On Apr 21, 2023 / 09:03, Daniel Wagner wrote:
> On Fri, Apr 21, 2023 at 08:33:46AM +0200, Hannes Reinecke wrote:
> > +_nvme_calc_rand_io_size() {
> > > + local img_size_mb
> > > + local io_size_mb
> > > +
> > > + img_size_mb="$(convert_to_mb "$1")"
> > > + io_size_mb="$(printf "%d" $((((img_size_mb * 1024 * 1024) / $(nproc) - 1) / 1024)))"
> > > +
> >
> > ... ending with ridiculous small io sizes on machines with lots of CPUs.
> > Please cap nproc by something sane like 32.
>
> Yeah, propably not really good long time strategy. I was wondering if we should
> make run_fio() variants smarter and do the size callculation there and not by
> the callee. If we do this, we could make the number of jobs dependend on CPUs
> and image size a bit nicer.
The usage of _run_fio_rand_io() look different for each test case. nvme/032
kills the fio process when it is no longer required. Then IO size reduction with
_nvme_calc_io_size() will not reduce runtime of nvme/032. I think nvme/040 has
same story, since _nvme_delete_ctrl will stop the fio process with I/O error.
On the other hand, nvme/045 and nvme/047 may have different usage. I'm not sure
if these test case needs I/O with all CPUs. It would be better to have other
run_fio() variant as Daniel mentioned, so that their runtime will not depend on
number of CPUs.
