CS Events

Abstract:

File systems are essential software used for storing data in end-user devices, IoT devices, servers, as well as high-performance computing systems. With the advent of ultra-fast storage technologies like NVMe that are slowly replacing traditional hard disks, performance bottleneck of storage system has shifted from hardware to software. A holy grail in modern storage research is to provide direct storage access (i.e., bypassing the OS) but without compromising data integrity, crash-consistency, security, and concurrency. However, prior approaches have built direct-access file systems that satisfy the first three properties but fail to provide concurrency, where tens and hundreds of cores access the file system simultaneously.

In this talk, I will first provide a detailed analysis of concurrency bottlenecks in state-of-the-art OS and user-level file systems. I will then present PARAFS, a direct-access user-level file system that increases concurrency, efficiently schedules I/O requests, and provides light-weight crash consistency. ParaFS achieves these goals by moving the file system into the device firmware, introducing concurrency at the firmware-level, and exploiting storage hardware's compute resources. ParaFS also implements an efficient multi-queue parallel I/O scheduler for efficient data dispatch and fairness. For crash-consistency, ParaFS exploits byte-addressability of modern non-volatile memory. Our evaluation of PARAFS shows more than 2x performance gains over state-of-the-art file systems.