Utilizing Control Independence and Multi-Threading to Exploit P

University Of Maryland, College Park, College Park MD

Investigators

Abstract

The primary objective of the project is to investigate techniques to utilize control independence and compiler-directed multi-threading to exploit instruction-level and thread-level parallelism from ordinary programs. The first step involves investigating the science of ILP that is present when utilizing control independence and multi-threading. Some of the questions that are addressed are: what is required to exploit parallelism, how much of the parallelism can be exploited by simple techniques such as predicated execution or data value prediction, how far should instructions be moved (statically or dynamically) to exploit it how is it impacted by practical considerations, how can it be enhanced by software transformations, on which portions of the code should software transformations be carried out and what should the nature of these transformations be, and what are the characteristics of the storage needed to exploit the available parallelism? The second step involves investigating hardware and software techniques to exploit the benefits of control independence and data speculation in a multi-threading environment. In particular, the use of novel methods to represent/specify control independence (and data independence) is investigated. Issues relevant to multi-threading, such as tradeoffs between the number of speculative paths and performance are also studied using detailed simulation models.