To BSP or not to BSP in heterogeneous NOWs

We compare the performance of "computation-centric" and "communication-centric" scheduling strategies for sharing a large bag of tasks in a heterogeneous network of workstations (HNOW) N, when one's goal is to accomplish as much work as possible during a prespecified fixed period. The computation-centric strategy strives to maximize the total amount of parallel computing; the communication-centric strategy strives to minimize the perceived overhead of interworkstation communications. We compare the strategies under the HiHCoHP model, which characterizes N via parameters that quantify the relative speeds of its workstations and of its underlying network. While both strategies achieve their respective goals, we find that the difference in their goals has significant unanticipated effects. To wit, the superiority of the optimal computation-centric schedule over the optimal communication-centric schedule increases with HNOW size and with coarser task granularity. Modern trends therefore favor the computation-centric scheduling strategy.