ITR: Feedback Control of Dynamic Computing Systems

Regents Of The University Of Michigan - Ann Arbor, Ann Arbor MI

Investigators

Abstract

Software systems have become increasingly complex, and are expected to provide service over wide operating ranges. To meet service goals, software developers provide many tuning parameters that must be set carefully for the system to provide acceptable performance and functionality. Tuning these parameters adequately often requires extensive trial-and-error experimentation by system administrators. As workloads and operating conditions continually change, the parameters must be endlessly re-tuned to provide good performance. This research project will apply feedback control to automate this parameter tuning loop in two example systems. The first, a wide-area replicated file system, provides variable consistency and performance through optimistic concurrency control. Feedback will be used to schedule replica reconciliation and data migration to provide acceptable performance and consistency to end users. The second system, a database engine, offers a large variety of tuning parameters to be considered. Feedback control will be used to enable database administrators to manage the growing complexity of their server systems. From these two examples and other work, a framework for feedback control of dynamic computing systems will begin to emerge. This framework will enable future computing systems to offer reliable and robust operation and predictable performance, without costly maintenance and continual manual tuning.