Automatic monitoring for interactive performance and power reduction.

The goal of interactive applications is to respond to user events under human perception bounds. However, existing operating systems do not have a way of dynamically evaluating the quality of the user experience. In this research we describe a mechanism that, by monitoring the communication between the tasks in the system, can automatically isolate execution episodes that directly impact the user. We use this technique to quantify the impact of multiprocessing on the response times of interactive applications and it forms the basis of a power management scheme that works by reducing the performance level of the processor when peak performance is unnecessary. We find that using two processors instead of one can improve the perceptible response-times of interactive applications by 20% to 30%. Our performance-setting algorithm for processors that support dynamic voltage scaling improves on previous algorithms in that it works equally well with irregular and multiprogrammed workloads and has the ability to ensure that interactive performance is within user specified parameters. Our simulations indicate that as a result of our algorithm, the energy used by the processor can be reduced by 20%--75% without a significant impact on the user experience.Ph.D.Applied SciencesComputer scienceUniversity of Michigan, Horace H. Rackham School of Graduate Studieshttp://deepblue.lib.umich.edu/bitstream/2027.42/131843/2/3000950.pd