A Comparison of SuperLU Solvers on the Intel MIC Architecture

In many science and engineering applications, problems may result in solving a sparse linear system AX=B. For example, SuperLU_MCDT, a linear solver, was used for the large penta-diagonal matrices for 2L) problems and hepta-diagonal matrices for 3D problems, coming from the incompressible blood flow simulation (see [1]). It is important to test the status and potential improvements of state-of-the-art solvers on new technologies. In this work, sequential, multithreaded and distributed versions of SuperLU solvers (see [2]) are examined on the Intel Xeon Phi coprocessors using offload programming model at the EURORA cluster of CINECA in Italy. We consider a portfolio of test matrices containing patterned matrices from LTEMM ([3]) and randomly located matrices. This architecture can benefit from high parallelism and large vectors. We find that the sequential Supertti benefited up to 45 % performance improvement from the offload programming depending on the sparse matrix type and the size of transferred and processed data.