An Adaptive Logging Framework for Persistent Memories

Abstract

Persistent memory is receiving a tremendous amount of attention recently from both academia and industry. Atomic and durable transactions have been studied to ensure crash consistency in persistent memory. However, whether to use undo or redo logging to execute those transactions is still a hotly debated topic. Redo logging seems appropriate for write-dominated workloads and transactions in high contention scenarios whereas undo logging seems appropriate for read-dominated workloads and transactions in low contention scenarios. This necessitates a priori knowledge on the workload and contention scenario to select an appropriate logging method between redo or undo to achieve better performance. In this paper, we argue that we can obtain the best of both worlds without the need of such a priori knowledge. Particularly, we present an adaptive logging framework that dynamically switches between redo and undo logging at runtime so that the performance is always better than that is obtained from a priori selection of either undo or redo logging. We formally model our framework, prove its correctness, and provide an extensive evaluation of it through a persistent memory emulation of TinySTM using 5 micro-benchmarks and 8 complex benchmarks from STAMP and STAMPEDE suites that are well-known and widely used in the literature. The results show significant benefits of our logging framework.