Adaptive parallelism under Equus

The authors describe adaptively parallel computations under Equus (T. Kindberg, 1990; A. V. Sahiner, 1991). These computations execute on a processor pool, and expand and contract as the number of processor nodes allocated to them varies over their run-time. They are based upon a hierarchical master-worker structure. The number of worker processes changes with the number of allocated nodes, and so does the number of processes that act as servers to them (such as the masters). The authors use an image-processing example to describe how workers and servers are added and withdrawn at run-time. The affected processes are synchronised, communication linkages are changed, and in some cases, state is transferred between them. Reconfigurations are transparent to worker (and other client) processes, but not all can be made transparent to servers.<
>