Avoid CV_TOO_CLOSE error by increasing minimum variable step size

[brunomaga]

Hi Michael @nrnhines:

I realised that if one makes a call to

CVode(cvode mem, tout, yout, &tret, itask);

and tout (the target time instant) and t0 (the current time instant) are too close, the following error message will show up:

[CV_TOO_CLOSE]
tout too close to t0 to start  integration. 

This will fail from stepping, and if called inside a while (t<tout) loop, it will hang the execution indefinitely. This issue is hard to reproduce and occurred to me only on very large networks, above 32K neurons.

A quick fix is to enforce an initial minimum step size, e.g. a call to:

CVodeSetInitStep(cvode_mem_f, step_size); 

for having a larger first step, or ideally

CVodeSetMinStep(cvode_mem_f, step_size); 

to guarantee that state is always advanced by some minimum timestep value. However, if step_size is too large, then the following error may occur:

[CV_CONV_FAILURE]
the corrector convergence failed repeatedly or with |h| = hmin. 

So the trick is to pick a step size big enough to advance the state, while being small enough to make the solver converge within the tolerance values provided. From my testing, 1e-5 is a good value.

This ticket requests the default minimum or initial variable step size in NEURON to be increased from the actual value of 0, to avoid CV_TOO_CLOSE errors.

[nrnhines]

So this issue is resolved if you call cvode.minstep(1e-5)?
Do you think this problem is occurring when t is very large so that a natural first step size after an event
initialization happens to be less than 2* tround = cv_mem->cv_uround * SUNMAX(SUNRabs(cv_mem->cv_tn), SUNRabs(tout));
note that cv_uround is UNIT_ROUNDOFF is DBL_EPSILON, I think about 2e-16

[brunomaga]

Yes, in my case. I solved it with:

CVodeSetMinStep(cvode_mem_f, 1e-5); 

In your speculative approach it will work fine I believe. This is simply because you perform a step with CVode( ..., CV_ONE_STEP) and the tout is a very big value (10000). CVode will use this big value as an estimation of the step size of the first iteration of that step, and tout is too far from your t anyway.
However, if you try any use case where a neuron needs to pause at a precise time instant (e.g. communication barrier in parallel executions), then CV_NORMAL or CV_ONE_STEP to that time instant may yield this error if the previous step finished immediately before in time.