Skip to navigation Skip to content
Careers | Phone Book | A - Z Index
Performance and Algorithms Research

November 2015 Results (2nd order)

Here are the results for the older, second order (operators.7pt.c) HPGMG-FV implementation. Each machine was allowed to use any amount of memory per node. Unfortuantely, due to scheduling and allocation limitations, some machines only evaluated a single per-node problem size. Similarly, 'Fraction of System' represents what fraction of the system was available for benchmarking.  In order to reproduce these results, users should use the Chebyshev smoother (-DUSE_CHEBY or --fv-smoother=cheby) and modify finite-volume/source/local.mk to use operators.7pt.c.  However, submission for new ranking must use the approach described for the fourth order.

HPGMG     HPGMG Fraction of Parallelization  DOF per Top500
Rank System Site DOF/s System MPI OMP GPU Process Rank
1 K RIKEN 2.83E+12 100% 82944 8   72M 4
2 Titan (CPU+GPU) Oak Ridge 9.16e+11 100% 16384 4 1 32M 2
 

(CPU-only)

Oak Ridge 2.53E+11 100% 32768 8   16M  
3 Mira Argonne 7.21E+11 100% 49152 64   16M 5
4 Edison NERSC 3.85E+11 100% 131072 1   4M 40
5 Stampede (CPU-only) TACC 1.49E+11 64% 8192 8   2M 10
6 Hopper NERSC 1.21E+11 86% 21952 6   2M 72
7 Piz Daint (CPU-only) CSCS 1.02E+11 78% 4096 8   18M 7
8 SuperMUC LRZ 7.13E+10 15% 2744 8   16M 23
9 BiFrost NSC 4.67E+10 100% 1260 16   176M -
10 Stampede (MIC-only) TACC 2.16E+10 8% 512 180   16M 7
11 Peregrine (IVB-only) NREL 1.08E+10 18% 512 12   2M -
12 Carver NERSC 1.35E+09 5% 125 4   2M -
13 Babbage (MIC-only) NERSC 8.24E+08 30% 27 180   16M -