Experiences with predicting resource performance on-line in computational grid settings

In this paper, we describe methods for predicting the performance of Computational Grid resources (machines, networks, storage systems, etc.) using computationally inexpensive statistical techniques. The predictions generated in this manner are intended to support adaptive application scheduling in Grid settings, and on-line fault detection. Wedescribe a mixture-of-experts approach to non-parametric, univariate time-series forecasting, and detail the effectiveness of the approach using example data gathered from "production" (i.e. non-experimental) Computational Grid installations.