A data-driven study of operating system energy-performance trade-offs towards system self optimization

This dissertation is motivated by an intersection of changes occurring in modern software and hardware; driven by increasing application performance and energy requirements while Moore's Law and Dennard Scaling are facing challenges of diminishing returns. To address these challenging requirements, new features are increasingly being packed into hardware to support new offloading capabilities, as well as more complex software policies to manage these features. This is leading to an exponential explosion in the number of possible configurations of both software and hardware to meet these requirements. For network-based applications, this thesis demonstrates how these complexities can be tamed by identifying and exploiting the characteristics of the underlying system through a rigorous and novel experimental study. This thesis demonstrates how one can simplify this control strategy problem in practical settings by cutting across the complexity through the use of mechanisms that exploit two fundamental properties of network processing. Using the common request-response network processing model, this thesis finds that controlling 1) the speed of network interrupts and 2) the speed at which the request is then executed, enables the characterization of the software and hardware in a stable and well-structured manner. Specifically, a network device's interrupt delay feature is used to control the rate of incoming and outgoing network requests and a processor's frequency setting was used to control the speed of instruction execution. This experimental study, conducted using 340 unique combinations of the two mechanisms, across 2 OSes and 4 applications, finds that optimizing these settings in an application-specific way can result in characteristic performance improvements over 2X while improving energy efficiency by over 2X