Differences

This shows you the differences between two versions of the page.

--- cluster:164 [2017/10/27 19:30]
hmeij07
+++ cluster:164 [2017/10/30 14:05]
hmeij07 [PPMA Bench]
@@ Line 564: / Line 564: @@
 ==== PPMA Bench ====
+  * Runs fastest when constrined to one gpu with 4 mpi threads
+  * Room for improvement as gpu and gpu memory are not fully utilized
+  * Adding mpi threads or more gpus reduces ns/day performance
+  * No idea if adding omp threads shows a different picture
+  * No idea how it compares to K20 gpus
 <code>
-PMMA Benchmark Performance Metric (x  nr of gpus)
+nvidia-smi -pm 0; nvidia-smi -c 0
+# gpu_id is done via CUDA_VISIBLE_DEVICES
+export CUDA_VISIBLE_DEVCES=[0,1,2,3]
+# on n78
+cd /home/hmeij/lammps/benchmark
+rm -f /tmp/lmp-run.log;rm -f *.jpg;\
+time /usr/local/mpich-3.1.4/bin/mpirun -launcher ssh -f ./hostfile  -n $STRING_1 \
+/usr/local/lammps-11Aug17/lmp_mpi-double-double-with-gpu -suffix gpu -pk gpu $STRING_2 \
+-in nvt.in -var t 310 > /dev/null 2>&1; grep ^Performance /tmp/lmp-run.log
+PMMA Benchmark Performance Metric ns/day (x  nr of gpus for node output)
-GTX on n78
+Lammps 11Aug17 on GTX1080Ti (n78)
 -n 1, -gpu_id 3
@@ Line 617: / Line 635: @@
 -n 4, -gpu_id 0
 -n 4, -gpuid 0123
+# comparison of binaries running PMMA
+# 1 gpu 4 mpi threads each run
+# lmp_mpi-double-double-with-gpu.log
+Performance: 49.833 ns/day, 0.482 hours/ns, 576.769 timesteps/s
+# lmp_mpi-single-double-with-gpu.log
+Performance: 58.484 ns/day, 0.410 hours/ns, 676.899 timesteps/s
+# lmp_mpi-single-single-with-gpu.log
+Performance: 56.660 ns/day, 0.424 hours/ns, 655.793 timesteps/s

DokuWiki

User Tools

Site Tools

Differences

Page Tools