Hiding communication overheads in dynamic load balancing for multicomputers

초록

We introduce new techniques for hiding communication overheads involved in dynamic load balancing for hypercubes, meshes and trees. The basic idea is either coalescing some phases of balancing to overlap the transfer of load units on different links or dividing each phase into steps to pipeline the transfer of load unit by unit for maximum utilization of links. Previous schemes perform transfer in a sequential way, leaving many links idle in some phases. The first technique (method1 technique in the paper) collects information about transfers on all the links in all phases and determines which can be overlapped with which. We construct new phases such that all the transfer actions that can be overlapped are combined into a phase. In some cases (especially severely uneven load distribution) the method1 is not so effective since processors should wait for incoming units before sending out. To remedy this, we introduce a new technique (method2 technique in the paper) in which each phase is split into steps and one unit is transferred in a step. Since each processor does not have to wait for all of the units to be sent to its neighbor before sending out, the method1 technique enhances the utilization of links. Our techniques are shown via simulation to result in 15-65% reduction (hiding) to known techniques, depending on the initial load distribution. (C) 2001 Elsevier Science Inc. All rights reserved.