[mpich-discuss] SMP and mpd.hosts

[Torquil Macdonald Sørensen]

Torquil Macdonald Sørensen wrote:
> Torquil Macdonald Sørensen wrote:
>> Hi all,
>>
>> I would like to ask for some help concerning the use of MPICH2 1.1 on 
>> a dual-core machine with Linux. My mpich2 innstallation uses the 
>> nemesis communication device, which I thought was able to utilise both 
>> cores, in addition to communicate through the network.
>>
>>  From before (maybe from mpich1?) I had a mpd.hosts file with the 
>> following syntax:
>>
>> dagda.uio.no:2
>> #styx.uio.no:2
>>
>> where the :2 was to indicate the number of CPUS. styx.uio.no is 
>> commented out because this time I want two processes to run on 
>> dagda.uio.no. But I am no longer sure of this syntax since I cannot 
>> find it in the documentation. I was hoping that the following would work:
>>
>> $ mpdboot -n 2
>> $ mpiexec -n 2 ./Executable
>>
>> However, mpdboot complains:
>>
>> "totalnum=2  numhosts=1
>> there are not enough hosts on which to start all processes"
>>
>> What it the common method to run a program so that it will be able to 
>> distribute itself across several cpu-cores in one machine, and also 
>> across the network, if I e.g. decomment styx.uio.no and use -n 4 in 
>> the above commands?
>>
>> Best regards
>> Torquil Sørensen
>>
> 
> It seems to be possible to get two processes by using "mpdboot -1 -n 2". 
>   I guess that means that it starts more than one mpd. In that case 
> "mpiexec -n prog" works. So I guess I was suffering a limitation of max 
> one process per mpd?
> 
> -Torquil
> 
> 

Ah now I see my mistake. In mpdboot -n 2, n stands for number of hosts, 
but in mpiexec it is number of processses... It works with

mpdboot -n 2
mpiexec -n 4 prog

when I have uncommented styx in the mpd.hosts. And I would run "mpdboot 
-n 1" together with "mpiexec -n 2 prog" to use both cores on only one 
machine.

Sorry for the noise... :-)

Best regards
Torquil Sørensen