Process and workload migration for a parallel branch-and-bound algorithm on a hypercube multicomputer

This paper describes the design and experimental evaluation of a novel parallel implementation of a branch-and-bound algorithm for solving the Traveling Salesperson Problem on a 32 node Intel hypercube. Issues studied experimentally are trade-offs in speed, memory, and communication costs as well as the effects of workload balancing and node utilization on speedup. Since the actual distribution of work among the parallel tasks of the TSP application cannot be predicted in advance, strategies and tradeoffs regarding the migration of processes from heavily loaded processors or the migration of work from heavily loaded processes can be studied. Toward this end, we have implemented operating system constructs for work and for process migration as extensions to the Intel iPSC hypercube's operating system. Furthermore, operating system support for the rapid sharing of intermediate values of the global objective function being optimized (i.e. 'tour' values in TSP) are provided.