Low-overhead fault-tolerance for java parallel applications on heterogeneous networked computers

摘要

In this paper, we implement a software package to support low-overhead fault-tolerance for Java parallel applications on Java message Passing System (JMPS) using a coordianted checkpoining algorithm and an efficient recovery algorithm, which is based on causal message logging for improving asynchrony during recovery. Currently, as the technologies of processors and communication networks have rapidly been developed, large-scale distributed systems, consisting of heterogeneous networked computers. provide high-perofrmance parallel computing environments. But as they scale up, their failure rate may also be higher. Therefore, they require the techniques to support low-overhead fault-tolerance. In this paper, the coordinated checkpointing algorithm and the efficient recovery algorithm are used for the requirement in distributed systems. Also as large-scale distributed systems consist of heterogeneous computers, the software package for providing their fault-tolerance must be platform-independent. Thus, we implement the package in Java, which is a hgihly portable programming language.Our performance evaluation results show that the software package implemented in this paper is performed at low-cost.