DMooseFS: Design and implementation of distributed files system with distributed metadata server

Abstract

Distributed file system is one of the key blocks of data centers. With the fast increase of user scale, metadata management in distributed file system should also be distributed to multiple nodes so as to achieve high scalability and reliability. However, existing works usually focus on the design of distributed metadata management approach, and very few efforts have been put on real implementation. In this paper, we present the design and implementation of DMooseFS, a system with simple and efficient distributed metadata management. Our work is based on the well-known distributed file system MooseFS. We extend it by introducing multiple metadata servers (MDS). The metadata is divided into non-overlapping parts using static subtree partitioning, and each MDS is assigned with one of the parts according to a hash algorithm. Such a hybrid design is efficient because it can avoid the drawbacks of static subtree partitioning and/or hash-based approaches. We implement DMooseFS by modifying existing implementation of MooseFS and test it in real deployment. Technical issues in implementation are also found and addressed. Experiment results from real world deployment show that our system is effective and efficient, compared with the original version of MooseFS.