Multi-node mirrored NVRAM in a virtualized environment

Abstract

The demand for increased data availability and reliability of storage systems has contributed to the design and deployment of multi-node data storage clusters. This paper presents a simulator of one such multi-node, multi-machine cluster. The simulator is architected by extending the NetApp, Inc. 2-node cluster architecture to an N-node design. Data availability is provided by mirroring client requests to a subset of peers in the multi-node cluster. Using this simulator, this thesis explores the relationship between the number of peers that each node mirrors to and the overall mirroring latency. This thesis also explores the performance cost incurred when, in response to a mirroring request from a peer node, a node stores the mirrored data in nonvolatile storage before acknowledgment. Using a workload consisting of multiple write requests to different nodes in the simulator, this thesis finds that there exists a linear relationship between the number of mirroring peers in a cluster and the resulting mirroring latency. Experiments using this workload also reveal a 40% increase in mirroring latency when the mirroring requests are stored on peer nodes persistent storage as opposed to volatile memory.