Communications of the ACM
| Hi-index | 0.00 |
One of the trends in supercomputer evolution lies in the direction of multiprocessor systems with increasing numbers of powerful processors. The user who is interested in attaining the maximum achievable computation rates on these machines must understand the philosophy underlying the parallel processing mechanisms provided on them. This paper focuses on the three types of parallelism currently available on the Cray: vectorization, microtasking, and macrotasking. Experiences with all three constructs are presented with a view to exposing the improvements possible with each of them. While this paper deals with a particular machine, the Cray X-MP/48, many of the observations, comments, and conclusions derived can be generalized to other shared-memory multiprocessor systems.