[Smeagol-discuss] MKL_NUM_THREADS

[ambavale sagar]

Dear Sankar,Thanks for your explanation. Is this 'MKL_NUM_THREADS=4' flag is similar to '-npool' flag? Have to use it in command or fix it by 'set' or 'export' command? One that I set by export at present is 'OMP_NUM_THREADS = 1' to avoid bad behavior of MKL libraries.RegardsSagar AmbavaleThe M.S. University of BarodaGujaratIndia Sankar <kesanakd at tcd.ie>
Subject: [Smeagol-discuss] to:Lamjed
To: smeagol-discuss at lists.tchpc.tcd.ie
Message-ID:
    <d62c60210902130854w37ad9646u16d43fccbdc17e3 at mail.gmail.com>
Content-Type: text/plain; charset="iso-8859-1"

>From the above message it is clear that node r19i0n3 crashed because of
overload on it frmm the job. As per my knowledge Smeagol needs large memory
allocation at the above said point. The same thing happened to me as well.
The only solution I can suggest you to use the -npernode flag instead of -np
flag in your submission script. Before all, you need to check the better
value for this npernode flag according to your system. My suggestion is
first to submit the smeagol job and execute 'qutil -u #Job No#'. You will
see all the node info . Then ssh to one of the running nodes and execute top
command. There you can see how much memory smegol is looking for before
crashing. Depending on that you can specify the -npernode flag.

For Eg: In my case when I do the 'top' on the specified node, the memory
smeagol needed was 7GB before crashing. Since each node on stokes hold 16GB
on 8 processors , I have to use only two processors per node (so that each
processor will get access to 8GB). So I'll add '-npernode 2' in my
submission script .
This is why when you run smeagol in seriel you won't get errors. In that
case each processor will get access to 16GB  and you are using only the
memory required for smeagol run. It will be fine if your sytem need more
than 8GB dynamic memory for smeagol run, but if your system needs less than
8GB it is advisable to use this flag to use the resorces properly. You can
also change the MKL_NUM_THREADS value in order to group the processors per
node. In my cas I use 'MKL_NUM_THREADS=4' and '-npernode 2'
I hope this explanation will help you. You are welcome to ask any further
question.

Sankar




      Add more friends to your messenger and enjoy! Go to http://messenger.yahoo.com/invite/
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.tchpc.tcd.ie/pipermail/smeagol-discuss/attachments/20090214/e95935fa/attachment.html