Efficient Exploitation of Speculation Under Data Flow Paradigms of Control.

摘要

Data flow architectures to exploit the parallelism in large scientific codes are now taking form. However, no approach to exploiting speculative, searching parallelism has been explored, even though (or perhaps because) the potential parallelism of such applications is tremendous. A view of speculation as a process which may proceed in parallel in a controlled fashion is explored, using examples from actual symbolic processing situations. The central issue of exploiting this parallelism is the dynamic containment of the resources necessary to execute large speculative codes. We show efficient structures (graph schemata and architectural support) for executing highly speculative programs (such as expert systems) under a data flow executing paradigm. Controls over cross-procedure parallelism in an extensible manner will be presented, with applications to the various current problems of data flow computation. Approaches to scheduling, prioritization and search tree pruning are considered, evaluated and compared. (kr)