Empirical Study of a Data ow Language on the CM - 51

This paper presents empirical data on the behavior of large dataaow programs on a distributed memory multiprocessor. The programs, written in the dataaow language Id90, are compiled via a Threaded Abstract Machine (TAM) for the CM-5. TAM reenes dataaow execution models by addressing critical constraints that modern parallel architectures place on the compilation of general-purpose parallel programming languages. It exposes synchronization, scheduling, and network access so that the compiler can optimize against the cost of these operations. The data presented in this paper evaluates the TAM approach in compiling dataaow languages on stock hardware. We present data on the instruction mix, speedup, scheduling behavior, and locality of large ID90 programs. It is shown that the TAM scheduling hierarchy is able to tolerate long communication latencies, especially when some degree of I-structure locality is present. We investigate how frame allocation strategies, k-bounded loops, and I-structure caching and distribution together aaect the overall eeciency. Finally we document some scheduling anomalies.