In parallel processing, it is more efficient to handle a job by decomposing it into finely-grained tasks than trying to manipulate the larger ones. However, if the tasks are finely grained, the overhead related to the synchronization of the processors and the shared memory accesses correspondently increases.
We propose an approach for reducing this overhead by constructing a parallel scheduler which considers the preparation time during plural searchings. In this scheduler, (1) the processes share two upper bounds derived by each searching, (2) the lower bound function is built taking into account the preparation time, and (3) the branching in the search tree is made by a heuristic node selection considering the node level.
An application of the proposed scheduler based on a multi-DSP system to the computation of the inverse dynamics for robot arms shows that the scheduling time can reduced to be more than twice shorter than that of the conventional scheduler, for the case of four processors.
抄録全体を表示