An Efficient Journaling Mechanism in Lustre File System for Fast Storage Devices

Abstract

Journaling mechanism is a widely used technique to enhance file system robustness against metadata and data corruptions and minimize file system recovery times after a system crash. However, we found that journaling for the object store in Lustre file system significantly negatively affects the performance. Since it requires that each write request wait for a journal transaction to commit, Lustre introduced serialization to the client request stream and imposed additional latency due to the journal transaction flush for each request. Therefore, threads which handle the I/O operations can only wait for the journal daemon thread to complete journal transaction commit.

In this paper, we provide an efficient journaling mechanism with two schemes to increase the overall efficiency of the Lustre file system. First, we present a parallel journal I/O commit in a cooperative manner with I/O operation threads. This enables the parallel I/O operations instead of the serialized I/O. Second, we use on-the-fly commit to perform the commit processing on the fly without any delayed operations. We implement our schemes in Lustre file system and evaluate them on four 4-core machine with high-performance NVMe SSDs. The experimental results show that our optimized Lustre file system improves the performance by up to 30.1% compared with the existing Lustre file system.