Parallelizing Shared File I/O Operations of NVM File System for Manycore Servers

Abstract

NOVA, a state-of-the-art non-volatile memory (NVM) le system, has limited performance due to its coarse-grained per-le lock when multiple threads perform I/Os to a shared le in a manycore environment. For instance, a writer lock blocks other threads attempting to access the same le, although they access di erent regions of a le. When multiple threads reading the same le share a cache line containing a reader counter, performance can be signi cantly degraded due to cache consistency protocol as we increase the number of readers. This paper proposes a ne-grained segment-based range lock (SRL) that divides a le into multiple segments and manages a lock variable dynamically for each segment. Consequently, write operations can be parallelized without blocking unless there is a con ict in accessing the same range in a le. Moreover, SRL maintains a reader counter per segment that allows multiple reader threads to perform read operations without causing a performance bottleneck. We evaluated an SRL-based NOVA on an Intel Optane DC persistent memory (PM) manycore server. The benchmarking results showed that the average write throughput of the SRL-based NOVA is 3× higher than the original NOVA, and the average read throughput scales linearly, while the original NOVA does not scale.

Scalable range locks for scalable address spaces and beyond

“…We note that the applicability of range locks extends beyond the virtual memory management subsystem. As Kim et al demonstrated recently [24], range locks can be used to optimize shared file I/O operations in a file system; we believe that the range locks we present in this paper can be used as a drop-in replacement for the implementation used in [24]. More generally, drawing from the original motivation behind the concept of range locks, the ideas presented in this paper appear to be a natural fit for parallel file systems; we plan to experiment with such systems in the future work.…”

“…In a recent and highly relevant work [24], Kim et al consider using range locks in the context of parallel file systems, and make a similar observation regarding the lack of scalability of the existing kernel range locks. They followed an alternative design for range locks, which was previously proposed by Quinson et al [33], in which the entire range is divided into (a preset number of) segments, each associated with a reader-writer lock.…”

“…Moreover, choosing the right granularity, i.e., the number of segments, is critical -too few segments would create contention on the underlying reader-writer locks, while too many segments would make range acquisition more expensive -yet, Kim et al do not discuss how the granularity should be tuned. Therefore, we believe the applicability of Kim et al 's scenarios is limited to the cases where the size of the entire range and the granularity of the access are known and static, which is precisely the case considered in [24]. Nevertheless, we include Kim et al 's range locks in our performance study in Section 7.…”

“…In addition, we implemented the recent proposal for range locks by Kim et.al. [24]. Those locks were proposed in the context of pNOVA, a variant of a non-volatile memory file system, hence we denote this version of range locks as pnova-rw.…”

“…Those locks were proposed in the context of pNOVA, a variant of a non-volatile memory file system, hence we denote this version of range locks as pnova-rw. As described in Section 2, pnova-rw operates with a present number of segments, each of a preset size [24]; in our experiments we set this number to 256 segment, spanning one array slot each. We also experimented with other number of segments, spanning multiple slots; although the results were quantitatively different, they lead to similar conclusions.…”