Concurrent file metadata structure using readers-writer lock

ABSTRACTLinux file systems serialize threads when writing shared files. Recent studies have attempted to adopt range locks on shared files to solve this serialization problem, allowing file I/O to be executed concurrently. However, we have found that even with a range lock, I/O throughput no longer increases after a certain number of cores and decreases rapidly on a manycore server. Through extensive performance profiling, we found the cascading tree lock problem that serializes concurrent accesses to the file metadata structure. A mutex lock-based locking mechanism for each file metadata structure serializes I/O requests in modern Linux file systems such as F2FS. In this paper, we present nCache, a novel file metadata cache framework using readers-writer lock that allows concurrent I/O operations for the shared file. nCache solves the I/O scalability problem in the manycore server while ensuring consistent updates. We implemented nCache in F2FS and evaluated it using FxMark on a 120-core server with high-performance NVMe SSDs. Our extensive evaluations show that nCache achieves maximum device throughput in FxMark's shared file I/O workload. It also shows 4.1x higher throughput compared to the baseline F2FS with range locks for realistic workloads.