TPAoPI - A Thread Partitioning Approach Based on Procedure Importance in Speculative Multithreading.

Thread partitioning is a core part of thread-level speculation (TLS) to achieve parallelization of irregular serial programs. The existing thread partitioning methods mostly adopt a unified partitioning scheme for all procedures in the same program, so that some procedures cannot obtain their best partition. This papers proposes a thread partitioning approach based on procedure importance (TPAoPI), which chooses to start with importance of procedures in irregular programs, adopting interdisciplinary research methods, creatively introducing the theory of importance in reliability theory, and calculates importance of procedures based on the characteristics of procedures. On the basis of the initial partition scheme, a manual intervention scheme suitable for the procedure is developed according to the importance of procedure, and a new performance evaluation model is used to theoretically evaluate obtained speedups, so that the best partitioning schemes of procedures are identified, exploring the intrinsic law that characteristics of procedures influence their best partitioning schemes. The paper does research from three aspects: calculating importance of procedures, generating the best partitioning schemes of procedures, and building a performance evaluation model, in order to fully exploit the potential parallelism for irregular programs. This paper provides a method for the research and applications of irregular programs parallelization and multi-core parallel computing. Experimental results show TPAoPI delivers an average 20.59% performance improvement than machine learning(ML)-based thread partition approach.