Regarding UNIVERSE_SIZE

[William Gropp]

At 04:28 PM 5/25/2005, Prakash Velayutham wrote:
>Hi,
>
>How is the implementation of UNIVERSE_SIZE in mpich2? Does anyone know. I 
>would like to know if it can be varied during a MPI run by the scheduler? 
>I am looking at it from MPI_Comm_spawn point of view. If I can modify this 
>variable from the scheduler's end when more processors are available and 
>check for this variable from inside the application, then I can 
>effectively use the spawn to get newer nodes from inside of scheduler 
>environment. I understand it is not that straight forward, but it is 
>hackable, right?

Section 5.5.1 of the MPI-2 standard says:

MPI UNIVERSE SIZE is initialized in MPI INIT and is not changed by MPI.

This was done to avoid the race conditions implicit in allowing 
MPI_UNIVERSE_SIZE to change.  The standard goes on to say that if the 
application wants more up-to-date information, it must communicate with the 
process management system.

Note that MPI_UNIVERSE_SIZE is not a maximum, so it doesn't limit the 
ability of the app to create more processes if the process management 
system allows that.

Bill

>Any words from developers, please?
>
>Regards,
>Prakash

William Gropp
http://www.mcs.anl.gov/~gropp