Fault-tolerant Distributed Applications In LiPS

Performing computations using networks of workstations is increasingly becoming an alternative to using a supercomputer. This approach is motivated by the the vast quantities of unused idle-time available in workstation networks. Unlike computing on a tightly coupled parallel computer, where a xed number of processor nodes is used within a computation, the number of useable nodes in a workstation network is constantly changing over time. Additionally, workstations are more frequently subject to outages, eg. due to reboots. The question arises how applications, adapting smoothly to this environment, should be realized. This paper shows how fault-tolerant distributed applications are implemented within LiPS version 2.4, a system for distributed computing using idle-cycles in networks of workstation. This system currently is used at the Universitt at des Saarlandes in Saarbr ucken to perform computationally intensive applications on a net of approximately 250 workstations. The LiPS system Library for Parallel Systems employs the tuplespace programming paradigm, as originally used in the Linda 1 programming language. Applications implemented using this paradigm easily adapt to changes in availability as they occur in workstation networks. Applications are enabled to terminate successfully in spite of failing nodes by periodically writing checkpoints, freezing the state of a computational process in a le, and keeping track of messages exchanged inbetween checkpoints in a message log. By integrating message logging with the tuplespace, individual processes may restart from their latest checkpoints, while other processes may continue to run unaaected. No other processes are aaected by a single process being restarted from a checkpoint. This also alleviates the need for application-wide synchronisation in order to generate a set of consistent checkpoints.