[yt-users] yt on kraken

[Harinarayan Krishnan]

Hi all,

On Kraken, I always build Yt on the /sw/xt filesystem which is accessible to the
login nodes before I copy the resulting installation to Lustre. So,
I can provide a module load command that points to the version I
create on there.

For example, loading would be as simple as: "module load yt/2.3-login-node"

I have two concerns about this:

First, users would have to be aware that the login-node variants of Yt
will not be available on compute nodes. Also, the login versions will
not have access to MPI.

Second and more importantly, login nodes are typically not meant to be
used as a place where one performs heavy computation and analysis.
They are primarily meant to be responsive and serve as a gateway to
the compute nodes.

I am all for trying to enable faster launches of Yt, but my concern is
that enabling this would open up the possibility of someone launching
large jobs and effectively rendering that login node useless for
anyone else. With a limited number of login nodes on Kraken, crippling
any one of them has the potential of making a lot other users unhappy.

I guess since users can potentially run yt on the login nodes right
now from their own local installation or the one available on Lustre,
this makes the my concern moot. I guess I can enable loading yt using
the xt filesystem and see if Kraken's administrators send me a nasty
email if someone abuses this feature :)

Best regards,
Hari



On Wed, Feb 22, 2012 at 5:13 PM, Patel, Pragneshkumar B
<pragnesh at utk.edu> wrote:
> Hi Nathan,
>
>> Would it be impossible to install the yt module on an NFS filesystem?  I gather that Hari Krishnan reads this user list, maybe he can weigh in on some best practices for using the yt
>> module on Kraken.
>
> It is possible to install the yt on NFS. But you can not access it from compute nodes. You can only use it on login nodes.
>
> Pragnesh
>
>
>
> On Feb 22, 2012, at 4:40 PM, Andrew Cunningham wrote:
>
>> Both the python interpreter and all modules associated with the "module load yt/dev" on kraken live on a luster filesytem.  In general I've found the kraken luster system to have reasonable bandwidth but high to intermittently *extremely high* latency for accessing files.  The yt import accesses lots of files and consequently the import from this yt installation ranges anywhere from annoyingly slow to unusable.  Compute nodes or lightly loaded head nodes are sometimes better, sometimes not it seems.
>>
>> I've install my own build in my home directory on that machine just because of this.  Because the home directory is not luster, an install in this location can load up normally.  I have not had the yt slow-to-start problem since I did this.
>>
>> The quota on the kraken home directory is too small for the yt installer script to download the whole kitchen sink.  However I did manage to install the hdf5, h5py, numpy, scipy, matplotlib, setuptools, cython and pip packages manually for the kraken system python 2.6 and the latest gcc modules.  That did fit in the alloted home space and did finally get all the dependencies yt needs working for me.
>>
>>
>>
>> On Wed, 22 Feb 2012, Britton Smith wrote:
>>
>>> Hi Nathan,
>>> It's most likely not hanging, but working on the import.  The filesystem on
>>> Kraken can be incredibly slow at times.  If you're only waiting a minute or
>>> so, you will probably need to give it a bit longer.
>>> Britton
>>> On Wed, Feb 22, 2012 at 5:15 PM, Nathan Goldbaum <goldbaum at ucolick.org>
>>> wrote:
>>>      Hi all,
>>>
>>>      I'm trying to run yt on a login node on kraken.  All I'm doing
>>>      is invoking python from the command line and importing yt.
>>>       However, python hangs on 'from yt.mods import *'.
>>>
>>>      The python interpreter also takes a very long time (~ a minute
>>>      or so) to start from the command line.
>>>
>>>      This happens using both the yt/dev and yt/2.3 module.
>>>
>>>      Anyone have any ideas?
>>>
>>>      Nathan Goldbaum
>>>      Graduate Student
>>>      Astronomy & Astrophysics, UCSC
>>>      goldbaum at ucolick.org
>>>      http://www.ucolick.org/~goldbaum
>>>
>>>      _______________________________________________
>>>      yt-users mailing list
>>>      yt-users at lists.spacepope.org
>>>      http://lists.spacepope.org/listinfo.cgi/yt-users-spacepope.org
>>>
>> _______________________________________________
>> yt-users mailing list
>> yt-users at lists.spacepope.org
>> http://lists.spacepope.org/listinfo.cgi/yt-users-spacepope.org
>>
>> !DSPAM:10175,4f458b0419704958719638!
>>
>
> _______________________________________________
> yt-users mailing list
> yt-users at lists.spacepope.org
> http://lists.spacepope.org/listinfo.cgi/yt-users-spacepope.org
>
>
> _______________________________________________
> yt-users mailing list
> yt-users at lists.spacepope.org
> http://lists.spacepope.org/listinfo.cgi/yt-users-spacepope.org