gap junctions in parallelized models

[shyam_u2]

Dear Ted,

Can you give some suggestions on implementing gap junctions in parallel NEURON. I have implemented it in serial mode. I want to convert it to parallel mode. Do you think that it is a good idea to create all cells(involving gap junctions) on a single host ? Or do you think that is there an alternative strategy.

Thanks,
Shyam

[ted]

Networks that involve gap junctions can be parallelized by using the "parallel transfer" methods class--see http://www.neuron.yale.edu/neuron/stati ... elTransfer. At least two published models have used these methods--see these papers, which are available from http://www.neuron.yale.edu/neuron/nrnpubs:

Hines, M.L., Eichner, H. and Schuermann, F.
Neuron splitting in compute-bound parallel network simulations enables runtime scaling
with twice as many processors.
Journal of Computational Neuroscience 25:203-210, 2008
ModelDB accession # 97917

and

Hines, M.L., Markram, H. and Schuermann, F.
Fully implicit parallel simulation of single neurons.
Journal of Computational Neuroscience 25:439-448, 2008
ModelDB accession # 97985

But I wonder if there is a more explicit guide for how to proceed.

[shyam_u2]

Dear Ted,

I see that Traub et al model (2005) was written serially and it was parallelized by parallel transfer methods. ( Correct me if I am wrong, because I am a beginner to Parallel NEURON). I have a big cerebellar model which is parallelized with parallel context class. I am currently working on basket cells and Gap junctions and this is the first time gap junctions are introduced in the model. So in order to parallelize the gap junctions of basket cells, do i need to modify the code of entire cerebellar model ? I am planning to create all cells(involving gap junctions) on a single host since the number of cells involving gap junctions are not that high. Do you think that this strategy is worth ?

Thanks,

Shyam

[ted]

I see that Traub et al model (2005) . . . was parallelized by parallel transfer methods.

If it involves gap junctions between cells on different hosts, parallel transfer would have been used.

in order to parallelize the gap junctions of basket cells, do i need to modify the code of entire cerebellar model ?

Presumably your program is organized something like this:
create cells
connect cells
instrumentation
simulation control
The necessary changes would of course be in the "connect cells" code, and would consist of adding one or more procedures that set up the gap junctions.

I am planning to create all cells(involving gap junctions) on a single host since the number of cells involving gap junctions are not that high.

Probably a bad idea because it would disturb load balance. Parallelization is most efficient if all hosts can be kept busy all the time. If cells connected by gap junctions tend to fire together, and they're all on one host, then during epochs when the gap-connected subnet is silent, that host will have less to do than the other hosts. Its idleness would mean a loss of up to 1/nhosts of the computational power of the entire parallel machine. During epochs when the gap-connected subnet is active, that host will have more to do than the other nhosts-1 machines, which means a loss of up to (nhosts-1)/nhosts of the power of the parallel machine.