On the impact of mode transition on phased transactional memory performance

Abstract

Several transactional memory implementations that employ state-of-the-art software and hardware techniques to deliver performance have been investigated in the last decade. Phased-based Transactional Memory (PhTM) systems run transactions in phases, such that all transactions in a phase execute in the same (hardware/software) mode.In PhTM, a runtime monitors the execution and decides when to change all transactions to another execution mode. Identifying the right moment to perform a mode transition is a central problem to achieve performance in PhTM systems.This article analyzes PhTM and provides a characterization of mode transitions and their impact on performance. We consider three PhTM implementations: (i) PhTM*, the first phased-based TM designed; (ii) Commit Throughput Measurement (CTM), a general-purpose runtime; and (iii) GoTM, a Graph-oriented runtime.We conduct a performance analysis to identify the drawbacks and benefits of each PhTM implementation with respect to their associated parameters. Results with speedups of up to 10× over the sequential baseline for CTM show that this mechanism generally shows better performance for a diverse set of applications.