A. Sedrakian, S. Petiton, and N. Emad (France)
Hybrid method, asynchronous communication, explicit restarting, parallel tight coupled algorithm.
The current variety of parallel programming model (MPI, OpenMP,...) and the growing performance requirements for scientific applications (weather forecast, nuclear, etc.) require numerical analysis research to explore hybrid tech niques. This later is a combination of several numerical techniques, or several copies of the same method parame terized differently in order to accelerate the convergence and/or to improve the accuracy of the solution of some large linear algebra problems. This paper shows how paral lel collaborative computations, can outperform traditional ones by solving the communication challenge between the involved methods. Our work focuses on the hybrid com putation of Krylov subspace methods, involved in solv ing most linear algebra problems underlying the previously mentioned large-scale applications. This paper proposes a strategy to realize an efficient hybrid computation, based on asynchronous non-blocking communication among the co-methods, and paves the way to a new approach to paral lelize the tight-coupled algorithms on architecture made of cluster of shared memory nodes. By evaluating the perfor mance of the hybrid Multiple Explicitly Restarted Arnoldi method [11] on IBM SP3 and SP4 platforms, we illustrate the effectiveness of our solution.
Important Links:
Go Back
