THE RUN-TIME EFFICIENCY OF PARALLEL ASYNCHRONOUS ALGORITHMS

Parallel asynchronous algorithms have been designed to reduce the run-time overhead caused by barrier synchronizations in many numerical and nonnumerical algorithms for multiprocessors. In these algorithms, critical sections are still needed for accessing shared writable data. The run-time efficiency of parallel asynchronous algorithms is mostly affected by the scheduling overhead, by the sharing of critical sections, and by shared-memory conflicts. At the software level, the concurrent accessing of critical sections results in the familiar lock-out problem; at the hardware level, memory access conflicts cause each processor to idle in wait cycles. A simple model to estimate the combined effects of hardware and software contention on the efficiency of asynchronous algorithms is presented in this paper.