Benchmark measurements of current UPC platforms

UPC is a parallel programming language based on the concept of partitioned shared memory. There are now several UPC compilers available and several different parallel architectures that support one or more of these compilers. This paper is the first to compare the performance of most of the currently available UPC implementations on several commonly used parallel platforms. These compilers are the GASNet UPC compiler from UC Berkeley, the v1.1 MuPC compiler from Michigan Tech, the Hewlet-Packard v2.2 compiler, and the Intrepid UPC compiler. The parallel architectures used in this study are a 16-node x86 Myrinet cluster, a 31-processor AlphaServer SC-40, and a 48-processor Cray T3E. A STREAM-like microbenchmark was developed to measure fine- and course-grained shared-memory accesses. Also measured are five NPB kernels using existing UPC implementations. These measurements and associated observations provide a snapshot of the relative performance of current UPC platforms.