[MITgcm-support] build options for Stampede
Holly Dail
hdail at MIT.EDU
Thu Feb 7 16:23:02 EST 2013
Matt -
Sorry for the slow response ... forgot to look for these last night.
And to be doubly disappointing, I can't find the code where I added these modifications. Sorry!
Holly
On Feb 6, 2013, at Feb 6 , 5:58 PM, Matthew Mazloff wrote:
> Hi Holly
>
> Yeah, its only for a few specific I/O routines. Do you have a copy of you mods?
>
> Thanks
> Matt
>
>
> On Feb 6, 2013, at 2:40 PM, Holly Dail <hdail at MIT.EDU> wrote:
>
>> Hi Matt -
>>
>> I don't have any suggestions for _fixing_ the problem, but I can tell you what I did that drastically improved my failure rate when I had a similar issue. Just modify the code to try opening the file multiple times (maybe 3 times?) before giving up. When I faced a similar issue it was so transient that repeating the call was effective. Can't remember if a sleep was required between the attempts ... probably helpful, but will also slow down your code.
>>
>> Holly
>>
>> On Feb 6, 2013, at Feb 6 , 11:54 AM, Matthew Mazloff wrote:
>>
>>> Hi Angela,
>>>
>>> Thanks - but unfortunately stampede is different from these two machines. The good news is it is far faster -- almost 4 times faster. The bad news is that it doesn't seam to like the I/O statements in the mdsio package.
>>>
>>> I find it randomly trips up writing (usually with MDS_WRITE_FIELD) when the file already exists.
>>>
>>> The error I get is usually a simple
>>>
>>> "forrtl: severe (121): Cannot access current working directory for unit 9, file "Unknown"
>>>
>>> Though sometimes it tells me that a file that definitely exists does not exist. Or it says this file cannot be written to.
>>>
>>> For example, the adjoint needs some average files and it writes out some "bar" fields. Yesterday it crashed writing a "bar" file after running fine for about 2 years…
>>>
>>> I also find that its processors do not like to share -- I have had to give each processor its own tiled forcing file because otherwise it randomly crashes…
>>>
>>> However, there have been a few runs that complete just fine….
>>>
>>> TACC consulting do not have any suggestions. I am hoping that with the proper compile flags I can make the run more robust.
>>>
>>> Any suggestions appreciated!
>>> Matt
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>> On Feb 6, 2013, at 5:42 AM, Angela Zalucha <azalucha at seti.org> wrote:
>>>
>>>> Matt: I have successfully used TACC's Lonestar and Ranger machines, but won't be in front of a real computer until tomorrow night, when I can send you the opt files.
>>>>
>>>> Angela
>>>>
>>>> p.s. I can also send you results for scaling tests, and would be interested to know how stampede works out for you. For instance I find that lonestar is a lot faster than ranger. I am actively proposing to XSEDE.
>>>>
>>>> Sent from my Verizon Wireless 4G LTE DROID
>>>>
>>>>
>>>> Matthew Mazloff <mmazloff at ucsd.edu> wrote:
>>>>
>>>> Hello,
>>>>
>>>> Has anyone made a compile options file for TACC's Stampede yet?
>>>>
>>>> http://www.tacc.utexas.edu/user-services/user-guides/stampede-user-guide
>>>>
>>>> I am getting random crashes on this machine due to disk access issues and suspect I may not be using an ideal option file.
>>>>
>>>> Thanks
>>>> Matt
>>>>
>>>>
>>>> _______________________________________________
>>>> MITgcm-support mailing list
>>>> MITgcm-support at mitgcm.org
>>>> http://mitgcm.org/mailman/listinfo/mitgcm-support
>>>
>>> _______________________________________________
>>> MITgcm-support mailing list
>>> MITgcm-support at mitgcm.org
>>> http://mitgcm.org/mailman/listinfo/mitgcm-support
>>
>>
>>
>>
>> Holly Dail
>> Postdoctoral Fellow
>> Harvard University
>> http://fas.harvard.edu/~hdail/
>>
>>
>> _______________________________________________
>> MITgcm-support mailing list
>> MITgcm-support at mitgcm.org
>> http://mitgcm.org/mailman/listinfo/mitgcm-support
>
>
> _______________________________________________
> MITgcm-support mailing list
> MITgcm-support at mitgcm.org
> http://mitgcm.org/mailman/listinfo/mitgcm-support
Holly Dail
Postdoctoral Fellow
Harvard University
http://fas.harvard.edu/~hdail/
More information about the MITgcm-support
mailing list