools that minimize the extra eeort required to construct application programs with explicit control of locality. There are signiicant tradeoos among t he important c o n s i d erations of portab-ility, eciency, a n d ease of coding. The m essage-passing programming m o d el is widely used because of its portability. Some a p plications, however, are too complex to b e c o d ed in a message-passing m o d e, if care is to b e t aken to m aintain a balanced computation load and a void redundant computations. The s h ared-memory programming m o d el simpliies coding, but i t i s n o t p o r t able and o f t en provides little control over interprocessor data transfer costs. Other more recent parallel programming m o d els, represented by s u ch languages and facilities as HPFF1, SISALL22, PCNN33, Fortran-MM4, Lindaa5, and s h ared virtual memory, address these problems in diierent w ays and t o v arying d egrees. None o f t hese models represents a n i d eal solution. Global Arrays GAs, the a p proach d escribed here, lead to b o t h s i m p l e c o d-ing a n d eecient execution for a class of applications that a p pears to be fairly common. The k ey concept of a GA …