Hi,
On 03/10/2016 01:42 PM, David A Case wrote:
> On Thu, Mar 10, 2016, Gerald Monard wrote:
>>
>> Could there be a way to have (decide on?) a list of OS + compilers for
>> which Amber _should_ compile correctly and all tests pass. I wouldn't
>> mind of course for a list of OS+compilers for which it would be "nice"
>> that it works also well.
>
> If you have not yet done so, please visit the Amber wiki page:
>
> http://ambermd.org/pmwiki/pmwiki.php/Main/Amber16Test
>
I've been there already, but there are so many different flavors:
- clang 3.5.1 + gcc 4.8.5, Gentoo Linux
- gcc 4.4.5, linux rhel 5.11
- GCC 4.8.5, CUDA 7.5, No MKL, MPICH 3.1.4, RHEL 7
- gcc 4.8.5 with MKL, linux
- gcc 5.2.0, linux rhel 6.7
- gcc 5.3, mac OSX
- intel 10.0.023, linux
- intel 11.1.056, linux rhel 5.11
- Intel 13.0.1 with MKL, linux
- intel 15.0.3, linux rhel 6.7
- intel 16.0.0, linux rhel 6.7
- intel 16.0.1 with MKL, linux
- pgi 15.4.0, linux rhel 6.7
There was also recent talks about dragonegg, the Amber web site mentions
cygwin too. I am sure that I am missing others.
My problem is the following: if one of these flavors reports a problem
for a part that I wrote and I don't have access to the corresponding
flavor, what should I do?
Example: Dan recently pointed out to me that SEBOMD tests don't pass on
Windows+Cygwin. I don't have access to that setup + it seems to be a
gfortran/cygwin bug. If cygwin is/must be supported, then I should spent
some (a lot of?) efforts on changing my code to make it work.
Thus my question: what must be supported? what could be supported? what
is "who cares"?
As far as I know, there is no mention on the doc or on the web site
about things like "you need gcc>4.0 or intel>10"
> I suspect that we (and our users) get relatively little benefit from all the
> work that goes into supporting the Intel and PGI compilers, especially the
> former, which has a different set of bugs in every release.
>
> It would be nice if some kind soul with some free time could run a pmemd
> benchmark (say jac) comparing Intel vs gnu5 on a somewhat modern chip.
> Also, is cpptraj time-constrained enough to warrant the extra optimizations
> that might come from a proprietary compiler? Do we know anything about clang
> vs gnu for cpptraj?
>
I don't have experience with gcc5, but on my linux cluster, Intel
compiler gives faster results for MD, especially when the MKL is used.
I'm talking here of course about things that are not cuda-enabled (QM
and QM/MM for example).
G.
> I'm willing to be persuaded: almost all my simulations are on GPUs now, where
> there is little need for proprietary compilers.
>
> ...dac
>
>
> _______________________________________________
> AMBER-Developers mailing list
> AMBER-Developers.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber-developers
>
--
____________________________________________________________________________
Prof. Gerald MONARD
SRSMC, Université de Lorraine, CNRS
Boulevard des Aiguillettes B.P. 70239
F-54506 Vandoeuvre-les-Nancy, FRANCE
e-mail : Gerald.Monard.univ-lorraine.fr
tel. : +33 (0)383.684.381
fax : +33 (0)383.684.371
web : http://www.monard.info
____________________________________________________________________________
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Thu Mar 10 2016 - 08:30:04 PST
