Re: [PATCH] usertools: suggest use of hwloc for new cpu

[Thomas Monjalon <thomas@monjalon.net>]

04/09/2023 12:20, Bruce Richardson:
> On Mon, Sep 04, 2023 at 11:11:20AM +0100, Ferruh Yigit wrote:
> > On 8/14/2023 10:25 AM, Konstantin Ananyev wrote:
> > > 
> > >> On Sun, Aug 13, 2023 at 08:52:01AM -0700, Stephen Hemminger wrote:
> > >>> On Sun, 13 Aug 2023 02:12:03 +0000
> > >>> "Varghese, Vipin" <Vipin.Varghese@amd.com> wrote:
> > >>>
> > >>>>>
> > >>>>> On Sat, 12 Aug 2023 06:27:20 +0530
> > >>>>> Vipin Varghese <vipin.varghese@amd.com> wrote:
> > >>>>>
> > >>>>>> Most modern processor now supports numa by partitioning NUMA based on
> > >>>>>> CPU-IO & Last Level Cache within the same socket.
> > >>>>>> As per the discussion in mailing list, suggesting the make use of
> > >>>>>> hw-loc for such scenarios.
> > >>>>>>
> > >>>>>> Signed-off-by: Vipin Varghese <vipin.varghese@amd.com>
> > >>>>>
> > >>>>> NAK, no scripting hwloc, it is ugly and creates a dependency that is not listed
> > >>>>> in DPDK packaging.
> > >>>>
> > >>>> There is no calls to hwloc within in thescript. Hence not clear what does ` NAK, no scripting hwloc it is ugly and creates a
> > >> dependency that is not listed in DPDK packaging.`.
> > >>>>
> > >>>> Requesting to cross check why NAK is shared for `print` as suggestion. Hence, I have disagree to this.
> > >>>
> > >>> Sorry, I misinterpreted what the print's were doing.
> > >>> Better off not to list exact flags, the lstopo may change and user may want different
> > >>> format anyway.
> > >>>
> > >>> How about something like this?
> > >>>
> > >>>
> > >>>  doc/guides/rel_notes/deprecation.rst | 5 +++++
> > >>>  usertools/cpu_layout.py              | 5 +++++
> > >>>  2 files changed, 10 insertions(+)
> > >>>
> > >>> diff --git a/doc/guides/rel_notes/deprecation.rst b/doc/guides/rel_notes/deprecation.rst
> > >>> index 317875c5054b..25a116900dfb 100644
> > >>> --- a/doc/guides/rel_notes/deprecation.rst
> > >>> +++ b/doc/guides/rel_notes/deprecation.rst
> > >>> @@ -185,3 +185,8 @@ Deprecation Notices
> > >>>    will be deprecated and subsequently removed in DPDK 24.11 release.
> > >>>    Before this, the new port library API (functions rte_swx_port_*)
> > >>>    will gradually transition from experimental to stable status.
> > >>> +
> > >>> +* cpulayout: The CPU layout script is unable to deal with all the possible
> > >>> +  complexities of modern CPU topology. Other existing tools offer more
> > >>> +  features and do a better job with keeping up with innovations.
> > >>> +  Therefore it will be deprecated and removed in a future release.
> > >>
> > >> Does the script really do that bad a job? While I can understand us looking
> > >> to recommend alternatives, I actually find the script in it's current form
> > >> really handy - much more so than working out the exact flags for lstopo
> > >> etc. Since it's not a large maintenance burden, I'd request we keep it
> > >> around - while still recommending lstopo to users.
> > > 
> > > +1
> > > I do use it on regular basis.
> > > It would be a pity if it will be gone.
> > >
> > 
> > I also use it time to time and find it useful.
> > 
> > But it is not accurate/correct for some AMD platforms (for various NPS
> > (Nodes per Socket) values).
> > So either it needs to be updated/improved or replaced.
> > 
> > Vipin sent a patch [1] to update it but it is question how much of this
> > logic belongs to DPDK, or should we rely on external tools dedicated for
> > this purpose.
> > 
> 
> I'd like to suggest that we take a slightly ambiguous position on this
> script. Specifically:
> 
> I think we should "recommend" but not "rely on" external tools for this.
> Specifically, I think that recommending use of hwloc is the best thing to
> do as it's better maintained and packaged for windows. However, for quick
> use in many situations, cpu_layout does the job as well or better in terms
> of simplicity of use and output.

We could also contribute to hwloc to have a similar simple output.