Career: A Prediction-based Approach to Responsive Distributed Interactive Applications

Northwestern University, Evanston IL

Investigators

Abstract

Over the next few years, large high resolution displays and immersive displays will enable and popularize new kinds of interactive applications that will drive a vast increase in the demand for computational and communications resources. These demands will be met by using increasingly wider area distributed computing environments such as computational grids. Unfortunately, there is a mismatch between the aperiodic soft real-time requirements of these distributed interactive applications and the shared, unreserved, highly dynamic and competitive environments they will run in. To achieve responsiveness these applications will have to adapt their behavior to the vagaries of their execution environments. A number of adaptation mechanisms have been developed or proposed, but there is comparatively little work on how to control these mechanisms to achieve soft real-time constraints. The goal of this project is to develop a control system that can competently advise an application as to how to make use of its adaptation mechanisms to achieve such constraints. The approach is to apply rigorous statistical prediction techniques to predict both how the application's resource demands will vary over time and how the environment's resource availability will vary over time. Such predictions, which are computed on demand at run-time, can then be used by the application or other user-level middleware services to choose an appropriate mapping for the application's tasks with which they can meet their constraints with high probability. The contributions of this project will include a statistical characterization of the dynamic behavior and predictability of the demand for CPU and network resources in distributed interactive applications, a statistical characterization of the dyunamic behavior and predictability of the availability of these resources in distributed computing environments, practical tools for predicting resource demand and availability online and providing application-level performance predictions and adaptation advice to distributed interactive applications, and courses that will introduce graduate students and undergraduates to statistical prediction techniques and data analysis within the context of computer systems research and practice. The project will also produce at least two Ph.D. dissertations. This project will build on the researcher's earlier work, which has shown , for a simplified problem, namely scheduling compute-bound real-time tasks using time series predictions of host load, that the approach described in this proposal can work and can lead to the kinds of contributions described above. In addition to producing scientific results, software artifacts, and educating students, the researcher believes that the path shown here will develop his reputation as the authority on prediction-based services for distributed interactive applications within the high performance distributed computing community.