Mitigating sync overhead in single-level store systems

Abstract

Emerging non-volatile memory technologies offer the durability of disk and the byte-addressability of DRAM, which makes it feasible to build up single-level store systems. However, due to extremely low latency of persistent writes to non-volatile memory, software stack accounts for the majority of the overall performance overhead, one of which comes from crash consistency guarantees. In order to let persistent data structures survive power failures or system crashes, some measures, such as write-ahead logging or copy-on-write, along with frequent cacheline flushes, must be taken to ensure the consistency of durable data, thereby incurring non-trivial sync overhead. In this paper, we propose two techniques to mitigate the sync overhead. First, we leverage write-optimized non-volatile memory to store log entries on chip instead of off chip, thereby eliminating sync overhead. Second, we present an adaptive caching mode policy in terms of data access patterns to eliminate unnecessary sync overhead. Evaluation results indicate that the two techniques help improve the overall performance from 5.88x to 6.77x compared to conventional transactional persistent memory.