Autonomic management of client concurrency in a distributed storage service

A distributed autonomic system adapts its constituent components to a changing environment. This paper reports on the application of autonomic management to a distributed storage service. We developed a simple analytic model which suggested potential benefit from tuning the degree of concurrency used in data retrieval operations, to suit dynamic conditions. We then validated this experimentally by developing an autonomic manager to control the degree of concurrency. We compared the resulting data retrieval performance with non-autonomic versions, using various combinations of network capacity, membership churn and workload patterns. Overall, autonomic management yielded improved retrieval performance. It also produced a distinct but not significant increase in network usage relative to one non-autonomic configuration, and a significant reduction relative to another.