A flexible processor allocation strategy for mesh connected parallel systems

Large mesh connected parallel systems callow multiple applications to run simultaneously on distinct, non-overlapping sub-meshes. This improves overall efficiency, since few (if any) individual applications can effectively use all of the available processors. The sub-mesh allocation strategy forms a crucial component in such systems and affects many performance criteria including system throughput, processor utilization, and application turnaround time. We propose a new sub-mesh allocation strategy called flexfold. When a request for an a×b sub-mesh arrives, flexfold first searches for a×b and b×a sub-meshes, as expected. If appropriate, it may additionally search for a/2×2b, 2a×b/2, 2b×a/2, and b/2×2a sub-meshes. At the expense of sometimes increasing their execution time, flexfold accommodates incoming applications earlier than other strategies. Simulation results indicate that a simple test suffices to determine when this trade-off is beneficial and flexfold achieves better overall performance compared to other algorithms. Flexfold is completely transparent to the application programmer and does not require additional architectural support beyond what already exists