A Comparison of Two Java Runtime Systems for Parallel Execution of ultithreaded Java Applications on Networks of Workstations

This paper assesses the performance of two Java frameworks for high performance computing (HPC) on networks of workstations (NOWs). The lottery-based work stealing algorithm is intrinsically distributed, and consequetly scalable to an extremely large number of participant workstations. Although proved to be near optimal for the distribution of well-structured multithreaded computations across large number of workstations, no claim has been made regarding the performance of the algorithm for a limited number of workstations or for relatively small computations. In this paper the performance of the lottery-based work stealing algorithm has been compared to that of a centralized work scheduling algorithm to determine whether the lottery-based algorithm is eÆcient across the board. It was found that, the centralized work scheduling algorithm is more eÆcient for small number of workstations. This conclusion suggests that HPC will bene t greatly by adopting a composite work scheduling algorithm. This composite algorithm could use the lottery-based work stealing algorithm to distribute computations across large networks, and a centralized algorithm to distribute within smaller networks.