HyCache: A Hybrid User-Level File System with SSD Caching

One of the bottlenecks of distributed file systems (DFS), e.g. Google File System [1] and Hadoop Distributed File System [2], is mechanical hard disk drives (HDD): their slow increase in bandwidth, slow decrease in latency, and exponential increase in capacity, have made modern storage devices quite unbalanced. Making things worse, the low bandwidth and high latency of HDD hinders the exploration of data locality, which is critical to distributed computing applications [3]. Even though non-volatile memory e.g. Solid State Drive (SSD), has been introduced for over a decade, HDDs are still dominant storage media in most systems because of their large capacities and low costs. We propose a middleware called HyCache to manage heterogeneous storage devices for distributed file systems. HyCache provides standard POSIX interfaces through FUSE [4] and works completely in the user space. We show that in the context of file systems, the overhead of user-level APIs (i.e. libfuse) is negligible with multithread support on SSD, and with appropriate tuning can even outperform the kernel-level implementation.