Time sharing in hypercube multiprocessors

In hypercube multiprocessors, multiple users are normally supported by dividing the cube into subcubes of different dimensions. A user request for a subcube may be denied depending on how other subcubes were previously allocated or because the allocation algorithm fails to recognize an existing free subcube that would satisfy the request. In both cases, the main consequence is a reduction in system utilization. To support multiple users while avoiding such problems, the authors propose to multiplex all the processors in the cube among the users. In this way, one can get full system utilization and offer all the resources of the system to the users. The authors have conducted experiments with this new approach on an Intel iPSC/860 system, comparing its performance with the one obtained in the conventional cube-partitioning approach. The results show that, for computationally intensive applications, the average execution time per user when multiplexing the processors is generally comparable to the same average when allocating subcubes to the users, and often significantly lower.<>