[gmx-users] Gromacs 2016.4 - the Intel compiler and 'make check'

[Mark Dixon]

Thanks - good to know that there is no clear winner between GCC and Intel 
with GROMACS performance.

As I'm building the software on behalf of a number of other people, and so 
don't have a typical simulation to do speed tests with, I can happily pick 
the lower-risk option and switch to GCC.

Regards,

Mark

On Thu, 26 Oct 2017, Mark Abraham wrote:

> Hi,
>
> It might, but Intel also works on the gcc and llvm compilers and their core
> business is hardware, not selling the compiler. Our experience on non-Phi
> is that performance is often quite close, but your actual hardware and
> simulation are probably also going to affect which compiler implementation
> works fastest for you.
>
> Mark
>
> On Thu, 26 Oct 2017 17:10 Mark Dixon <m.c.dixon at leeds.ac.uk> wrote:
>
>> Hi Mark,
>>
>> Many thanks for the reply.
>>
>> Am I going against the flow by using the Intel compiler with GROMACS? I've
>> been using it so far because of - the potentially foolhardy idea - that it
>> might generate a faster executable than GCC on modern Intel processors.
>>
>> Best,
>>
>> Mark
>>
>> On Thu, 26 Oct 2017, Mark Abraham wrote:
>>
>>> Hi,
>>>
>>> Thanks for the report - we should look into that combination. It's highly
>>> likely that there's some minor issue that different optimization
>> capability
>>> is creating or exposing. The functionality covered by that test is only
>>> used for a few analysis tools, and the fact that gcc passes fine suggests
>>> you should be confident in the code and icc.
>>>
>>> Mark
>>>
>>> On Thu, 26 Oct 2017 13:56 Mark Dixon <m.c.dixon at leeds.ac.uk> wrote:
>>>
>>>> Hi there,
>>>>
>>>> Is there a recommended compiler for GROMACS, please?
>>>>
>>>> I'm trying to validate my install on a CentOS 7.4 Intel Broadwell system
>>>> by running the tests shipped in the GROMACS source tar ball (and the
>>>> separate regression tests).
>>>>
>>>> If I use GCC (4.8.5 or 7.2.0), everything passes but, if I use the Intel
>>>> compilers (tested 16.0.2 and 17.0.1), it keeps failing on
>>>> CorrelationsTest. I've tried pruning it down to the minimum build:
>>>>
>>>>    cmake ../gromacs-2016.4 -DGMX_BUILD_OWN_FFTW=ON
>>>>    make -j12
>>>>    make check
>>>>
>>>> And I see the following:
>>>>
>>>> 16/27 Test #16: CorrelationsTest .................***Exception:
>> Numerical
>>>> 0.25 sec
>>>> [==========] Running 20 tests from 2 test cases.
>>>> [----------] Global test environment set-up.
>>>> [----------] 10 tests from AutocorrTest
>>>> [ RUN      ] AutocorrTest.EacNormal
>>>> [       OK ] AutocorrTest.EacNormal (39 ms)
>>>> [ RUN      ] AutocorrTest.EacNoNormalize
>>>> [       OK ] AutocorrTest.EacNoNormalize (73 ms)
>>>> [ RUN      ] AutocorrTest.EacCos
>>>> [       OK ] AutocorrTest.EacCos (53 ms)
>>>> [ RUN      ] AutocorrTest.EacVector
>>>> [       OK ] AutocorrTest.EacVector (66 ms)
>>>> [ RUN      ] AutocorrTest.EacRcross
>>>>
>>>> I delved into the build directory, hoping to find a log with a bit more
>>>> information and found the string 'Floating point exception', but nothing
>>>> more.
>>>>
>>>> Any advice you can give would be appreciated!
>>>>
>>>> Thanks,
>>>>
>>>> Mark