Serialization Management Driven Performance in Best-Effort Hardware Transactional Memory Systems

Abstract

Serialization Management is the Best-Effort Hardware Transactional Memory (BE-HTM) counterpart to Software Transactional Memory (STM) Contention Management. A serialization manager uses non-speculative serialization to provide a forward-progress guarantee while simultaneously attempting to provide high application performance. Historically, non-speculative serialization management has been done through a simple policy of allowing a fixed number of retries. This thesis investigates the proposition that application performance can be improved through better Serialization Management. This thesis explores seven serialization managers and their tuning parameters on Blue Gene/Q's BE-HTM system using the Stanford Transactional Applications for Multi-Processing (STAMP) and the Recognition, Mining and Synthesis (RMS-TM) benchmark suites. It presents the first large-scale investigation of Serialization Management for BE-HTM in the literature. This investigation experiments with a large number of values for each tuning parameter on multiple platforms. The main finding is that program performance can be improved by changing the serialization manager. However, performance is actually dominated by the tuning of parameters for each manager and this tuning depends on the benchmark, the thread count, and the platform.