Eliminating FUSE Bottleneck by Implementing Multi-threaded Framework on Distributed File System

Abstract

A popular Filesystem in Userspace (FUSE) is widely used as a data accessing protocol on many modern distributed file systems. FUSE works as a bridge for transferring requests from user application to FUSE-based file system. Actually, it receives requests from user application and then processes some corresponding FUSE operations sequentially by a single FUSE thread. Unfortunately, when a FUSE-based distributed file system performs on workload that consists of a relatively huge number of small-file operations or a very large number of clients, it suffers from overhead of request handling caused by the single FUSE threat. This overhead is the bottleneck of the whole system which significantly degrade the overall performance. In this paper, we propose a multi-threaded FUSE framework that receives and processes requests in parallel which can eliminate the bottleneck caused by that original single FUSE thread. As long as there are requests still available waiting inside the FUSE queue, the other new FUSE thread will be automatically created to receive request and performs some specific FUSE operations simultaneously. We incorporated our mechanism into a GlusterFS distributed file system. The experiment results of our proposed mechanism indicated that depending on the workloads and hardware used, the performance upgradation is improved by 32% on small-file writing workload and 35% on small-file reading workload.