Using Utility Functions to Control a Distributed Storage System

Provisioning, and later optimizing, a storage system involves an extensive set of trade-offs between system metrics, including purchase cost, performance, reliability, availability, and power. Previous work has tried to simplify provisioning and tuning tasks by allowing a system administrator to specify goals for various storage metrics. While this helps by raising the level of specification from low-level mechanisms to high-level storage system metrics, it does not permit trade-offs between those metrics. This dissertation goes beyond goal-based requirements by allowing the system administrator to use a utility function to specify his objectives. Using utility, both the costs and benefits of configuration and tuning decisions can be examined within a single framework. This permits a provisioning system to make automated trade-offs across system metrics, such as performance, data protection, and power consumption. It also allows an automated optimization system to properly balance the cost of data migration with its expected benefits. This work develops a prototype storage provisioning tool that uses an administrator-specified utility function to generate cost-effective storage configurations. The tool is then used to provide examples of how utility can be used to balance competing objectives (e.g., performance and data protection) and to provide guidance in the presence of external constraints. A framework for using utility to evaluate data migration is also developed. This framework balances data migration costs (decreases to current system metrics) against the potential benefits by discounting future expected utility. Experiments show that, by looking at utility over time, it is possible to choose the migration speed as well as weigh alternate optimization choices to provide the proper balance of current and future levels of service.
Sponsored in part by the National Science Foundation (NSF).