New View : Performance Test

[crtrott]

Need to test:

• MiniApps: MiniMD, MiniFE, Lulesh, MiniAero
• LAMMPS: couple test cases
• Trilinos: FENL, Tpetra CG-PerformanceTest
• NALU: crtrott version

[crtrott]

LAMMPS Data:

Run with OMP_NUM_THREADS=24
For GCC OMP_PROC_BIND=TRUE
For Intel KMP_AFFINITY=compact

mpirun -np 2 -bind-to socket -map-by socket ../src/lmp_VARIANT -i in.lj -k on t 24 -sf kk -pk kokkos neigh full comm device -var x 3 -var y 3

Lennard Jones (lammps/bench/in.lj); 1000 steps ; for Cuda set binsize 2.8
Times are given in seconds OLD_VALUE/NEW_VALUE
Compiler 864k/force 864k/neigh 32k/force 32k/neigh 32k/other
GCC/4.8.4 34/34 6.9/6.3 1.18/1.17 0.26/0.24
GCC/4.9.2 33.7/32.6 6.4/6.2 1.18/1.16 0.24/0.24
Intel/15.0.2 32.7/32.7 6.7/6.7 1.13/1.14 0.24/0.24
Cuda/7.5.18 10.5/10.5 2.6/2.6 0.48/0.47 0.16/0.16

[crtrott]

NALU data:

Trilinos: 872a11a5c30f31c41ea1da86ad035239b1788ce8
Nalu-crtrott:9d30a9f9a448919c9c1a4cad393bf5da64aac056

run command:mpirun --bind-to socket -map-by socket -n $1 ${NALU_PATH}/naluX -i nalu_conduction.i

Compiler Assembly Solve
GCC/4.7.2 5.2/5.5* 2.61/2.57
GCC/4.9.2 5.1/4.5 2.64/2.60
GCC/5.1.0 4.9/5.9* 2.62/2.54
Intel/15.0.2 4.8/4.6 2.65/2.61

*some of the assembly data (i.e. exp-view with GCC 5.1.0 and GCC 4.7.2 showed significant spread, while all the other data was pretty consistent. Closer investigation shows that it is due to a single kernel: AssembleNodeSolver while the other kernels are fine, and generally a bit faster with the new view implementation than the old.