So you know this is something that we are working on for the next major
release of Open MPI (v 1.3). More details on some of the discussion can
be found here:
Torje Henriksen wrote:
> Specifying nodes several times in the hostfile or with the --host
> parameter seems to just add up the number of slots availible for the
> given node. It doesn't seem to affect the mapping of the ranks. I think
> this is due to how the hostfile is read into the structure that holds this
> information in the source code.
> Adding the host several times to the hostfile was the first thing I tried,
> and I've also gotten that suggestion from others, so it might seem that it
> would make sense to make it work that way.
> I've hacked the source to be able to take a ranks-parameter in the
> hostfile like this:
> node0 ranks=0,1,3
> node1 ranks=2,4,5
> so I guess it's not a problem any more, but I would love to know if there
> is a way of doing it without changing the source code.
> You're very right about the unix scripting part. It makes sense to create
> the hostfile this way.
> On Mon, 1 Oct 2007, Christian Bell wrote:
>> How about a hostfile such as
>> % cat -n ~/tmp/hostfile
>> 1 node0
>> 2 node0
>> 3 node1
>> 4 node0
>> 5 node1
>> 6 node1
>> Looks like the function to express the mapping is not anything simple. If it's
>> an expressible function but too complicated for open mpi, you'll have to make
>> your own script to generate the function. This shouldn't be hard to do with
>> any standard unix scripting.
>> . . christian
>> On Mon, 01 Oct 2007, Torje Henriksen wrote:
>>> Oh man, sorry about that, and thanks for the fast response.
>>> Let me try again, please :)
>>> I want to manually specify what ranks should run on what node.
>>> Here is an example of a mapping that I can't seem to be able to do, since
>>> it isn't a round-robin type of mapping.
>>> hosts ranks
>>> node0 0,1,3
>>> node1 2,4,5
>>> No matter what I do, I either get
>>> node0: 0,1,2
>>> node1: 3,4,5
>>> node0: 0,2,4
>>> node1: 1,3,5
>>> Hope I got it right this time, and thank you again.
>>> On Mon, 1 Oct 2007, jody wrote:
>>>>> hosts ranks
>>>>> node0 1,2,4
>>>>> node1 3,4,6
>>>> I guess there must be a typo:
>>>> You can't assign one rank (4) to two nodes....
>>>> And ranks start from 0 not from 1.
>>>> Check this site,
>>>> there might be some inforegarding your problem.
>>>> users mailing list
>>> users mailing list
>> (QLogic Host Solutions Group, formerly Pathscale)
>> users mailing list
> users mailing list