[Mauiusers] Problem with Torque/Maui

S Ranjan sranjan at ipr.res.in
Wed Jan 24 18:15:53 MST 2007


Garrick Staples wrote:

>On Thu, Jan 25, 2007 at 05:37:56AM +0530, S Ranjan alleged:
>  
>
>>Garrick Staples wrote:
>>
>>    
>>
>>>On Wed, Jan 24, 2007 at 08:51:11AM +0530, S Ranjan alleged:
>>>
>>>
>>>      
>>>
>>>>Hi
>>>>
>>>>I have torque pbs_server running on the headnode, which is also the 
>>>>submit host.  There are 32 other compute nodes, mentioned in 
>>>>/var/spool/torque/server_priv/nodes file.  There is a single queue at 
>>>>present.  Sometimes, mpi jobs requesting for 28/30 nodes, land up 
>>>>running on the head node, though the head node is not a compute node at 
>>>>all.  netstat -anp shows several sockets being openend for the job, and 
>>>>eventually the head node hangs up. 
>>>>
>>>>Appreciate any help/suggestion on this.
>>>>  
>>>>
>>>>        
>>>>
>>>Which MPI?  MPICH?  I'd guess mpirun is using the default machinefile
>>>that is created when mpich is built, and not the hostlist provided by
>>>the PBS job.
>>>
>>>Run mpirun with "-machinefile $PBS_NODEFILE" or use OSC's mpiexec
>>>instead of mpirun: http://www.osc.edu/~pw/mpiexec/
>>>
>>>_______________________________________________
>>>mauiusers mailing list
>>>mauiusers at supercluster.org
>>>http://www.supercluster.org/mailman/listinfo/mauiusers
>>>_____________________________________________________________________
>>>
>>>The mail server at Institute for Plasma Research has scanned this
>>>email for Virus using ClamAV 0.88.4
>>>_____________________________________________________________________
>>>
>>>
>>>
>>>
>>>      
>>>
>>We are using Intel mpi 2.0.  We are using mpiexec -n 28 ......      
>>inside the pbs script.
>>However, for mpdboot (executable in the mpi 2.0 binary dir), we are 
>>running it before running the pbs script. The exact syntax being used is
>>
>>mpdboot -n 32 -f mpd.hosts --rsh=ssh -v
>>
>>mpd.hosts file, residing in the user's home directory,  contains the 
>>names of the 32 compute nodes (excluding the head node).
>>    
>>
>
>There is your problem, you want to use the list of nodes assigned to
>your job.  So you'll want something like this:
>  np=$(wc -l < $PBS_NODEFILE)
>  mpdboot -n $np -f $PBS_NODEFILE --rsh=ssh -v
>
>But I still recommend using OSC's mpiexec instead.
>
>  
>
>------------------------------------------------------------------------
>
>_______________________________________________
>mauiusers mailing list
>mauiusers at supercluster.org
>http://www.supercluster.org/mailman/listinfo/mauiusers
>  
>
Actually for us, mpd.hosts and $PBS_NODEFILE contain the same set of 
machines - the 32 compute nodes.  However, since our headnode is also 
the submit host, mpdboot actually needs to be started with -n33, instead 
of -n 32. 

This is because,  mpdboot starts on the headnode anyway, and using -n 
32, actually decreases the compute nodes by one, as the mpdboot counting 
goes as <headnode> + 31  <compute nodes>


 I'll download OSc's mpiexec and try the -machinefile option - I presume 
it is compatible with INTEL mpi 2.0's mpiexec, because the dynamic 
libraries loaded would be from INTEL's library.

Thanks

Sutapa




More information about the mauiusers mailing list