The utilization of idle compute cycles has been known as most promising and cost-effective way to build a large scale high performance computing system, but not widely used because of the lack of effective idleness prediction techniques. In this paper, we argue PCs at university computer labs have a great potential for the utilization of idle CPU cycles, and propose two techniques for predicting idle cycles of those PCs: heuristic and statistical. Based on these techniques, we present the design and implementation of an idle compute cycle prediction service for computational grids. Our experimental results show that the utilization of idle compute cycles is a viable approach to cost-effective large scale computational grids.