Scheduling Fork-join Computations on Distributed-memory Multiprocessor Systems

In this paper, we develop two analytic models for scheduling fork-join computations on distributed-memory multiprocessor system. The rst model allows for arriving tasks to multiplex processors with existing tasks, while the other model does not allow multiplexing. Both models assume that tasks block for I/O and that the overhead of rescheduling a task to another processor is prohibitive. We compare the performance of both models in terms of the overall job completion time. We show that multiplexing have a performance advantage over non-multiplexing over a wide range of conditions. We point out how the results can be used in the design of an adaptive multiprocessor scheduler.