Evaluation of concurrent pools

Abstract

Performance considerations affecting the design of a mechanism that preserves locality and avoids high-latency remote references called the concurrent pools data structure are explored. The effectiveness of three different implementations of concurrent pools is evaluated. Experiments performed on a BBN Butterfly multiprocessor under a variety of workloads shown that the three implementations perform similarly well for light workloads, but that with stressful workloads it appears that a simple algorithm can provide better performance than a complex algorithm, designed to keep remote accesses to a minimum. Implementations can benefit by taking into account information on the nature of the operations performed by each process to help balance the elements among processes that need them.<>