Optimizing Shared File Access in Distributed Parallel File System

Distributed parallel file systems are widely deployed in high-performance computing environments. However, they suffer from poor I/O performance under the case of high contention. We analyze and find that the root cause of the problem is due to unnecessary lock range expansion which leads to frequent false lock conflicts. To mitigate the false sharing issue, we propose a new locking mechanism with byte-range granularity. To evaluate the effectiveness of the mechanism, we implemented a prototype distributed file system name NxFS. Experiments conducted on a 4-node cluster showed that our method outperforms Lustre by up to 4.1x.