Optimal load sharing in dynamically heterogeneous systems

Heterogeneity of processor speed and time availability is introduced to the paradigm of load sharing among a number of autonomous and independently scheduled heterogeneous computers that communicate via message-passing interconnection system. A divisible job originating at one of the system sites is to be partitioned and executed concurrently on a suite of selected processors, to the extent of the time and processing speed allocation that each computer offers for load sharing, in order to minimize job completion time. The originating computer, after invoking a poll-and-reserve load sharing protocol to obtain the multiple-interval time constraints and corresponding processing speeds declared by each processor on load sharing at the time, uses the time and speed availability data to select an optimal suite of processors and load partitioning thereon which achieves the earliest completion time. The load is characterized as being arbitrarily divisible, a paradigm suitable for loads encountered in many applications including image processing, real-time systems, and the processing of massive experimental data. An efficient algorithm for determining the optimal load partitioning is presented.