International audienceP2PDC is an environment for high performance peer to peer computing that allows direct communication between peers. This environment is based on P2PSAP, a self adaptive communication protocol. P2PDC is suited to the solution of large scale numerical simulation problems via distributed iterative methods. dPerf is a performance prediction environment for parallel and distributed applications, with primary interest in programs written in C, C++, Fortran for P2PDC. The dPerf performance prediction tool makes use of static and dynamic analyses combined with trace-based simulation. In this paper, we present a decentralized version of P2PDC and show how dPerf predicts performance for the P2PDC environment. We present new features of P2PDC aimed at making it more scalable and robust. Through experiments with P2PDC and dPerf, we show how to properly choose a peer to peer computing system which can match the computing power of a cluster
International audienceRecently, a new environment for high performance peer-to-peer distributed computing was proposed. This environment , named P2PDC, addresses stable or volatile systems communicating in a decentralized manner using the self-adaptive protocol P2PSAP. P2PDC is devoted to task parallel applications like numerical simulation problems or optimization problems solved via parallel or distributed iterative algorithms. For distributed applications meant to run with P2PDC, a performance prediction tool named dPerf was proposed. dPerf combines static and dynamic analysis with trace-based simulation to provide scientist with information about the execution of their large scale numerical simulation applications. dPerf addresses real parallel and distributed numerical simulation and optimisation applications written in C, C++ or Fortran for P2PDC. This paper introduces an enhancement of the dPerf tool which provides scalable performance prediction results. Scaling is done with respect to (i) network configuration and (ii) number of peers. Scaling predictions based on network configuration is achieved through trace-based simulation, where various architectures can be studied. Scaling predictions based on the number of peers implies analyzing the communication topology and modifying trace files prior to simulation. We present experimental results obtained for the obstacle problem, a C/P2PDC implementation of the code used in mechanics and finance. Prediction for this application is computed under real conditions, with a reduced slowdown and by providing user with scalable results
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.