normal Slow mdruns with gromacs 2018.1

  • nanogod
  • nanogod's Avatar Topic Author
  • Offline
  • Senior Boarder
More
6 years 3 days ago #7584 by nanogod
Slow mdruns with gromacs 2018.1 was created by nanogod
Hi,

anyone else noticed that performance slows significantly down with cuda 9.1 and gromacs 2018.1? I dont know from which version since i jumped from gromacs 2016 and cuda 8.

Best regards
Morten

Please Log in or Create an account to join the conversation.

  • nanogod
  • nanogod's Avatar Topic Author
  • Offline
  • Senior Boarder
More
6 years 3 days ago #7585 by nanogod
Replied by nanogod on topic Slow mdruns with gromacs 2018.1
Before update md.log

On 1 MPI rank, each using 12 OpenMP threads

Computing: Num Num Call Wall time Giga-Cycles
Ranks Threads Count (s) total sum %
Neighbor search 1 12 15001 20.363 830.191 9.4
Launch GPU ops. 1 12 300001 13.433 547.687 6.2
Force 1 12 300001 14.810 603.804 6.8
Wait GPU local 1 12 300001 68.948 2811.045 31.8
NB X/F buffer ops. 1 12 585001 15.172 618.568 7.0
Write traj. 1 12 39 0.212 8.642 0.1
Update 1 12 300001 47.775 1947.824 22.0
Constraints 1 12 300001 11.899 485.122 5.5
Rest 24.429 995.973 11.3
Total 217.041 8848.857 100.0

GPU timings
Computing: Count Wall t (s) ms/step %
Pair list H2D 15001 1.581 0.105 1.8
X / q H2D 300001 17.238 0.057 19.4
Nonbonded F kernel 285000 53.969 0.189 60.8
Nonbonded F+prune k. 12000 2.783 0.232 3.1
Nonbonded F+ene+prune k. 3001 0.890 0.297 1.0
F D2H 300001 12.268 0.041 13.8
Total 88.729 0.296 100.0

Average per-step force GPU/CPU evaluation time ratio: 0.296 ms/0.049 ms = 5.991

Core t (s) Wall t (s) (%)
Time: 2604.493 217.041 1200.0
(ns/day) (hour/ns)
Performance: 3582.743 0.007
Finished mdrun on rank 0 Sun Mar 18 07:26:17 2018

Please Log in or Create an account to join the conversation.

  • nanogod
  • nanogod's Avatar Topic Author
  • Offline
  • Senior Boarder
More
6 years 3 days ago - 6 years 3 days ago #7586 by nanogod
Replied by nanogod on topic Slow mdruns with gromacs 2018.1
After update



Computing: M-Number M-Flops % Flops
Pair Search distance check 113111.867628 1018006.809 1.1
NxN RF Elec. + LJ [F] 49815.906048 2490795.302 2.8
NxN RF Elec. + LJ [V&F] 503.095984 38235.295 0.0
NxN LJ [F] 1887833.749440 84952518.725 94.3
NxN LJ [V&F] 19075.148672 1239884.664 1.4
NxN RF Electrostatics [F] 72.985920 2262.564 0.0
NxN RF Electrostatics [V&F] 0.742000 26.712 0.0
Shift-X 556.777116 3340.663 0.0
Bonds 278.400928 16425.655 0.0
Angles 777.602592 130637.235 0.1
Propers 297.600992 68150.627 0.1
Impropers 96.000320 19968.067 0.0
Virial 557.452161 10034.139 0.0
Stop-CM 111.422232 1114.222 0.0
Calc-Ekin 1113.554232 30065.964 0.0
Lincs 604.806048 36288.363 0.0
Lincs-Mat 9446.494464 37785.978 0.0
Constraint-V 1209.608064 9676.865 0.0
Constraint-Vir 30.242016 725.808 0.0
Total 90105943.656 100.0


R E A L C Y C L E A N D T I M E A C C O U N T I N G

On 1 MPI rank, each using 12 OpenMP threads

Computing: Num Num Call Wall time Giga-Cycles
Ranks Threads Count (s) total sum %
Neighbor search 1 12 15001 52.342 2134.340 9.8
Force 1 12 300001 400.974 16350.539 75.2
NB X/F buffer ops. 1 12 585001 44.190 1801.930 8.3
Write traj. 1 12 39 0.197 8.014 0.0
Update 1 12 300001 11.249 458.707 2.1
Constraints 1 12 300001 19.712 803.810 3.7
Rest 4.732 192.953 0.9
Total 533.396 21750.293 100.0

Core t (s) Wall t (s) (%)
Time: 6400.751 533.396 1200.0
(ns/day) (hour/ns)
Performance: 1457.834 0.016
Finished mdrun on rank 0 Sun Mar 25 09:01:56 2018
Last edit: 6 years 3 days ago by nanogod. Reason: explanation added

Please Log in or Create an account to join the conversation.

  • nanogod
  • nanogod's Avatar Topic Author
  • Offline
  • Senior Boarder
More
6 years 22 hours ago #7588 by nanogod
Replied by nanogod on topic Slow mdruns with gromacs 2018.1
Just a short update. I installed gromacs 2016.5 and i got an 100% performance improvement compared to gromacs 2018.1. Cuda is not the issue. Anyone has an idea what is wrong? 2018. is supposed to be a big performance improvement for a system like mine (12 cores + 1 GTX1080)

Please Log in or Create an account to join the conversation.

More
5 years 11 months ago #7589 by peterkroon
Replied by peterkroon on topic Slow mdruns with gromacs 2018.1
Hi Morten,

which mdp settings do you use? If you're not already try using the new-rf parameters from cgmartini.nl/index.php/force-field-parameters/input-parameters . Beyond that I'd refer you to the Gromacs userlist. They'll want at the very least your mdp settings, a log file and a description of your system.

I don't think anybody in the lab runs a lot of things with 2018(.1) yet. We tend to try and finish projects using the same Gromacs version.

Please Log in or Create an account to join the conversation.

Time to create page: 0.106 seconds