[molpro-user] Killing Molpro Jobs using SGE

[Javier Díaz Montes]

Hi,
Rocks is just an easy way to manage a cluster. Mainly, reinstall and  
autoconfigure computer nodes.
I am using SGE, because is a good scheduler and queue manager. I think  
that it is needed, if you want manage queues and a lot of jobs. Moreover,  
A scheduler allows you interoperate with other tecnologies like Grid  
Computing.
When, I run a job, SGE allocate the job in the best nodes acording to the  
scheduling policies. Then, The first node perform as the master of the  
parallel ejecution, and the rest of the nodes will be the slaves. I can  
obtain the list of nodes from SGE and give it to mpirun command (it is not  
magic ;)).

I have not installed mpd, because we just use MPICH or OPENMPI. mpd is a  
process manager for MPICH, but it is not essential. However, It could be a  
solution to kill all mpi processes.

Thanks,
Javi



En Tue, 04 Dec 2007 14:51:54 +0100, Miguel Guilherme Fernandes de Souza  
<guilhermefsmiguel at gmail.com> escribió:

> OK Mr. Javier,
> i saw that you have a solution and a problem together...
> it supposed to be a solution, your rocks cluster, but it is a problem
> because you don´t know what they have done ... and i don´t know why you  
> use
> sge...
> actually you realy don´t need to specify, on your molpro command line the
> --mpirun-machinefile.
>
> this indicates that molpro only instance one process on each node, and as
> other people said, maybe it is normal to have more than one process on  
> each
> node
>
> MPP nodes  nproc
>  compute-0-19.loc    1
>  compute-0-5.loca    1
>  compute-0-22.loc    1
>  compute-0-12.loc    1
>
> As i know, and use, the cluster is built using the mpd command. If this
> rocks cluster do something magic that includes the machines on a cluster  
> for
> you, you don´t need to specify wich machine are you going to include on
> molpro just simply use it.
>
> try to give some command as mpdtrace and see what happen.
> rocks , as i saw on a internet description is a complete "cluster on a  
> CD"
> solution for x86 and IA64 Red Hat Linux COTS clusters. Building a Rocks
> cluster does not require any experience in clustering, yet a cluster
> architect will find a flexible and programmatic way to redesign the  
> entire
> software stack just below the surface (appropriately hidden from the
> majority of users). Although Rocks includes the tools expected from any
> clustering software stack (PBS, Maui, GM support, Ganglia, etc), it is
> unique in its simplicity of installation.
>
> let´s just ask some questions:
> please, try to run mpdtrace command and send-me what it returns ok ?
>
> best regards,
>
>
> On Dec 4, 2007 10:41 AM, Javier Díaz Montes <javier.diaz at uclm.es> wrote:
>
>> Hi,
>> I can not kill the jobs using kill command, because I run the jobs from
>> the Frontend of the cluster. Therefore, I would have to login in each  
>> node
>> to run the kill command.
>> I have a cluster built using Rocks. I have openmpi-1.1-1.
>> I want an instance of molpro for each node, because I have AMD64  
>> machines.
>> I run molpro using the next command:
>>
>>        molprop -n $NSLOTS --mpirun-machinefile $TMPDIR/machines <  
>> infile >
>> outfile
>>
>> How can I specify that I only want a instance per node? Because I think
>> that it consider an instance per node, but I do not know why this create
>> two processes. This is the tree:
>>
>> 27481 ?        R    654:36  \_
>> /home/programs/molpro/molpro-Linux-x86_64-i8-2006.1
>> /molprop_2006_1_i8_p4_mpi.exe
>> 27483 ?        S      0:04      \_
>> /home/programs/molpro/molpro-Linux-x86_64-i8-2006.1
>> /molprop_2006_1_i8_p4_mpi.exe
>>
>> These are the first lines of the Molpro output:
>>
>> ARMCI configured for 6 cluster nodes. Network protocol is 'TCP/IP
>> Sockets'.
>>
>>  MPP nodes  nproc
>>  compute-0-19.loc    1
>>  compute-0-5.loca    1
>>  compute-0-22.loc    1
>>  compute-0-12.loc    1
>> .......
>>
>> Regards,
>> Javi
>>
>>
>