[MPICH] Changing the comm size at runtime

[Patrick Gräbel]

Here is the code, posted on a pastebin: http://rafb.net/p/5L20tD24.html

The head comment explains how to to start this "experimental" code.

Meanwhile I succeeded in creating a "huge" intracomm object from
incoming intercomm objects _without_ using MPI_Intercomm_create. The
master's code looks like this:

-------------------------------------------
intras[0] = MPI::COMM_WORLD.Dup();
for(int i = 0; i < num_slaves; i++)
{
  inters[i] = intras[i].Accept(port,MPI::INFO_NULL,0);
  intras[i + 1] = inters[i].Merge(false);
  intras[i + 1].Send(&num_slaves,1,MPI::INT,i + 1,0);

}
-------------------------------------------

The master uses the intracomm of the last merge to accept the next
intercomm object. For each accepted slave the master reports the number
of slaves being awaited. The slaves themselves do something like this:

-------------------------------------------
inters[0] = world.Connect(port.c_str(),MPI::INFO_NULL,0);
intras[0] = inters[0].Merge(true);
// receive the actual number of slaves being awaited
intras[0].Recv(&num_slaves,1,MPI::INT,0,0);
// lower ranks have to connect all higher ranks
for(int i = 0; i < num_slaves - intras[0].Get_rank(); i++)
{
  inters[i + 1] = intras[i].Connect(port.c_str(),MPI::INFO_NULL,0);
  intras[i + 1] = inters[i + 1].Merge(false);
}
-------------------------------------------

E.g. an Allgather call works over the arbitrary number of slaves
including the master. After disconnection of all slaves the master is
able to accept a new set of slaves.

I wonder if this solution is equivalent to your suggestion...

Greetings
Patrick

Rajeev Thakur schrieb:
> Can you send us the example code?
> 
> Rajeev 
>