Computational Laboratory for Analysis, Modeling, and Visualization (CLAMV)

 

HPC Benchmarks

1. Scaling Example Gaussian16

File: glnb2_cb7-op1-wat.com
Number of Atoms : 151
Level: MO6-2X/6-31G*
Method : Frequency Calculation in Water Continuum
Communication: Shared Memory

n : Number of processor cores
T_n : Elapsed Time (min)
C_n : Computational Costs (min), C_n = n * T_n
S_n : Speed-Up, S_n = T₁ / T_n
E_n : Efficiency, E_n = S_n / n = T₁ / C_n

n	Tn	Cn	Sn	En
1	3995	3995	1.000	100 %
2	2001	3996	1.997	99.8 %
4	1015	4047	3.936	98.4 %
8	523	4149	7.639	95.5 %
16	272	4268	14.69	91.8 %
32	155	4810	25.77	80.5 %
64	95	5764	42.05	65.7 %
128	91	10980	43.90	34.3 %

2. Pinning Example Gaussian16

File: glnb2_cb7-op1-wat.com
Number of Cores : All jobs 64 cores
Command: numactl -C [CoreRange] g16 < glnb2_cb7-op1-wat.com > glnb2_cb7-op1-wat.log

Pinning	Cores	Elapsed Time (min)
No	(0-127) scattered	95
Yes	0-63	149
Yes	32-63,96-127	112
Yes	32-63,96-127 +GPU-Job* 0-31 +GPU-Job** 64-95	123

*: numactl -C 0-31 gmx mdrun -v -nt 32 -gpu_id 0 -nb gpu -deffnm production-md
**: numactl -C 64-95 gmx mdrun -v -nt 32 -gpu_id 1 -nb gpu -deffnm production-md
(Testjob from Kalyan)

3. Pinning Example Gromacs GPU Jobs

File: Nose-Hoover
Number of Cores : All jobs 32 cores
Command: numactl -C [CoreRange] gmx mdrun -v -nt 32 -gpu_id [GPU_ID] -deffnm step3 -nsteps 1000000 -nb gpu -bonded gpu -pme gpu

Test	Pinning	No. Jobs	Cores	GPU	Performance (ns/day)
I	No	1	(0-127), scattered	0	41.7
II	Yes	1	0-31	0	151.5
III	Yes	1	0-31, membind=0	0	?
IV	Yes	2	0-31 32-63	0 1	147.0 60.6
V	Yes	2	0-31 32-63	0 0	95.1 95.2
VI	Yes	2	0-31 64-95	0 1	151.0 150.9
VII	Yes	3	0-31 64-95 +Gaussian 32-63,96-127	0 1 -	141.4 143.1 -

Same Gromacs Job on node channel05 with 12 cores, GPU0 performs 60.6 ns/day

HPC Introduction Download the PDF file more info »