Too many connections to MUSIC output proxies created when NEST runs MPI-parallel

[heplesser]

The following code should connect each parrot neuron exactly once to the MUSIC output proxy:

/nrns /parrot_neuron 2 Create def
/out /music_event_out_proxy << /port_name (out_spikes) >> Create def
nrns out Connect

Running the code on a single process, this works fine, and the following code

Rank 0 eq { (                         Src Tgt Prt Rcpt) = } if
(Rank: ) Rank cvs join ( --> Connections: ) join
<< >> GetConnections GetStatus { [[ /source /target /port /receptor ]] get } Map pcvs join =

prints this output:

                         Src Tgt Prt Rcpt
Rank: 0 --> Connections: [[1 3 0 0] [2 3 0 0]]

This is the same output we get if out were a spike recorder.

Running on two MPI processes, one would expect that each rank has one connection, from the neuron it holds to the local device. But we get two connections on each rank:

                         Src Tgt Prt Rcpt
Rank: 0 --> Connections: [[2 3 0 0] [2 3 1 0]]
Rank: 1 --> Connections: [[1 3 0 0] [1 3 1 0]]

If out is a spike recorder, we get one connection per rank as expected.

The consequence of this is that spikes are sent "double up". The example here leaves out the setting of music_channel to keep the reproducer minimal; it only affects the value of rport.

This bug must have existed for a while and most likely affects all MUSIC out proxies; I am not sure about MUSIC in proxies.

[heplesser]

I believe the problem is located in ConnectionManager::connection_required(), specifically at

nest-simulator/nestkernel/connection_manager.cpp

Lines 1244 to 1253 in 470e1a4

	if ( target->local_receiver() )
	{
	// Connections to nodes with one node per process (MUSIC proxies
	// or similar devices) have to be established by the thread of the
	// target if the source is on the local process even though the
	// source may be a proxy on tid.
	if ( target->one_node_per_process() and source->has_proxies() )
	{
	return CONNECT_TO_DEVICE;
	}

As a brute-force solution, I have experimented with the following fix:

  if ( target->local_receiver() )
  {
    if ( target->one_node_per_process() )
    {
      if ( kernel().node_manager.is_local_node( source ))
      {
        return CONNECT_TO_DEVICE;
      }
      else
      {
      return NO_CONNECTION;
      }
    }

It seems to solve the problem above, but I have not checked for side effects.

[heplesser]

@mdjurfeldt Just an alert about this bug I found, in case MUSIC users contact you about it.

[jougs]

Your fix looks reasonable to me and I can't think of a bad side-effect. Could you please send a PR? Thanks!

[github-actions]

Issue automatically marked stale!