Adaptive multi-resource prediction in distributed resource sharing environment

Resource prediction can greatly assist resource selection and scheduling in a distributed resource sharing environment such as a computational grid. Existing resource prediction models are either based on the auto-correlation of a single resource or based on the cross correlation between two resources. In this paper, we propose a multi-resource prediction model (MModel) that uses both kinds of correlations to achieve higher prediction accuracy. We also present two adaptation techniques that enable the MModel to adapt to the time-varying characteristics of the underlying resources. Experimental results with CPU load prediction in both workstation and grid environment show that on average, the adaptive MModel (called MModel-a) can achieve from 6% to more than 90% reduction in prediction errors compared with the autoregressive (AR) model, which has previously been shown to work well for CPU load predictions.