Fastpara and peerring: two systems in support of data parallel computing

FastPara and PeerRing are two separate systems to support data parallel applications in two different targeted environments. FastPara provides support for developing and running data parallel applications over dedicated and distributed computing clusters. FastPara allows users to express the data parallel computing logic in a declarative way without writing code to deal with low-level communication and control details. Based on the user's declaration, FastPara can generate software components necessary for distributing the computation to computing clusters. At runtime, FastPara provides a series of critical functions for the application's execution in the distributed clusters: process coordination among computing clusters; efficient input and output data communication; history run based workload scheduling and an efficient file data access strategy. We built two applications on top of FastPara to demonstrate its usability and performances. PeerRing is targeted for carrying out data parallel application in a Peer-to-Peer based computing environment. PeerRing focuses on providing two runtime functionalities: dynamic load balancing among computing nodes and fault tolerance for leaving/failure nodes. PeerRing organizes computing nodes into a ring structure and restricts communications to be among neighboring nodes in the ring. This novel design makes it easy to deal with the dynamic and diversified nature of P2P systems and helps reduce communication and control overhead. A series of performance experiments were carried out to prove the runtime efficiency of PeerRing systems.