Abstract:
Intel has officially released the Optane DC Persistent Memory based on 3D-Xpoint technology in April 2019, which provides new opportunities for building efficient persistent memory storage systems. However, existing software is far from fully exploiting the hardware performance, due to the ignorance of utilizing the byte-addressable feature of persistent memory. This paper proposes a hybrid data page management (HDPM) mechanism. It manages file data by selectively using the copy-on-write technique and log-structure, so as to fully utilize the byte-addressable feature of persistent memory. It can avoid the redundant copy overhead as in traditional approaches when processing un-aligned or small-sized writes. To guarantee the read performance unaffected, HDPM introduces reverse-scanning mechanism, which avoids the additional data copying when rebuilding data pages from the log. HDPM also introduces a multi-stage garbage collection mechanism for log cleaning. When a single log is too large, it’s automatically reclaimed by read/write system calls. When the persistent memory space is limited, a background thread asynchronously reclaims the log space with a lock-free approach, without affecting the normal read/write performance. Experiments show that HDPM provides high write performance. Compared with NOVA, a state-of-the-art persistent memory file system, HDPM exhibits 58% lower write latency at most with the small-sized and write-intensive workload, and provides comparable performance for read operations. Our evaluation with Filebench shows that HDPM outperforms NOVA by 33% at most with 10 concurrent threads.