Optimizing the file system with variable-length I/O for fast storage devices

Abstract

Recently, the demand for fast storages is rapidly increasing in HPC environments such as cloud platforms, social network services and desktop users. But HDD-based storages cannot satisfy these demands, and a variety of high performance storages providing lower I/O latency and higher I/O bandwidth have been eagerly developed. Although merely adopting these fast devices in the storage system can take some advantages, it cannot fully utilize their high performance. So, proper optimizations are needed. In this work, we focus on the granularity of the I/O request from the application layer to the block layer. We found that the I/O operation at the page granularity causes huge performance degradation in the case of small size random I/O patterns that are often observed during the execution of mail servers, DB servers, etc. This is because it allows non-requested data to be transferred. Therefore, we propose new file system design that contains two optimizations; 1) an extended I/O interface maintaining user requested data size over all the layers in I/O subsystem, and 2) sub-page mechanism for minimizing non-requested data transfer effectively. We have implemented our approach in the Linux file system. The experimental results show that our solution achieves 1.6 to 6.3 times performance gains.