Single cell in parallel

General issues of interest both for network and
individual cell parallelization.

Moderator: hines

Post Reply
bll5z6
Posts: 29
Joined: Thu May 26, 2016 10:27 am

Single cell in parallel

Post by bll5z6 »

I have a single cell model with 2500 or so compartments that I would like to speed up. It is implemented using BMTK's bionet, which runs NEURON on the back end. I was reading some older posts and understand it's possible to parallelize a single cell, but that development time is substantial. Is that still the case? Is there a way to speed up a complex single cell model with less development overhead?
ted
Site Admin
Posts: 6286
Joined: Wed May 18, 2005 4:50 pm
Location: Yale University School of Medicine
Contact:

Re: Single cell in parallel

Post by ted »

2500 compartments isn't particularly large.

Easiest parallel simulation is via multithreaded execution, which usually doesn't require any change to your model specification code. Some NMODL mechanisms may need slight revision to make them threadsafe; there are a few mechanisms that are inherently not threadsafe. Because of multithreaded execution's overhead, speedup will be less than the number of cores devoted to the task; the practical limit beyond which there is no gain is about 15, but most models max out at around 4 or 5.

To learn more, see this exercise from the 2018 NEURON Summer Course https://neuron.yale.edu/neuron/docs/mul ... lelization. That particular exercise does not discuss error tolerance scale factors, which can be important if one or more of your model's state variables are very small compared to others (e.g. intracellular calcium concentration).

If you plan to run many simulations, e.g. for optimization or to explore parameter space, consider using NEURON's bulletin-board style parallelization, which is very efficient as long as all processors have enough work to keep busy (i.e. number of runs is larger than the number of processors).
Post Reply