Speculative parallelization using software multi-threaded transactions

摘要: With the right techniques, multicore architectures may be able to continue exponential performance trend that elevated of applications all types for decades. While many scientific programs can parallelized without speculative parallelism appears key continuing this general-purpose applications. Recently-proposed code parallelization such as those by Bridges et al. and Thies al., demonstrate scalable on multiple cores using speculation divide into atomic units (transactions) span threads in order expose data parallelism. Unfortunately, most software hardware Thread-Level Speculation (TLS) memory systems transactional memories are not sufficient because they only support single-threaded units. Multi-threaded Transactions (MTXs) address problem, but require expensive currently proposed literature. This paper proposes a Software MTX (SMTX) system captures applicability MTX, existing machines. The SMTX yields harmonic mean speedup 13.36x native with four 6-core processors (24 total) running speculatively