Am 15.04.2017 um 12:27 schrieb Riccardo (Jack) Lucchetti:
On Sat, 15 Apr 2017, Artur Tarassow wrote:
> Dear gretl users,
>
> I have a three laptops (some of them are about 5 years old but have
> fast CPUs implemented) and a desktop computer available, and was
> thinking to set up a small-scale cluster.
>
> It seems to me that the MPI interface which gretl supports is just
> designed for such a task. Is this correct?
Yes.
> Does anybody have some experience with it?
Some. In order to make the most of it, keep in mind that the most
stringent constraint on performance is the ethernet connection. So,
you're going to have good results as far as you manage to split the
workload between processes in such a way to minimise the amount of data
they have to trade between one another. Sometimes it's relatively simple
to do so, sometimes it requires some ingenuity.
HTH,
Thank you for your answer, Jack. I was thinking to setup an experiment
estimating big VARs combining a rolling-window plus bootstrap approach
where at each window (in total about 500) some heavy-load bootstrap is done.
I guess in this case it wouldn't make much sense to distribute the
bootstrap part over all PCs but rather the rolling-window task.
Best,
Artur