Runtime Tuning of STM Validation Techniques

Since the end of the megahertz race in the processor industry and the switch to multicore processors, Software Transactional Memory (STM) has sparked ample interest as a programming model for the now mainstream parallel processing systems. Unfortunately, STM systems still exhibit significant performance overhead over the more traditional, lock-based programming models. A large part of this overhead is due to the STM systems spending significant amount of time validating the state of the shared memory that is visible to a transactions. Significant improvements to the validation systems for STM have been published recently. Transactional Locking II and Lazy Snapshot validation are two time-based validation techniques that have so far shown the most promising performance for a wide variety of applications. Unfortunately, the performance of these techniques depends heavily on the application: number of concurrent jobs, length of the transaction and the read/write ratio of the shared objects can significantly favor one technique over the other. Moreover, for long-running applications that change their behavior over time, neither of these two techniques is optimal. In this paper, we present a runtime tuning strategy that uses profiling to determine the most profitable validation technique. Our runtime tuning strategy can behave as an arbitrary mix of Transactional Locking II and Lazy Snapshot techniques depending on the state of the STM system. We evaluate our technique on a set of STM benchmarks and show that our strategy performs within a couple of percent of the best validation strategy for a given static workload scenario, and that it outperforms both of the above techniques by up to 18% in long-running, dynamically-changing scenarios.