RUBIC: Online Parallelism Tuning for Co-located Transactional Memory Applications

Abstract

With the advent of Chip-Multiprocessors, Transactional Memory (TM) emerged as a powerful paradigm to simplify parallel programming. Unfortunately, as more cores become available in commodity systems, the scalability limits of a wide class of TM applications become more evident. Hence, online parallelism tuning techniques were proposed to adapt the optimal number of threads of TM applications. However, state-of-the-art solutions are exclusively tailored to single-process systems with relatively static workloads, exhibiting pathological behaviors in scenarios where multiple multi-threaded TM processes contend for the shared hardware resources. This paper proposes RUBIC, a novel parallelism tuning method for TM applications in both single and multi-process scenarios that overcomes the shortcomings of the preciously proposed solutions. RUBIC helps the co-running processes adapt their parallelism level so that they can efficiently space-share the hardware. When compared to previous online parallelism tuning solutions, RUBIC achieves unprecedented system-wide fairness and efficiency, both in single- and multi-process scenarios. Our evaluation with different workloads and scenarios shows that, on average, RUBIC enhances the overall performance by 26% with respect to the best-performing state-of-the-art online parallelism tuning techniques in multi-process scenarios, while incurring negligible overhead in single-process cases. RUBIC also exhibits unique features in converging to a fair and efficient state.