Evaluation of sparse linear algebra operations in Trilinos
- authored by
- Mohammad Siahatgar, Gabriele Von Voigt
- Abstract
The performance of numerous scientific libraries and applications depends heavily on efficiency of sparse linear algebra operations. In this paper, we survey the performance of several parallel sparse vector and matrix kernels provided in the Trilinos framework on supercomputer systems Cray XC30/40 and IBM Blue Gene/Q. The linear algebra operations in Trilinos are handled by one of the two packages Epetra or Tpetra. While the former is the mostused, the latter is the target of future developments and supports larger scale problems as well as shared memory parallelism. We compare the results obtained from both packages together with the MPI only and hybrid solutions. The hybrid parallelism is managed by the package Kokkos, which aims for performance portability among different architectures. We report the efficiency of a single node of the system and demonstrate the scalability behavior of the benchmarks up to 38,400 cores of the HLRN-III systems. Furthermore, for the Intel processors used in the Cray system we present measurements of the energy consumption of the kernels and compare the Energy-to-Solution between different compilers and parallel programing paradigms. In addition, we discuss the effect on the performance and the energy consumption by linking the vendor provided libraries compared to the user-compiled versions. These extensive comparisons obtained on the top most performant supercomputer systems help users and developers as a starting point for determining an optimal development strategy.
- Organisation(s)
-
Institute of Data Science
L3S Research Centre
- Type
- Conference contribution
- Pages
- 1381-1391
- No. of pages
- 11
- Publication date
- 2016
- Publication status
- Published
- Peer reviewed
- Yes
- ASJC Scopus subject areas
- Artificial Intelligence, Applied Mathematics
- Sustainable Development Goals
- SDG 7 - Affordable and Clean Energy
- Electronic version(s)
-
https://doi.org/10.7712/100016.1893.11500 (Access:
Open)
