Load balancing and work load minimization of overlapping parallel tasks

In this paper, we propose a unique problem in the assignment of overlapping tasks to processors on a parallel machine, with the twin objectives of minimizing workloads while maintaining good load balance. This problem arises in some applications in VLSI CAD, e.g. parallel compiled VHDL simulation. We assume that the parallel application can be decomposed into a set of tasks, each in turn comprising a finite number of subtasks. Overlapped computations arise as a result of replication of subtasks across tasks in order to reduce the amount of communication performed in fine grained parallel applications. The uniqueness of the problem stems from the fact that overlapping computation on tasks assigned to the same processor is only performed once. Theoretical results on NP-hardness and bounds on the utilization of overlap are provided. A heuristic solution is also proposed. An important application area in VLSI-CAD, parallel compiled event driven VHDL simulation is introduced. Results of the application of our heuristics to this problem are reported on a SUN Sparcserver 1000 multiprocessor.