Parallel finite element analysis using the Intel Xeon Phi

This paper describes the porting of the open source engineering software ParaFEM to the Intel Xeon Phi processor. The results of a preliminary performance study are presented for a new open source ParaFEM mini-app written especially for the purpose. The main findings of the study are that: (i) The original MPI-based software scales linearly on up to 56 of the 60 available cores; (ii) A new mixed mode MPI/OpenMP implementation boosts performance by a factor of 4 when using 4 threads per core on 1-16 cores (for the largest problem that fits in the Xeon Phi memory) and (iii) The best Xeon Phi solution time is ~2 times faster than the host; here comprising 2 x 12 core “standard” Intel Xeon processors. It appears that scaling beyond 16 cores and 4 threads per core is limited by the amount of work available to each of the 240 threads. The authors propose a number of strategies that can be explored to reduce the memory footprint of individual finite elements so that a much larger problem can be tackled. With more “parallelizable work” per thread, we expect to be able to further improve the performance of the mixed mode MPI/OpenMP implementation. This work will be of interest to researchers and engineers who may wish to evaluate the Intel Xeon Phi for scientific computing, particularly those using the finite element method.