[MITgcm-devel] model crashes when opening scratch units

[Matthew Mazloff]

In my experience 
using 
define TARGET_CRAYXT
always has them write to the same disk.

Your way sounds nice -- I'll check out 
#define SINGLE_DISK_IO

Thanks!
Matt



On Apr 18, 2015, at 9:43 AM, "Menemenlis, Dimitris (329D)" <Dimitris.Menemenlis at jpl.nasa.gov> wrote:

> Same disk or different disk?  I had a similar problem on Pleiades and solved by writing only one scratch file from process 0. Followed by a sleep/barrier, Then reading it from all the others.
> 
> Dimitris Menemenlis
> 818-625-6498
> 
>> On Apr 17, 2015, at 11:06 AM, Matthew Mazloff <mmazloff at ucsd.edu> wrote:
>> 
>> If you have too many tasks they can get out of sync and one moves to the next parameter initialization and overwrites scratch file before the other is ready.
>> 
>> define TARGET_CRAYXT
>> 
>> makes every task write its own scratch files. So it can make for a messy directory, but its safe
>> 
>> 
>> 
>> 
>>> On Apr 17, 2015, at 8:45 AM, Dimitris Menemenlis <dmenemenlis at gmail.com> wrote:
>>> 
>>> What is problem with scratch files?  In case this is relevant, the following flag is available
>>> in CPP_EEOPTIONS.h to reduce number of scratch files on single-disk systems:
>>> 
>>> C--   Flag defined for eeset_parms.F and open_copy_data_file.F
>>> C     to write scratch files from process 0 only
>>> #define SINGLE_DISK_IO
>>> 
>>>> On Apr 17, 2015, at 1:06 AM, Martin Losch <Martin.Losch at awi.de> wrote:
>>>> 
>>>> Hi Matt,
>>>> 
>>>> thanks, that’s the flag I was looking for, it works for me too (except for all the “scratch*” files).
>>>> 
>>>> M.
>>> 
>>> 
>>> _______________________________________________
>>> 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