An execution service for a partitionable low bandwidth network

As the amount of scientific data grows to the point where the Internet bandwidth no longer supports its transfer it becomes necessary to make powerful computational services available near data repositories. Such services allow remote researchers to start long-running parallel computations on the data. Current execution services do not provide remote users with adequate management facilities for this style of computing. This paper describes the PEX system. It has an architecture based on partitionable group communication. We describe how PEX maintains replicated state in the face of processor failures and network partitions, and how it allows remote clients to manipulate this state. We present some performance numbers, and close with discussing related work.