Using a more recent SUNDIALS version

[sanjayankur31]

The copy of SUNDIALS that is bundled with neuron is quite old, old enough that the API for SUNDIALS seems to have changed since. It'll be good if neuron could be updated to use the latest SUNDIALS code, to take advantage of its optimisations and bug fixes. This will also make it much easier for downstream distributions to provide neuron.

The current version of SUNDIALS is 3.2.1: https://computation.llnl.gov/projects/sundials/sundials-software

[brunomaga]

I just had the same conversation with @nrnhines . I managed to update CVODES to version 3.2.1 on a subset of NEURON code, and will start working on porting it fully to NEURON on the 19th November. Hopefully by the end of the month will be completed.

[sanjayankur31]

Ah, thank you very much. I was about to embark on it myself, but given my lack of experience with sundials, it'd be quite the task. Would you have a fork where you're working on this that I can watch for updates?

[brunomaga]

I will create a branch when I start the development on the 19th, and let you know.

[brunomaga]

You can follow the development in this fork:
https://github.com/brunomaga/nrn
Beta version is done, we are now in testing stage. @nrnhines will then decide about the merge.

[sanjayankur31]

Thanks very much. I'll test out and give any feedback if it comes up.

[sanjayankur31]

@brunomaga: any idea when the update will be merged to the neuron code base?

[sanjayankur31]

I get errors building for MPI from your branch, unfortunately. This, specifically: https://github.com/brunomaga/nrn/blob/master/src/nrncvode/cvodeobj.h#L7 -> the nvector/nvector_serial.h file does not seem to exist in the source anywhere. I'll keep looking, though.

[nrnhines]

the cvode3 branch of this repository compiles (at least on my desktop). The build is
git checkout cvode3
sh build.sh
./configure --prefix=pwd --with-nrnpython --with-paranrn --disable-rx3d
make -j install

The neurondemo yields
quantitatively similar results for cvode and ida. I need to resolve nrntest repository failures for 12 tests.
At least a few of those are double precision non-identitity at the 1e-9 resolution level but a few may be
exhibiting substantive errors.

[nrnhines]

One change to the cvode3 branch was required to resolve a nrntest issue (2733468)
I believe all the other nrntest failures are due to the cvode3 generally computing different time steps
than the older cvode.

[sanjayankur31]

Thanks, @nrnhines : I'll switch to the cvode3 branch then. (I was following @brunomaga 's fork till now).

[brunomaga]

@sanjayankur31 I'm in the process to porting the code to SUNDIALS4.
For the time being, check #165 if you intend to run CVODE on very large networks.

[alexsavulescu]

superseded by #1328