<div dir="ltr">Hi,<br><br><div class="gmail_quote"><div dir="ltr">On Thu, Aug 6, 2015 at 1:21 PM Kutzner, Carsten <<a href="mailto:ckutzne@gwdg.de">ckutzne@gwdg.de</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">Hi,<br>
<br>
I wanted to quickly report back on tune_pme for 5.1, what are the issues<br>
so we could discuss on how best to proceed.<br></blockquote><div><br></div><div>Many thanks!</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
In principle, tune_pme works, however as a user one may have to set extra<br>
options so that actually tuning can be done.<br>
<br>
The issues are:<br>
<br>
a) A valid mdrun executable needs to be specified first, see:<br>
<a href="https://gerrit.gromacs.org/#/c/4771/" rel="noreferrer" target="_blank">https://gerrit.gromacs.org/#/c/4771/</a></blockquote><div><br></div><div>OK. I think we need to make this a first-class property of tune_pme, and have proposed such at <a href="https://gerrit.gromacs.org/4963">https://gerrit.gromacs.org/4963</a></div><div><br></div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
b) On GPU nodes, mdrun typically segfaults, due to timers not being reset<br>
at a neighborsearching step, see:<br>
and <a href="http://redmine.gromacs.org/issues/1781" rel="noreferrer" target="_blank">http://redmine.gromacs.org/issues/1781</a></blockquote><div><br></div><div>Fixed at <a href="https://gerrit.gromacs.org/4964">https://gerrit.gromacs.org/4964</a></div><div><br></div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
c) One needs to specify -ntomp somevalue, otherwise it is 1 by default<br>
and mdrun refuses to run.<br></blockquote><div><br></div><div>Can you provide an example here? I'm not sure what's the problem, to assign blame or to fix! :-)</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
Issue b) is probably the most important one to address, which could be either<br>
done in mdrun or in tune_pme.<br>
Tune_pme would have to determine from the .tpr what the actual neighborsearching<br>
frequency will be and adjust the counter resetting appropriately.<br></blockquote><div><br></div><div>Not very workable - nstlist varies according to what hardware is available, and one can't really know that without running mdrun. (e.g. you could run gmx tune_pme from a non-CUDA install and it has no idea that mdrun can use a GPU).</div><div><br></div><div>Mark</div></div></div>