[dpdk-dev] [PATCH v6 0/8] fib: implement AVX512 vector lookup

[Richardson, Bruce]

> -----Original Message-----
> From: Thomas Monjalon <thomas at monjalon.net>
> Sent: Wednesday, July 15, 2020 1:45 PM
> To: Medvedkin, Vladimir <vladimir.medvedkin at intel.com>
> Cc: Kinsella, Ray <mdr at ashroe.eu>; Stephen Hemminger
> <stephen at networkplumber.org>; dev at dpdk.org; david.marchand at redhat.com;
> jerinj at marvell.com; Ananyev, Konstantin <konstantin.ananyev at intel.com>;
> Richardson, Bruce <bruce.richardson at intel.com>; Mcnamara, John
> <john.mcnamara at intel.com>; O'Driscoll, Tim <tim.odriscoll at intel.com>
> Subject: Re: [dpdk-dev] [PATCH v6 0/8] fib: implement AVX512 vector lookup
> 
> 15/07/2020 14:29, Medvedkin, Vladimir:
> > On 15/07/2020 12:59, Thomas Monjalon wrote:
> > > 15/07/2020 12:35, Medvedkin, Vladimir:
> > >> On 15/07/2020 10:47, Thomas Monjalon wrote:
> > >>> 14/07/2020 16:38, Stephen Hemminger:
> > >>>> "Kinsella, Ray" <mdr at ashroe.eu> wrote:
> > >>>>> On 13/07/2020 23:19, Stephen Hemminger wrote:
> > >>>>>> Did anyone else see the recent AVX512 discussion from Linus:
> > >>>>>>     "I hope AVX512 dies a painful death, and that Intel starts
> fixing real problems
> > >>>>>>      instead of trying to create magic instructions to then
> create benchmarks that they can look good on.
> > >>>>>
> > >>>>> Yup - I saw this one.
> > >>>>> Sweeping statements like these are good to provoke debate, the
> truth is generally more nuanced.
> > >>>>> If you continue to read the post, Linus appears to be mostly
> questioning microprocessor design decisions.
> > >>>>>
> > >>>>> That is an interesting discussion, however the reality is that the
> technology does exists and may be beneficial for Packet Processing.
> > >>>>>
> > >>>>> I would suggest, we continue to apply the same logic governing
> adoption of any technology by DPDK.
> > >>>>> When the technology is present and a clear benefit is shown, we
> use it with caution.
> > >>>>>
> > >>>>> In the case of Vladimir's patch, the user has to explicitly
> > >>>>> switch on the AVX512 lookup with RTE_FIB_DIR24_8_VECTOR_AVX512.
> > >>>>
> > >>>> Using what is available makes sense in DPDK.
> > >>>
> > >>> Why does it require explicit  enabling in application?
> > >>> AVX512 is not reliable enough to be automatically used when
> available?
> > >>
> > >> It is reliable enough. User have to explicitly trigger to avx512
> > >> lookup because using avx512 instructions can reduce the frequency
> > >> of your cores. The user knows their environment better. So the
> > >> scalar version is used so as not to affect the frequency.
> > >
> > > So the user must know which micro-optimization is better for a code
> > > they don't know. Reminder: an user is not a developper.
> > > I understand we have no better solution though.
> > > Can we improve the user experience with some recommendations, numbers,
> etc?
> > >
> >
> > In case where a user is a developer (dpdk users are mostly devs,
> > aren't
> > they?) who uses the fib library in their app may decide to switch to
> > avx512 lookup using rte_fib_set_lookup_fn() when they know that their
> > code is already using avx512 (ifdef, startup check, etc).
> > In other case an app developer, for example, could provide to user
> > command line option or some interactive command to switch lookup
> function.
> > I'd recommend to run testfib app with various "-v" options to evaluate
> > lookup performance on a target system to make a decision.
> 
> I think this is the difference between a library for hackers, and a
> product for end-users.
> We are not building a product, but we can make a step in that direction by
> documenting some knowledge.
> I don't know exactly what it means in this case, so I'll let others
> suggest some doc improvements (if anyone cares).
> 

We have got a patchset in the works to try and make AVX-512 use simpler for 20.11,
by providing both developer APIs and end-user cmdline flags to control this
centrally for DPDK, rather than having each library provide its own magic hooks
to optionally enable this support. As part of that set, we'll see about what
doc updates need to be made also - again covering both developer and end-app user.

Hopefully we can get that set out soon to get early feedback and reach a good
conclusion.

/Bruce