Effective fault tolerance for agent-based cluster computing

Abstract

Long-lived parallel applications running on workstation clusters are vulnerable to single-node or multiple-node failures. Fault recovery is therefore important in avoiding immature program termination. However, much of the runtime overhead imposed by fault tolerance schemes is generally due to the cost of transferring the checkpoint states of applications by disk I/O operations. In this paper, we propose a fault tolerant model in which disk I/O operations are not required because checkpoint states are transferred between replicated parallel applications. We also describe how the resource consumption of the replicated applications can be effective. To achieve this, applications are constructed as computation agents that can reconfigure dynamically according to resource availability. The fault tolerant model has been implemented and tested on a workstation cluster and a Fujitsu AP3000 multi-processor machine. The measurements of our experiments have showed that efficient fault tolerance can be achieved by replicating parallel applications on workstation clusters.