Tunable Parallel Experiments in a GridRPC Framework: Application to Linear Solvers

摘要

The use of scientific computing centers becomes more and more difficult on modern parallel architectures. Users must face a large variety of batch systems (with their own specific syntax) and have to set many parameters to tune their applications (e.g., processors and/or threads mapping, memory resource constraints). Moreover, finding the optimal performance is not the only criteria when a pool of jobs is submitted on the Grid (for numerical parametric analysis for instance) and one must focus on the wall-time completion. In this work we tackle the problem by using the Diet Grid middleware that integrates an adaptable PaStiX service to solve a set of experiments issued from the simulations of the Aster project.