Proc. 1st Int'l EURO-PAR Conference, Stockholm, Sweden, August 29-31, 1995.
One way to obtain higher computational performance of a computer H is to use a parallel coprocessor unit C which is attached to H and can efficiently solve computationally demanding applications. If such a unit C can be used for concurrent execution of several independent executable programs, the problem of efficient run-time (i.e. dynamic) allocation of programs to C arises. Our aim is to find an allocation strategy that maximizes throughput of the overall system H-C. We present two dynamic allocation approaches for C with mesh-connected architecture. In both cases we reduce the allocation problem to the weighted bipartite matching problem which must be solved during the run-time each time a set of waiting executable programs is to be allocated space in C. The two approaches differ in the way they search for free space in C.