Speculatively Multithreaded Architectures

Abstract

Using the increasing number of transistors to build larger dynamic-issue superscalar processors for the purposes of exposing more parallelism has run into problems of diminishing returns, great design complexity, and high power dissipation. While chip multiprocessors (CMPs) alleviate these problems by employing multiple smaller, power-efficient cores to utilize the available transistors, CMPs require parallel programming which is significantly harder than sequential programming. Speculatively multithreaded architectures address both the programmability issues of CMPs and the power–complexity–performance problems of superscalar processors. Speculatively multithreaded architectures partition a sequential program into contiguous program fragments called tasks which are executed in parallel on multiple cores. The architectures execute the tasks in parallel by speculating that the tasks are independent, though the tasks are not guaranteed to be independent. The architecture provides hardware support to detect dependencies and roll back misspeculations. This chapter addresses the key questions of how programs are partitioned into tasks while maximizing inter-task parallelism and how inter-task control-flow and data dependencies (register and memory dependencies) are maintained especially in the distributed multicore organization employed by the speculatively multithreaded architectures.