Performance Rating via the Feast Indices

Abstract.

Based on algorithmic and computational studies, we present the Feast Indices which are new indicators for the realistic performance of many recent processors, as typically used in ‘low cost’ PC's/workstations up to (parallel) supercomputers. Since these tests are very specifically designed for the evaluation of modern numerical simulation techniques, with a special emphasis on `large scale' FEM computations and iterative solvers of Krylov-space/multigrid type, they examine new aspects in comparison to the standard benchmark tests (LINPACK, SPEC FP95, NAS, STREAM) and allow new qualitative and particularly quantitative ratings of the various hardware platforms. We explore the computational efficiency of certain Linear Algebra components, hereby applying the typical sparse approaches and additionally the sparse banded style from Feast which enables us to exploit a significant percentage of the available Peak performance. The tests include ‘simple’ matrix-vector applications as well as complete multigrid solvers with very robust smoothers which are necessary for the efficient treatment of highly adapted meshes.