In large-scale grid environments, accurate failure prediction is critical to achieve effective resource allocation while assuring specified QoS levels, such as reliability. Traditional methods, such as statistical estimation techniques, can be considered to predict the reliability of resources. However, naive statistical methods often ignore critical characteristic behavior of the resources. In particular, periodic behaviors of grid resources are not captured well by statistical methods. In this paper, we present an alternative mechanism for failure prediction. In our approach, the periodic pattern of resource failures are determined and actively exploited for resource allocation with better QoS guarantees. The proposed scheme is evaluated under a realistic simulation environment of computational grids. The availability of computing resources are simulated according to real trace that was collected from our large-scale monitoring experiment on campus computers. Our evaluation results show that the proposed approach enables significantly higher resource scheduling effectiveness under a variety of workloads compared to baseline approaches.