BUG: MPI simulations break for `dt` of certain size

[ntolley]

As the title states, MPI simulations fail under some specific conditions. Oddly it only breaks when dt is too large

I found dt=0.1 breaks, but dt=0.05 passes for the code below

Also it may or may not be related to #662 but this issue only comes up for the very particular case of adding a poisson drive with cell_specific=False, other drives do not evoke this error.

net = jones_2009_model()
weights_ampa_noise = {'L2_basket': 0.01, 'L2_pyramidal': 0.002,
                      'L5_pyramidal': 0.02}
net.add_poisson_drive('noise_global', rate_constant=2.0, location='distal',
                       weights_ampa=weights_ampa_noise, space_constant=100,
                       n_drive_cells=1, cell_specific=False)
with MPIBackend(n_procs=2):
    dpl = simulate_dipole(net, tstop=100)

A quick search to the issue suggests that we should put p.parallel.set_maxstep(10) somewhere to resolve this? https://github.com/neuronsimulator/nrn/issues/933

[rythorpe]

We actually do call ParallelContext.set_maxstep() on each rank here. (Note that ParallelContext.set_maxstep() will set the max step interval between synchronizations to the minimum netcon delay, not necessarily to the upper maximum of 10 entered as an arg in this function.)

I suspect that you've actually created a scenario that is incompatible with integration across parallel machines: your space_constant is set to a large value and thus gives you a very small delay between cells that cannot be resolved between parallel machines with a large dt. See Michael Hines's comment here. @nrnhines is there a way around this, or is the solution here to put bounds on our user-accessible variables governing delay and integration time step?