Brief Announcement: A Shared Disk on Distributed Storage

摘要

A shared disk implementation on distributed storage requires consistent behavior of disk operations. Deterministic consensus on such behavior is impossible when even a single storage node can fail. Atomic registers show how consistency can be achieved without reaching consensus, but suffer from a crash consistency problem. The presented shared disk algorithm, based on atomic registers and probabilistic consensus, can survive multiple storage node failures, as long as a majority of nodes respond.