[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Condor-users] Mac Users of Condor Anybody out there?



Ian Stokes-Rees wrote:
> I understand Alan Sill from TTU is also interested in knowing how much
> of the grid software stack will run on OS X, in particular server
> packages -- he has successfully compiled and installed most of the OSG
> client-side software to run on OS X.  And as I said, either Clemson or
> RENCI (IIRC) are already running Condor pools which are accessible by
> OSG, however these aren't officially published, so you have to know some
> secret incantation to put into your classad to get them to run there.

In general, OS X machines will work great in a Condor pool.

In OSG, you can not (currently) run a homogeneous Mac OS X pool. The
issue is that the full OSG Compute Element stack is not available for
that platform.

However, you can build a heterogeneous pool. The node you install the CE
on can be a supported flavor of Linux, and then you can have Mac OS X
compute nodes behind that CE node. The biggest issue with this solution
is having jobmanager-fork and jobmanager-condor map to different
architectures.  For example, if a user wants to build a piece of
software for your CE, he is likely to use jobmanager-fork. When he
starts submitting jobs they will fail because the binary he built is for
the wrong architecture. The same problem goes for things like
information providers (your CE will be advertised as a Linux CE). That
being said, Clemson, RENCI and others on OSG have this setup. If you
look in the Condor.pm jobmanager adapter, you will find that by default,
the Condor jobmanager sets requirements on the jobs to stay within the
same architecture, and the same file system domain. You can break out of
this requirement by submitting your own requirements. For example, to
submit a job to run on a Windows machine, add this to the Globus RSL:

( condorsubmit = (universe vanilla)(requirements
'OpSys==\"WINNT51\" && Arch==\"INTEL\"')(should_transfer_files
YES)(when_to_transfer_output ON_EXIT) )

-- 
Mats Rynge
Renaissance Computing Institute <http://www.renci.org>