Dynamically allocating sets of fine-grained processors to running computations

An approach to using general-purpose parallel computers is explored which involves mapping hardware resources onto computations instead of mapping computations onto hardware. Problems such as processor allocation, task scheduling and load balancing change significantly under this approach and may become amenable to new attacks. The implementation of this approach used by the FFP machine, whose computation and communication resources are repeatedly partitioned into disjoint groups that match the needs of available tasks from moment to moment, is described. Several consequences of this system are examined.<
>