An optimised lazy task creation technique for iterative and recursive computations

摘要

This paper describes a new runtime technique for lazily exploiting parallelism in iterative and recurisive computations. The technique is based around the Tasklet, a construct used to represent a task that could be executed in parallel if that is necessary to keep the system busy, or inline if not. The major advantage of this technique, when compared to previous schemes, is that only one Tasklet is created for a whole iterative or recusive computation, rather than one for each it-eration of a loop or for every potentially parallel cal in a recursion. The techniques are used by NIP [1,2], a runtime system for parallel platforms (including networks of workstations), that supports implicitly parallel languages combining both functional and object-oriented programming. In the paper, the design and usage of Tasklets is described, and performance results presented.