Performance Characteristics of OpenMP Language Constructs on a Many-core-on-a-chip Architecture

Recent emerging many-core-on-a-chip architectures present massive on-chip parallelism through hardware support for multithreading. In order to achieve fast development of parallel applications that exploit this massive intrachip parallelism to achieve highly sustainable performance, suitable programming models are needed. OpenMP, the industry de facto standard for writing parallel programs on shared memory systems, could become a reasonable candidate. To increase our understanding of the behavior and performance characteristics of OpenMP programs on many-core-on-a-chip architectures, this paper presents a performance study of basic OpenMP language constructs on the IBM Cyclops64 architecture, which consists of 160 hardware thread units in a single chip. Compared with previous work on conventional SMP systems [1], the overhead of OpenMP language constructs on C64 many-core architecture is at least one order of magnitude lower.