[MPICH] Changing the comm size at runtime

[Patrick Gräbel]

Quoting Rajeev Thakur <thakur at mcs.anl.gov>:
> This would also work. I think the slaves would need to call Accept in the
> loop, not connect, because they become part of the communicator that the
> master uses to call Accept.
>

That makes sense, because the growing intra communicator does a  
collective accept on a new, single connecting slave. Actually  
"connect" instead of "accept" works also, but I guess it shouldn't! Is  
that a weird effect I should forget as soon as possible? As for me I  
can not explain the conforming behaviour...

Thanks
Patrick