Maximum dt

[QCaudron]

Hello,

I'm running a few performance tests using Neuron, with the important factor being dt. I manage to reduce dt (by factors of two only, apparently) from the default 0.025, down to 0.0125 and further, but I can't seem to go any higher than 0.025 ms. Is there some restriction or something I'm missing ? Please note I'm still very new to Neuron !

While on the topic, can I ask what kind of PDE solver methods Neuron uses ?

Thank you very much,
Quentin

[ted]

I'm running a few performance tests using Neuron, with the important factor being dt. I manage to reduce dt (by factors of two only, apparently)

Not quite. A correct statement would be "you can reduce dt by whole number factors".

NEURON's standard run system has two parameters that affect fixed time step integration:
dt, which is the time step used by the numerical integrator, and steps_per_ms, which controls the intervals at which points are plotted. The standard run system's run() procedure calls another procedure setdt(), which forces dt to be compatible with steps_per_ms. setdt() ensures that dt = 1/(N*steps_per_ms) where N = 1, 2, 3 . . .

While on the topic, can I ask what kind of PDE solver methods Neuron uses ?

The second derivative in space is approximated by the central difference method. Each section has its own discretization parameter called nseg which specifies the number of internal nodes, evenly spaced along its length, at which solutions will be computed. You may think of each internal node as being located at the center of a "segment" (NEURON's term for "compartment"). Consequently solutions have second order accuracy at locations that correspond to the centers of segments, and it is advisable to use odd values for nseg so that each section has a node located at 0.5 (midway along its length). It may be helpful to read chapters 4 and 5 of The NEURON Book. If you don't have the book, please be sure to read
"The NEURON Simulation Environment" (the expanded version of our first article in Neural Computation)
and
"NEURON: a Tool for Neuroscientists"
which are available from links at http://www.neuron.yale.edu/neuron/docs

The default integration method is implicit Euler, but Crank-Nicholson and a suite of adaptive order/adaptive time step methods are also available. The latter are described in chapter 4 of The NEURON Book and usage is detailed in the Programmer's Reference.

[QCaudron]

Thanks Ted, that was very helpful !

Quentin