A Low-Cost Consistency Protocol for Replicated Directory Data in Cluster-Based Storage Systems

Abstract:

We discuss a low-cost consistency protocol for replicated metadata in
cluster-based storage systems. Our goal in
maintaining the consistency is to minimize the efforts for porting
applications from single systems to the cluster-based environment. The
access patterns to the replicated metadata in
recent cluster-based storage systems
offer a new opportunity for strong consistency
without sacrificing performance in common cases. We design and
implement a protocol for the atomicity, serialization and recovery of
operations in the face of arbitrary sequences of failures. The
correctness of the protocol is checked with randomized failure
injections into a prototype implementation. We measure the impact of the
consistency protocol on the
performance and scalability of an example cluster-based file system. The
measurement of micro
benchmarks shows that the protocol adds little overhead to common
operations; while the measurement of trace-based operation mixes shows
a speedup of 15.7 in a cluster of 16 nodes.