An Adaptive Thread Partitioning Approach in Speculative Multithreading

Abstract

Thread partition is a core part of Speculative Multithreading (SpMT) technique. The existing thread partition approaches mostly adopt one unique thread partitioning scheme for unknown programs, resulting in high misspeculation ratio, restricting the programs’ speedup improvement due to inappropriate partitioning schemes. This paper which introduces an adaptive thread partition approach (AdapTPA), takes the relationship between program complexity and thread partitioning scheme as the research entry point, and uses the irregular programs as the research carrier, and utilizes formal analysis, probability statistics, mathematical modeling and simulation experiments to reveal the rule that program’s characteristics affect speedup performance, and generates a compound thread partitioning scheme for one program, and selects and executes the most suitable thread partitioning scheme according to the runtime context and the program’s complexity, so to achieve the expected maximum speedups. With the method of path statistics on one program’s control flow graph, the program’s complexity calculation model is set up; A candidate thread partitioning scheme set is constructed on the foundation of classical thread partitioning approaches; Using expert knowledge to guide production rules, a scheme selection mechanism that complies with program complexity is explored. Compared to the heuristic rules-based (HR-based) thread partitioning method, the experiment results show that AdapTPA delivers an average 18.24% performance improvement.