On Mon, Aug 12, 2024 at 05:02:11PM +0200, Morten Brørup wrote:
> > From: Bruce Richardson [mailto:bruce.richard...@intel.com]
> >
> > On Mon, Aug 12, 2024 at 04:10:49PM +0200, Morten Brørup wrote:
> > > > From: Bruce Richardson [mailto:bruce.richard...@intel.com]
> > > >
> > > > The default number of ethernet queues per port is currently set to
> > > > 1k which is more than enough for most applications, but still is lower
> > > > than the total number of queues which may be available on modern NICs.
> > > > Rather than increasing the max queues further, which will increase
> > > > the memory footprint (since the value is used in array dimensioning),
> > > > we can instead make the value a meson tunable option - and reduce the
> > > > default value to 256 in the process.
> > >
> > > Overall, I agree that this tunable is not very exotic, and can be exposed
> > > as
> > suggested.
> > > The reduction of the default value must be mentioned in the release notes.
> > >
> >
> > Yes, good point. I'll add that in any next version.
>
> ACK.
>
> >
> > >
> > > > # set other values pulled from the build options
> > > > dpdk_conf.set('RTE_MAX_ETHPORTS', get_option('max_ethports'))
> > > > +dpdk_conf.set('RTE_MAX_QUEUES_PER_PORT',
> > > > get_option('max_queues_per_ethport'))
> > >
> > > Please rename RTE_MAX_QUEUES_PER_PORT to _PER_ETHPORT, so it resembles
> > MAX_ETHPORTS. For API backwards compatibility, you can add:
> > > #define RTE_MAX_QUEUES_PER_PORT RTE_MAX_QUEUES_PER_ETHPORT
> > >
> >
> > Agree that would more consistent. That would probably best be a separate
> > patch, since we'd want to convert all internal use over. Will make this a
> > two-patch set in next version.
>
> ACK. And agree about two-patch series.
>
> >
> > >
> > > I wonder if it would be possible to have separate max sizes for RX and TX
> > queues? If it can save a non-negligible amount of memory, it might be useful
> > for some applications.
> > >
> >
> > That is an interesting idea. It would certainly allow saving memory for
> > use-cases where you want a large number of rx queues only, or tx queues
> > only. However, the defaults are still likely to remain the same. The main
> > issue I would have with it, is that it would mean having two build time
> > options rather than 1, and I'm a bit concerned at the number of options we
> > seem to be accumulating in DPDK.
> >
> > Overall, I'm tending towards suggesting that we not do that split, but I'm
> > open to being convinced on it.
>
> I would guess that many applications have an asymmetrical split of number of
> RX/TX queues. So I would argue that:
> The reason to make this configurable in meson is to conserve memory, so why
> only go half the way if there is more memory to be conserved?
>
> The distros will use oversize maximums anyway, but custom built applications
> might benefit.
>
> Here's a weird thought:
> Perhaps RX and TX maximums can be controlled individually by changing
> rte_config.h, and they can be overridden via one meson configuration
> parameter to set both (to the same value).
>
> >
> > >
> > > With suggested changes (splitting RX/TX maximums not required),
> > > Acked-by: Morten Brørup <m...@smartsharesystems.com>
> > >
>
> My ACK remains; splitting RX/TX maximums is not Must Have, it is Nice To Have.
>
Let me see how much in involved in trying to split...
