Throttling and speculating on parallel architectures

Throttling of parallelism is of importance to dynamic parallelism models in which the number and sizes of possible parallel code segments (tasks, processes, threads, etc.) are unknown at compile time, and which, if left uncontrolled, may lead to explosive parallelism with resulting slowdown rather than speedup. The primary goal of throttling is preventing a dynamically unfolding parallel execution from creating so many parallel code segments that either: the system runs out of resources with which to manage the parallel segments; the system begins to thrash as a result of the increased process management and communication requirements; or the system simply wastes CPU time creating extra code segments which end up being executed sequentially anyway. On the other hand, speculative parallelism involves the parallel execution of code segments which are not yet known to be required. It can be exploited in two basic forms: parallel search and parallel test. Speculative parallelism can be used when there are ideal resources and the currently executing set of codes is not expected to yield any more parallel threads of execution.<
>