Synaptic noise model and variable timestep solver

[iraikov]

Hello,

I wish to use the fluctuating synaptic conductance model by Destexhe et al 2001 (Gfluct2) in a dentate gyrus network derived from http://www.opensourcebrain.org/projects/dentate .
Because this mechanism uses dt, it seems that it is not compatible with NEURON's variable timestep solver.

However, the cells that use this mechanism are a relatively small number; the majority of the cells in the model are granule cells with about 70-80 compartments, and using the variable timestep solver for them would result in a large performance gain.

My question is then, is it possible to model synaptic noise in a way compatible with variable timestep? If not, is there the technical possibility of using fixed timestep for the cells that use Gfluct2 and variable timestep for the rest? Thanks for all and any help.

[ted]

the majority of the cells in the model are granule cells with about 70-80 compartments, and using the variable timestep solver for them would result in a large performance gain.

If you're looking for faster execution, you may be disappointed. Global variable timestep won't do much unless the cells are nearly synchronous--otherwise, there will always be a few cells that need a smaller dt than the rest of the net does. Local variable timestep seems attractive for such cases, but actual speedup is often modest. Parallelization with MPI usually does much more for run times.

I wish to use the fluctuating synaptic conductance model by Destexhe et al 2001 (Gfluct2) in a dentate gyrus network derived from http://www.opensourcebrain.org/projects/dentate .
Because this mechanism uses dt, it seems that it is not compatible with NEURON's variable timestep solver.

True, but the real problem with such mechanisms is more of a conceptual issue. As written, it draws a new random value at every new solution time. This is tolerable if time steps are fixed, because the effective power spectrum of the noise source remains constant throughout the simulation, but what is the meaning of a series of random values that are sampled at varying intervals? It is as if the power spectrum of the noise source is also varying, and it varies in a way that is not controlled by the designer of the model, but in a way that is completely accidental--depends on whether this or that cell needs shorter dt.

A possible workaround is to implement the noise source in such a way that its output is independent of integration time step. This can be done by taking advantage of NEURON's event delivery system, so that the noise source updates its "output" (current, if it's a current source, or conductance, if it's a fluctuating conductance) at intervals specified by the user.

[iraikov]

Thanks for the reply, your workaround suggestion seems sensible. You are right that I am only speculating about performance based on the results from single cell simulations with variable timestep. The 'normal' state of this network is for the granule cells to be mostly silent, with occasional highly localized inputs, which is why I think it might be sensible to at least try variable timestep. Speaking of the local timestep method, does it work with parallel transfer gap junctions?

If you're looking for faster execution, you may be disappointed. Global variable timestep won't do much unless the cells are nearly synchronous--otherwise, there will always be a few cells that need a smaller dt than the rest of the net does. Local variable timestep seems attractive for such cases, but actual speedup is often modest. Parallelization with MPI usually does much more for run times.
the real problem with such mechanisms is more of a conceptual issue. As written, it draws a new random value at every new solution time. This is tolerable if time steps are fixed, because the effective power spectrum of the noise source remains constant throughout the simulation, but what is the meaning of a series of random values that are sampled at varying intervals? It is as if the power spectrum of the noise source is also varying, and it varies in a way that is not controlled by the designer of the model, but in a way that is completely accidental--depends on whether this or that cell needs shorter dt.

A possible workaround is to implement the noise source in such a way that its output is independent of integration time step. This can be done by taking advantage of NEURON's event delivery system, so that the noise source updates its "output" (current, if it's a current source, or conductance, if it's a fluctuating conductance) at intervals specified by the user.

[ted]

Speaking of the local timestep method, does it work with parallel transfer gap junctions?

If it does, it could only be if the gap connects two sections that belong to the same cell.