Sca1eXFS: Getting scalability of XFS back on the ring

This paper addresses the scalability issue of XFS journaling in the manycore system. In this paper, we first identify two primary causes for XFS scalability failure: the contention between in-memory logging and on-disk logging and the contention among the multiple concurrent in-memory loggings. We then propose three key techniques to address the scalability issues of XFS; (i) Double Committed Item List, (ii) Per-core In-memory Logging, and (iii) Strided Space Counting. Contention between the in-memory logging and the on-disk logging is mitigated using the Double Committed Item List. Contention among the in-memory logging operations is addressed with Per-core In-memory Logging. Strided Space Counting addresses the contention on updating the global journaling state. We call the newly developed filesystem ScaleXFS. In modified varmail workload, the latency of metadata operation, unlink Q, decreases to 1/6th from 0.59 ms to 0.09 ms under 112 threads against stock XFS. The throughput of ScaleXFS corresponds to 1.5 x, 2.2 x, and 2.1 x of the throughput of the stock XFS under varmail, dbench and mdtest, respectively.