Cell splitting in neural networks extends strong scaling (Hines et al. 2008)

 Download zip file   Auto-launch 
Help downloading and running models
Neuron tree topology equations can be split into two subtrees and solved on different processors with no change in accuracy, stability, or computational effort; communication costs involve only sending and receiving two double precision values by each subtree at each time step. Application of the cell splitting method to two published network models exhibits good runtime scaling on twice as many processors as could be effectively used with whole-cell balancing.
1 . Hines ML, Eichner H, Schürmann F (2008) Neuron splitting in compute-bound parallel network simulations enables runtime scaling with twice as many processors. J Comput Neurosci 25:203-10 [PubMed]
Citations  Citation Browser
Model Information (Click on a link to find other models with that property)
Model Type: Realistic Network;
Brain Region(s)/Organism: Generic;
Cell Type(s):
Gap Junctions:
Simulation Environment: NEURON;
Model Concept(s): Methods;
Implementer(s): Hines, Michael [Michael.Hines at Yale.edu];
balcomp.hoc *
cell_templates.hoc *
clear.hoc *
finit.hoc *
fortmap.hoc *
gidcell.ses *
mosinit_orig.hoc *
onecell.ses *
prcellstate.hoc *
printcon.hoc *
spkplt.hoc *
vclampg.hoc *
vcompclamp.hoc *
vcompsim.hoc *
NEURON version of Traub et al J Neurophysiol. 2005 Apr;93(4):1829-30.
Single-column thalamocortical network model exhibiting gamma oscillations
sleep spindles, and epileptogenic bursts.

See: http://senselab.med.yale.edu/senselab/modeldb/ShowModel.asp?model=45539

Prepare for running with
nrnivmod mod

See the comparison of each cell type with the fortran output with
nrngui onecell.hoc

Selecting the "Exact traub style ri" which forces the NEURON connection
coefficients to be exactly the same as computed by the Traub's equivalent
circuit style from the FORTRAN demonstrates that cell and channel properties
are reliably represented in this NEURON translation of the FORTRAN code.


Since the NEURON and Fortran models do not produce quantitatively
identical results there is always some question as to whether simulation
differences are due to substantive parameter translation errors or
can be attributed to different numerical methods.
It must be realized that our experience has
been that every test into a new runtime domain has exhibited
discprepancies that were ultimately resolved by fixing
translation errors.

And also that our comparison
tests are only between NEURON and an already significantly modified
FORTRAN code. The bulk of the FORTRAN modifications are toward more
generic FORTRAN syntax to allow the original ifc compatible FORTRAN to
run under g77. Most of the execution places where the g77 version
differs from the ifc version are straight forward transformations of
bulk array assignment into equivalent elementwise assignment via do loops.
Did we get them all? Did we assign over ALL the elements in each array?
We did manually review all ifc to g77 editing changes but
a few cases involved our judgement with regard to whether there was a bug
in the original ifc fortran version.  The modified FORTRAN used for the
NEURON comparisons is available from this model=45539 page. As is, the
g77 FORTRAN model can only be run as 14 processes, one for each cell type
and a full model run takes 20 hours or so. Simplifying to 1/10 the number
of cells gives a model that takes approximate 1.5 hours for 100 ms of
simulation time. Our last network bug, based on significant spike raster
plot discrepancies, was found using a 10 ms run.

We consider the translation of the 14 individual cell types to be quite
reliable based on the quantitative similarity of the g77 and NEURON
isolated 100 ms cell trajectories at the spike detection location for
0 and 0.3 nA constant current stimulation into the soma. Note that
quantitative similarity demands compartment coupling of exactly the
same values used by the FORTRAN version algorithms (imitated in NEURON
using the "traub_exact()" algorithm where some branch points had the
form of "wye" equivalent circuits, some had the "delta" form, and all had
a different view of how resistance from child to parent should be computed.)

Network topology and chemical synapse parameter reliability is limited to the
diagnostic power of our specific tests. For quantitative comparisons we
printed the precise FORTRAN network topology to files and used that information
to define the NEURON network connections. For 10 ms with a 1/10 size network
we focused on quantitative similarity of the spike raster plots. The FORTRAN
version has a spike resolution time of 0.1 ms and all synaptic conductance
trajectories are step functions with that resolution (the underying
dt is 50 times smaller, dt = 0.002). We prepared a special version of the
NEURON executable to force spike threshold detection on 0.1 ms boundaries
to allow convenient comparison of spike rasters. For the first 10 ms
we judged whether spike discrepancies were due to FORTRAN-NEURON spikes
straddling the 0.1 ms boundaries or whether the discrepancy was likely to
be due to a topology or synaptic parameter error. The judgement was based
on the details of the voltage trajectory at the spike detector compartment.
We believe that careful analysis of the first
10 ms of the  100 ms spike raster overlap plot for the
FORTRAN (fat red marks) and NEURON (thin black marks) in combination with
the spike trajectory sensitivity of suppyrRS cells with respect to number
of spikes in their burst after the first spike will convince that we
have gone as far as possible with quantitative spike location similarity
as a diagnostic technique. Further diagnostics will likely have to be
based on specific questions in regard to qualitative discrepancies and
a focus on the NEURON model itself as the tool for exploration in terms
of certain properties added or subtracted from successive runs. Unfortunately
that can only be done in response to a specific suspicion on the part of the
user. Enumerated below are the known discrepancies between the representations
of the Traub model in FORTRAN and NEURON:

The NEURON nmda saturation is turned off. See the NMDA_saturation_fact in
the FORTRAN groucho.f file and the nrntraub/mod/traub_nmda.mod file.
Warning: NEURON will not mimic the FORTRAN
merely by setting the factor to 80.

The groucho.f axon_refrac_time is normally set to 1.5. Our quantitative
tests temporarily set this parameter to 0.5. The NEURON spike detection
algorithm defines a spike as a positive going transition past the
trigger value.

Enumerated below are those major components (of which we
are aware) that are in the model but have not been tested in terms of their
quantitative equivalence to FORTRAN:
gap junctions.
long term nmda properties and effects.
ectopic spikes
random current stimulation

The bottom line: The spike rasters for a
full g77 FORTRAN run (gap junctions and current stimulation present) and
a full NEURON run with its own independent random variables
(no "traub_exact" connection coefficients, dt = 0.025 (ms),
secondorder = 2, and spike detection with dt resolution) is presented.