Performance test of parallel linear equation solvers on Blue Waters--Cray XE6/XK7 system
Parallel linear equation solvers are one of the most important components determining the scalability and efficiency of many supercomputing applications. Several groups and companies are leading the development of linear system solver libraries for HPC applications. In this paper, we present an objective performance test study for the solvers available on a Cray XE6/XK7 supercomputer, named Blue Waters, at National Center for Supercomputing Applications (NCSA). A series of non-symmetric matrices are created through mesh refinements of a CFD problem. PETSc, MUMPS, SuperLU, Cray LibSci, Intel PARDISO, IBM WSMP, ACML, GSL, NVIDIA cuSOLVER and AmgX solver are employed for the performance test. CPU-compatible libraries are tested on XE6 nodes while GPU-compatible libraries are tested on XK7 nodes. We present scalability test results of each library on Blue Waters, and how far and fast the employed libraries can solve the series of matrices. Keywords-parallel linear equation solver; dense direct solver; sparse direct solver; sparse iterative solver; CPU-compatible library; GPU-compatible library