Thursday, May 9, 2002
In this talk I will describe how such resource-intensive, interactive applications can bound latency in spite of resource scarcity and variability. I advocate a solution based on multi-fidelity computation, which provides a simple yet powerful programming model for applications to dynamically trade off output quality for resource consumption. Four case studies show that the cost of porting legacy applications to this model is modest.
I then describe how a middleware layer can effectively support such computations without any special support from the underlying OS, using predictive resource management. I describe the architecture of our system prototype, and present experimental results that show that it successfully improves latency in real applications.
The last part of the talk describes history-based demand prediction, a technique that is essential to our resource management framework. I will present a case study of GLVU -- a virtual walkthrough application -- and show how its CPU demand can be predicted with a time-varying linear model.