[MPICH] Master tied to one node
Ralph M. Butler
rbutler at mtsu.edu
Wed Jul 6 10:10:03 CDT 2005
Hi Alok:
I will make some assumtions about how you are testing:
1. mpich2 (instead of mpich1)
2. on unix/linux
3. mpd as the process manager
4. mpiexec for running jobs
The help msg from mpiexec (mpiexec -h or mpiexec --help) offers
a couple of options depending on your needs.
For example, if you just want to get rank 0 off the local node where
you are running mpiexec, you can use the -1 option (that's a one and not
a lower case l).
Using a ring of three mpd's on my cluster nodes named bp400, bp401, and
bp402, I tried this:
(bp400:95) % mpiexec.py -n 1 hostname
bp400
(bp400:96) % mpiexec.py -1 -n 1 hostname
bp401
However, if you want to run rank 0 on a particular host, the help
message shows us this example:
mpiexec -n 1 -host foo master : -n 4 -host mysmp slave
So, I decided to try forcing it to bp402:
(bp400:97) % mpiexec.py -n 1 -host bp402 hostname
bp402
I hope you find this helpful.
--ralph
> Date: Wed, 06 Jul 2005 14:44:54 +0530
> From: Alok G Singh <alokgs at cdacb.ernet.in>
> To: mpich-discuss at mcs.anl.gov
> Subject: [MPICH] Master tied to one node
>
>
> Is it possible for the master task (taskid 0) to be tied to a specified (in some way, IP address, hostname) processor ? From some preliminary experiments that I did, the master always runs on the node where the job is fired (the submit node is also part of the cluster). Is this always the case or is it just coincidence ?
>
> What is required is to make the master process run only a special type of node (one which has local storage, in this case). Any help would be much appreciated.
>
> TIA,
> Alok
More information about the mpich-discuss
mailing list