Skip to Main Content
This paper consider the situation where compute time can be purchased for the database machine of a simple online bookstore, which we model using the standardized TPC-W benchmark. A service level agreement (SLA) defines the value of the system, using throughput, response time, and expected response time as metrics. The autonomous agent must weigh the potential gain (or loss) in value defined by the SLA against the cost of purchasing (or relinquishing) compute time. This paper implement such an autonomous agent on a simulated partitionable system and show that it is possible to outperform many static choices of compute power, over a range of test workloads and resource costs. The agent uses only raw, low-level system statistics, without the need for custom instrumentation of the middleware or operating system.