Dynamic Communication-Efficient Parallel Sorting on SMPs

This paper proposes a dynamic communication-efficient pattern and introduces a new efficient data-exchange process in parallel sorting, called the DCES (Dynamic Communication-Efficient parallel Sorting) algorithm, to improve the communication time. In this approach, we present the dynamic communication pattern by using a ldquoBroadcast-Checkerrdquo table, which can reduce total iterations to one iteration (in best case), or 2, 3, ..., or up to log2P(log2P+1)/2 (in worst case), while total (static) iterations of the recently study are fixed = log2P(log2P+1)/2. Finally to evaluate the sorting performance, we implemented our DCES algorithm on the SGI Origin2000. Our investigated experimental results have been compared to those of the best of existing algorithms (LBM: Load Balanced Merge sort). In the experiments, the proposed DCES algorithm yielded the improved results over those of the LBM algorithm at least 24% on the system of size P = 4 and at least 34% on the system of size P = 8.