[gmx-users] Vacuum simulation in Gromacs 5

[Jan Jirsák]

Hi Szilárd,
thank you for you reaction!

I prepared a simple input which reproduces the behavior, see

https://owncloud.cesnet.cz/public.php?service=files&t=9e8fc3924fd109f6d48a552e20ceaf11

With -nt 1 it successfully finishes in ca 12 seconds, with -nt 2 it
gets stuck after SETTLE citation.

Regards, Jan

2015-06-15 21:51 GMT+02:00 Szilárd Páll <pall.szilard at gmail.com>:
> On Sun, Jun 14, 2015 at 6:54 PM, Jan Jirsák <janjirsak at gmail.com> wrote:
>> Hi,
>>
>> I did the test and found out that -nt 8 is even slower than -nt 1 !!
>
> FYI: "-nt" is mostly backward compatibility option and for clarity
> it's best to use either -ntmpi or -ntomp, depending on what you mean.
>
>> However, I think that simulation hasn't even properly started with 8
>> threads and got stuck somewhere in the beginning.
>>
>> Details:
>> I used a short run (1000 steps) for testing. mdrun -nt 1 finished
>> after ca 11 hours, whereas not a single checkpoint or run log record
>> have been saved for mdrun -nt 8 (and simulation is still running). Not
>> even starting energies were displayed - last record in a log file is a
>> SETTLE citation.
>
> That's probably a stuck which should not happen and I'd consider that
> a bug. Can you share your input files?
>
>> In top I noticed both processes used up to 100% CPU, but considerably
>> differed in the use of memory. Relevant lines:
>> PID USER      PR  NI    VIRT    RES    SHR S  %CPU %MEM     TIME+ COMMAND
>> 117710 jan       20   0 5313776 4.642g   9940 R  98.9  3.7  52:47.19
>> mdrun
>> 117686 jan       20   0  547344  24456   8712 R  92.7  0.0  54:48.92 mdrun
>
> That definitely does not look good.
>
> BTW, in contrast with Mark's opinion I do think top is indicative. If
> you wanted to use 8 cores and told mdrun to do so but you get load
> only on one core, you can be (almost) certain that something is not
> right.
>