Ismail Akturk

Replication of Metadata in Distributed Storage Systems: Asynchronous Replication Across Multi-master Servers Ismail Akturk

In this study, we examine various metadata server layouts for distributed data storage systems in terms of high availability, scalability and performance. A central metadata server is a single point of failure leading to low availability. Ensuring high availability requires replication of metadata servers. Synchronously replicated metadata servers layout introduces synchronization overhead which degrades the performance of data operations. We propose an asynchronously replicated multi-master metadata servers layout which ensures high availability, scalability and provides better performance. We discuss the implications of asynchronously replicated multi-master metadata servers on metadata consistency and con?ict resolution. Furthermore, we introduce an asynchronous multi-master replication tool that is deployed in the state-wide distributed data storage system called PetaShare, and compare performance of all three metadata server layouts: central metadata server, synchronously replicated multi-master metadata servers and asynchronously replicated multi-master meta- data servers.