Carnegie Mellon University Parallel Data Lab Technical Report CMU-PDL-10-113. October 2010.
Matthew Wachs, Gregory R. Ganger
Electrical and Computer Engineering
Carnegie Mellon University
Pittsburgh, PA 15213
Workloads that share a storage system should achieve predictable, controllable performance despite the activities of other workloads. One desirable way of expressing performance goals is as bandwidth guarantees. Unfortunately, storage bandwidth is difficult to allocate and manage among workloads, because total system capacity depends on both the workloads' access patterns and on any interference between them. This report demonstrates a new approach to supporting soft bandwidth guarantees, building on explicit performance insulation that bounds interference among workloads and its effect on performance and total system capacity. Combining dynamic disk head timeslicing and slack assignment, this approach eliminates almost all avoidable guarantee violations, leaving just those fundamental ones faced by individual workloads whose locality change too significantly. Experiments with a prototype show an order-of-magnitude decrease in the number of guarantee violations compared to traditional token-bucket based throttling.
KEYWORDS: QoS, quality of service, performance guarantees, storage performance, virtualization, performance insulation, performance isolation, shared storage
FULL TR: pdf