Co-Scheduling of Disk Head Time in Cluster-Based Storage (CMU-PDL-08-113)

Disk timeslicing is a promising technique for storage performance insulation. To work with cluster-based storage, however, timeslices associated with striped data must be co-scheduled on the corresponding servers. This paper describes algorithms for determining global timeslice schedules and mechanisms for coordinating the independent server activities. Experiments with a prototype show that, combined, they can provide performance insulation for workloads sharing a storage cluster—each workload realizes a configured minimum efficiency within its timeslices regardless of the activities of the other workloads. Acknowledgements: We thank the members and companies of the PDL Consortium (including APC, Cisco, DataDomain, Facebook, EMC, Google, HP, Hitachi, IBM, Intel, LSI, Microsoft, NetApp, Oracle, Seagate, Symantec, VMware) for their interest, insights, feedback, and support. We also thank Intel, IBM, Network Appliances, Seagate, and Sun for hardware donations that enabled this work. This material is based on research sponsored in part by the National Science Foundation, via grants #CNS-0326453 and #CCF-0621499, by the Department of Energy, under award number DE-FC02-06ER25767, by the Army Research Office, under agreement number DAAD19–02–1–0389, and by an NDSEG Fellowship sponsored by the Department of Defense.