Optimizing File Systems for High-Performance Storage Devices

Abstract

High-performance storage technologies such as solid-state drives (SSDs) provide low-latency, high throughput, and high I/O parallelism to legacy storage systems. SSDs access data without mechanical overhead, and they often leads to order-of-magnitude improvements in performance over legacy storage devices such as hard disk drives (HDDs). However, replacing HDDs with SSDs while keeping the software I/O stack or not exploiting SSD features does not lead to maximum performance.

In this dissertation, we optimize file systems to fully exploit the SSD features (e.g., low-latency and high I/O parallelism). First, we analyze and explore I/O strategies in the existing file systems on low-latency SSDs. The file systems issue and complete several I/O requests when blocks are not contiguous, which does not take advantage of the low-latency of SSDs. To address this problem, we propose efficient I/O strategies, which transfer requests from discontiguous host memory buffers in the file systems to discontiguous storage segments in a single I/O request. Thus, they enable file systems to fully exploit the performance of low-latency SSDs.

Second, we investigate the locking and I/O parallelism in the existing file systems on highly parallel SSDs. In the file systems, the coarse-grained locking to access shared data structures is used and I/O operations are serialized by a single thread. For these reasons, the file systems often face the problem of lock contention and underutilization of I/O bandwidth on multi-cores with highly parallel SSDs. To address these issues, we enable concurrent updates on data structures and parallelize I/O operations.

We implement our techniques in EXT4/JBD2 and evaluate them on low-latency and highly parallel SSDs. The experimental results show that our optimized file system improves the performance compared to the existing EXT4 file system.