OptiTuner: An Automatic Distributed Performance Optimization Service and a Server Farm Application

The next generation of real-time performance-sensitive systems is expected to be more distributed and dynamic. They will have multiple “knobs" that affect performance and resource allocation. Settings of these knobs will need to be jointly and dynamically optimized in the face of a changing workload and resource supply. Centralized approaches, such as Q-RAM, can efficiently maximize global utility of systems of multiple resources and constraints. As such systems grow in scale, complexity, and distribution, distributed solutions may be preferred at run-time. This paper uses optimization decomposition, recently applied at length in network theory, to break complex system-wide global optimization problems into less coupled subproblems that can be optimally solved at run-time in a distributed fashion. We develop a software service, called OptiTuner, that monitors the current performance and the resource availability in a performance-sensitive system and executes the distributed global optimization algorithm derived from optimization decomposition to maximize utility subject to resource and performance constraints. We apply OptiTuner to perform distributed energy optimization in a real-time Web server farm. We show that our distributed solutions achieve roughly up to 14% lower energy consumption than previous literature in our server farm testbed comprising of 18 machines.