Fault-tolerance for PastryGrid middleware

This paper analyses the performance of a decentralized and fault-tolerant software layer for Desktop Grid resources management. The grid middleware under concern is named PastryGrid. Its main design principle is to eliminate the need for a centralized server, therefore to remove the single point of failure and bottleneck of existing Desktop Grids. PastryGrid (based on Pastry) supports the execution of distributed application with precedence between tasks in a decentralized way. Indeed, each node can play alternatively the role of client or server. Our main contribution is to propose a fault tolerant mechanism for Pastry-Grid middleware. Since the management of PastryGrid is distributed over the participants without central manager, its control becomes a challenging problem, especially when dealing with faults. The experimental results on the Grid'5000 testbed demonstrate that our decentralized fault-tolerant system is robust because it supports high fault rates.