[MITgcm-devel] monitor output with uneven number of threads

[Martin Losch]

Hi Jean-Michel,

thanks, I could have thought about that myself! That was actually it. I find it quite strange, that the optimisation actually changes the I/O streams. I’ll check the modified build-options file in.

Martin

> On 06 Jun 2016, at 18:31, Jean-Michel Campin <jmc at mit.edu> wrote:
> 
> Hi Martin,
> 
> Did you try without any compiler optimisation (with "-devel" and/or "-ieee") ?
> 
> If this is an issue of compiler optimisation, you could try to
> put mon_init.F & mon_set_iounit.F into the NOOPTFILES list.
> 
> Cheers,
> Jean-Michel
> 
> On Mon, Jun 06, 2016 at 05:27:59PM +0200, Martin Losch wrote:
>> Hi,
>> 
>> I use the cray compiler (cce) with ???impi??? (intel mpi) on our CS 400. I use MPI and multi-threads. I want to (and I think I do) use 13 MPI tasks with 9 threads each. (ntx=3 and nty=3 in eedata, maybe that???s the problem and I should use 9 and 1?)
>> 
>> Martin
>> 
>>> On 06 Jun 2016, at 15:38, Jean-Michel Campin <jmc at mit.edu> wrote:
>>> 
>>> Hi Martin,
>>> 
>>> I tried with 2 MPI procs and 3 threads each and it looks fine
>>> (was with gfortran+openmpi).
>>> Are you simply trying multi-threads but without MPI ?
>>> and which compiler was it ?
>>> 
>>> Cheers,
>>> Jean-Michel
>>> 
>>> On Fri, Jun 03, 2016 at 01:49:23PM -0400, Jean-Michel Campin wrote:
>>>> Hi Martin,
>>>> 
>>>> This should be easy to reproduce. I will check and will let you
>>>> know what I find.
>>>> 
>>>> Cheers,
>>>> Jean-Michel
>>>> 
>>>> On Fri, Jun 03, 2016 at 05:10:13PM +0200, Martin Losch wrote:
>>>>> Hi,
>>>>> 
>>>>> I run the llc90 configuration with 117 tiles (30x30 each). Now I would like to use a hybrid configuration (to avoid MPI_sends/recv as much as possible), and I though it would be neat to run 9 threads, ie. 90x90 chunks on one mpi-task. I increase MAX_NO_THREADS to 9 and everything works fine, I think, at least the IO looks OK.
>>>>> 
>>>>> But now I see that everything that goes to standardMessageUnit (=6) goes into the STDOUT.* files, as expected, except for the monitor output. I checked, the mon_ioUnit is also 6
>>>>> A little bit of poking around suggests, that it works fine with an even number of threads (e.g. 4, 8 ,12), but with an uneven number of threads (9, 13) I get this behavior that the master thread on the master mpi process writes the monitor output not into STDOUT.0000 but in the the stdout of the job-script.
>>>>> I have no idea where to look for the problem. Any pointer are highly appreciated.
>>>>> 
>>>>> Martin
>>>>> 
>>>>> 
>>>>> _______________________________________________
>>>>> MITgcm-devel mailing list
>>>>> MITgcm-devel at mitgcm.org
>>>>> http://mitgcm.org/mailman/listinfo/mitgcm-devel
>>>> 
>>>> _______________________________________________
>>>> MITgcm-devel mailing list
>>>> MITgcm-devel at mitgcm.org
>>>> http://mitgcm.org/mailman/listinfo/mitgcm-devel
>>> 
>>> _______________________________________________
>>> MITgcm-devel mailing list
>>> MITgcm-devel at mitgcm.org
>>> http://mitgcm.org/mailman/listinfo/mitgcm-devel
>> 
>> 
>> _______________________________________________
>> MITgcm-devel mailing list
>> MITgcm-devel at mitgcm.org
>> http://mitgcm.org/mailman/listinfo/mitgcm-devel
> 
> _______________________________________________
> MITgcm-devel mailing list
> MITgcm-devel at mitgcm.org
> http://mitgcm.org/mailman/listinfo/mitgcm-devel