[Gretl-devel] Re: MPI bootstrap failures depending on number of iterations
Am 03.11.2021 um 17:32 schrieb Riccardo (Jack) Lucchetti:
On Wed, 3 Nov 2021, Sven Schreiber wrote:
>
> I'm struggling with a hard-to-pinpoint problem involving MPI - this is
> in relation to the johansensmall package and a pending update.
>
> I cannot yet give very much detail, but in a certain special context for
> a lower number of specified bootstrap iterations the thing runs fine,
> whereas when cranking up that iteration number, it basically fails with
> the typical obscure message from gretlmpi that an external command
> failed. So the algorithm and code per se do not seem to be wrong, or
> else it would always fail.
>
> My suspicion for now is related to some kind of memory limitation or a
> similar constraint. Is that conceivable? How would one check for that?
Could it be that there is an extreme corner case, that occurs with
very small probability, that makes some command fail with error under
very extreme circumstances?
All ideas and brainstorming is welcome, thanks.
This would explain why the problem occurs only with a very high number
of replications.
However, when I run the thing sequentially with the same (high, but not
that high, actually) number of replications, there is no problem.
Therefore I believe it must be something more MPI-specific.
Another thought: Maybe some subtle interaction thing with the RNG in MPI?
Well, I know this is all very vague and shooting in the dark. I'll try
to investigate further.
thanks
sven