No, I haven’t tried specifying anything yet with ‘numactl’. chris On May 16, 2016, at 9:54 AM, Harms, Kevin <[email protected]> wrote: Chris, for flat+quadrant mode, did you try running with 'mpirun numactl -m 1 ./app.exe' ? kevin
Hello,
I placed some updated slides on the KNL confluence page. The pdf now includes additional comparisons and some timings for the WEST code used in one of the ESP projects.
I noticed that “cache+quadrant” mode can in some cases yield ~2x speedups in runtime for the planewave DFT codes compared to “flat+quadrant” with most of the runtime difference coming from the 3D FFT related kernels. Maybe explicitly allocating memory to HBM will recover performance in “flat+quadrant” mode?
chris
On May 9, 2016, at 10:16 AM, Christopher Knight <[email protected]> wrote:
Hello,
Here’s quick summary of what I’ve seen so far with LAMMPS and QBox on KNL. WEST code is work in progress.
For the LAMMPS rhodo benchmark, I see core-to-core speedup of 3.39x on KNL and a node-to-node speedup of 5.40x using 64 MPI ranks and 2 OpenMP threads on KNL. Using Intel optimized code on KNL (can’t run on BG/Q), I see core-to-core speedup of 10.49x and node-to-node speedup of 12.47x.
For the LAMMPS reaxc benchmark using multi-threaded package, I see core-to-core speedup of 4.05x on KNL and a node-to-node speedup of 5.19x using 16 MPI ranks and 8 OpenMP threads on KNL.
For QBox, I’m using inputs related to an ESP project. I see core-to-core speedups in the range of 2.5 – 3.5x on KNL and a node-to-node speedups in range of 3.8 – 4.5x depending on model, system, and compilation flags. Runtimes sensitive to FFT interface selected for MKL.
chris
_______________________________________________ Intel-nda mailing list [email protected] https://lists.jlse.anl.gov/mailman/listinfo/intel-nda
<knl_LAMMPS_QBOX_May2016.pdf>