Design of Object Storage Using OpenNVM for High-performance Distributed File System

Abstract

The current trend for high-performance distributed file systems is object-based architecture that uses local object storage to store the file data. The IO performance of such systems depends on the local object storage that manages the underlying low-level storage, such as Fusion IO ioDrive, a flash device connected through PCI express. It provides OpenNVM flash primitives, such as atomic batch write and sparse addressing. We designed an object storage using OpenNVM whose goal is to maximize IOPS/bandwidth performance. Using the sparse address space, it is possible to design object storage as an array of fixed-size regions. Using atomic batch write, the object storage supports the ACID properties in each write. Our prototype implementation achieves 740, 000 IOPS for object creations using 16 threads, which is 12 times better than DirectFS. The write performance achieves 97.7% of the physical peak performance on average.