[Bioclusters] mpiblast on the smallest possible cluster

Lucas Carey bioclusters@bioinformatics.org
Wed, 28 Jan 2004 17:49:03 -0500


MPICH (and LAM) will cycle through the nodes in order if you specify more than are in the machines file (or running your lamd). 
-np 5 on two single-cpu machines would give you this
cpu0: processes 0 2 4
cpu1: processes 1 3
you can see this if you run 'mpirun -v -np $NP' with LAM's mpirun. Not sure about mpich.
this also allows you to run with any -np you want on a single system, though you're computer won't be interactivly usable with -np larger than 2 or 3.
-Lucas

On Wed, Jan 28, 2004 at 07:16:57PM -0300, Fernan Aguero wrote:
> +----[ Lucas Carey <lcarey@odd.bio.sunysb.edu> (28.Jan.2004 11:57):
> |
> | You should be fine using -np 3 instead of -np2. MPICH will
> | start up the third process on the master. Also, try the
> | cvs version if you have any troubles, as it contains a lot
> | of bugfixes. You can get a nightly tarball from:
> | http://mpiblast.sf.net/mpiblast-latest.tar.gz
> | -Lucas
> |
> +----]
> 
> Thanks Lucas,
> 
> that did it! Although I have yet to understand the output of
> mpiblast --debug, I can now see CPU activity increase in the
> master, coincident with the start of a mpiblast process.
> 
> I'm still confused as to what is happening ... -np 3 with 2
> database segments? But perhaps it's just something like:
> process 0 -> master (management)
> process 1 -> node (process execution)
> process 2 -> master, now acting as a node (process execution)
> 
> Is this documented somewhere? Where have I missed
> reading? :) 
> 
> Fernan