EFFICIENT PARALLEL COMPUTATION OF ALL-PAIRSN-BODY ACCELERATION BY DO-LOOP FOLDING

Abstract

The computational load inside a do loop is equalized by folding the loop appropriately if the amount of load is a linear function of the loop index. Using this idea, we develop an efficient parallel computation scheme of Newtonian all-pairs N-body acceleration vectors with help from OpenMP architecture. Using a consumer PC with a quad-core eight-thread processor, the new parallel scheme runs 4.2–4.9 times faster than a serial computation when the number of particles exceeds a few hundred.