An analysis of idle CPU cycles at university computer labs

Grid computing has a great potential for grand challenge scientific problems such as Molecular Simulation, High Energy Physics and Genome Informatics. Exploiting under-utilized resources is crucial for a cost-effective, large-scale grid computing platform (i.e., computational grid), but there has been little research work on how to predict what resources will be underloaded in the near future. In this paper, we analyze idle CPU cycles of PCs at university computer labs and present techniques for predicting idle cycles to be efficiently scheduled for parallel/distributed computing. Our experiments with eight month monitoring data show that the accuracy of our prediction techniques is over 85%. Especially, the ratio of critical failure, which predicts that what is actually busy be idle, was only 3.2% out of total subject PCs during the experimental period.