[gmx-users] g_tune_pme big standard deviation in perf.out output

[Carsten Kutzner]

Dear Yanbin,

On Dec 30, 2010, at 9:20 PM, WU Yanbin wrote:
> I'm simulating a SPC/E water box with the size of 4nm by 4nm by 4nm. The command "g_tune_pme" was used to find the optimal PME node numbers, Coulomb cutoff radius and grid spacing size. 
> 
> The following command is used:
> g_tune_pme -np 24 -steps 5000 -resetstep 500 ...
> rcoul=1.5nm, rvdw=1.5nm, fourierspacing=0.12
> 
> The simulation is done with no error. Below is the output:
> ---------------------------
> Line tpr PME nodes  Gcycles Av.     Std.dev.       ns/day        PME/f    DD grid
>    0   0   12          2813.762      187.115        9.604        0.361    4   3   1
>    1   0   11          2969.826      251.210        9.112        0.510   13   1   1
>    2   0   10          2373.469      154.005       11.385        0.445    2   7   1
>    3   0    9          2129.519       58.132       12.665        0.601    5   3   1
>    4   0    8          2411.653      265.233       11.248        0.570    4   4   1
>    5   0    7          2062.770      514.023       13.490        0.616   17   1   1
>    6   0    6          1539.237       89.189       17.547        0.748    6   3   1
>    7   0    0          1633.318      113.037       16.548          -      6   4   1
>    8   0   -1(  4)     1330.146       32.362       20.276        1.050    4   5   1
> ---------------------------
> 
> The optimal -npme is 4.
> 
> It seems to me that the "Std. dev" is too huge.
This is the standard deviation resulting from multiple runs with the
same settings. If you do not specify "-r" for the number of repeats 
explicitly to g_tune_pme, it will do two tests for each setting. For
the optimum of 4 PME nodes the standard deviation is 2.4 percent of the 
mean, thus not large at all.

> Can anyone tell me the meaning of "Gcycles Av." and "Std. dev" and their relations to the accuracy of "ns/day"?
Both the number of CPU cycles as the ns/day values are determined from
the md.log output file of the respective runs. g_tune_pme does the averaging
for you, but you can also look at the individual results, these log files
are still there after the tuning run. The standard deviation is printed
only for the Gcycles - maybe it is a good idea to also print the standard
deviation for the ns/day values. If the standard dev is X percent of the
mean for the cycles, then it is also X percent of the mean ns/day.

> 
> Another question:
> I tried 
> g_tune_pme -np 24 -steps 1000 -resetstep 100 ... (the default value of g_tune_pme)
> rcoul=1.5nm, rvdw=1.5nm, fourierspacing=0.12
> 
> The optimal -npme is 6, different from "-npme=4" as obtained with big "-nsteps".
> Should I increase "-nsteps" even more to get better estimate, or what else parameters should I try?
> 
In principle the results will become more exact, the longer the test runs
are. For your system it seems that the load between the processes is not yet
optimally balanced after the default 100 steps so that -resetstep 500 gives
you a more accurate value. I think the -steps 5000 value is large enough, 
but another test with a higher resetstep value would answer your question.
Since you already know that 7-12 PME nodes will not perform well, I would
try

g_tune_pme -np 24 -steps 5000 -resetstep 5000 -min 0.16 -max 0.25 ...

Regards,
  Carsten

> Do let me know if the questions are not made clear.
> Thank you.
> 
> Best,
> Yanbin
> -- 
> gmx-users mailing list    gmx-users at gromacs.org
> http://lists.gromacs.org/mailman/listinfo/gmx-users
> Please search the archive at http://www.gromacs.org/Support/Mailing_Lists/Search before posting!
> Please don't post (un)subscribe requests to the list. Use the 
> www interface or send it to gmx-users-request at gromacs.org.
> Can't post? Read http://www.gromacs.org/Support/Mailing_Lists


--
Dr. Carsten Kutzner
Max Planck Institute for Biophysical Chemistry
Theoretical and Computational Biophysics
Am Fassberg 11, 37077 Goettingen, Germany
Tel. +49-551-2012313, Fax: +49-551-2012302
http://www.mpibpc.mpg.de/home/grubmueller/ihp/ckutzne