A Prediction-based Approach to Distributed Interactive Applications

 

Abstract

Interactive applications that benefit ordinary users are becoming increasingly computationally intensive, forcing their computations to migrate from the individual desktop into the resources of distributed computing environments. Simultaneously, such environments are enabling new kinds of interactive applications by making massive computations possible at all. Finally, advances in grid and cluster computing have the potential to bring these possibilities within reach of everyone, enabling a new form of personal computing, one not bound by what is in the PC box. A distributed computing environment is both enticing and challenging. It provides myriad resources for an interactive application to get its work done and please its users, but the supply of those resources varies dynamically because they are shared with other applications and users, and generally cannot be reserved. Furthermore, the resource demand of the application varies greatly because a human is in the loop. To achieve good interactive performance, the application or its supporting middleware must adapt to this dynamicity.  Our work focuses on measuring, understanding and predicting that dynamic behavior. Our general approach is the use of statistical signal processing, with the goal being to present applications and middleware with accurate predictions with which they can make enlightened adaptation decisions that improve the user experience of interactive applications. This approach also leads to pure scientific knowledge about the nature of computer systems and users. In this talk, I will first introduce the notion of distributed interactive application using a scientific visualization framework and an immersive audio environment as examples. Then I will discuss our signal processing approach as it pertains to host load prediction, running time prediction, and soft real-time scheduling in our publicly available RPS system. Finally, I will conclude with a brief overview of our current research directions, including work on wavelet-based techniques, network measurement and prediction, windows monitoring and data reduction, and relational approaches to this information. If time and circumstances permit, the talk will include some demonstrations of our systems.

Short Bio of the Speaker