Re: [dpdk-dev] [PATCH] config/x86: add support for AMD platform

[Thomas Monjalon <thomas@monjalon.net>]

18/11/2021 14:52, Bruce Richardson:
> On Thu, Nov 18, 2021 at 01:25:38PM +0100, Thomas Monjalon wrote:
> > I request a techboard decision for this patch.
> > 
> > 
> > 02/11/2021 20:04, Thomas Monjalon:
> > > 02/11/2021 19:45, David Marchand:
> > > > On Tue, Nov 2, 2021 at 3:53 PM Aman Kumar <aman.kumar@vvdntech.in> wrote:
> > > > >
> > > > > -Dcpu_instruction_set=znverX meson option can be used
> > > > > to build dpdk for AMD platforms. Supported options are
> > > > > znver1, znver2 and znver3.
> > > > >
> > > > > Signed-off-by: Aman Kumar <aman.kumar@vvdntech.in>
> > > > > ---
> > > > >  dpdk_conf.set('RTE_CACHE_LINE_SIZE', 64)
> > > > >  dpdk_conf.set('RTE_MAX_LCORE', 128)
> > > > >  dpdk_conf.set('RTE_MAX_NUMA_NODES', 32)
> > > > > +
> > > > > +# AMD platform support
> > > > > +if get_option('cpu_instruction_set') == 'znver1'
> > > > > +    dpdk_conf.set('RTE_MAX_LCORE', 256)
> > > > > +elif get_option('cpu_instruction_set') == 'znver2'
> > > > > +    dpdk_conf.set('RTE_MAX_LCORE', 512)
> > > > > +elif get_option('cpu_instruction_set') == 'znver3'
> > > > > +    dpdk_conf.set('RTE_MAX_LCORE', 512)
> > > > > +endif
> > > > 
> > > > I already replied to a similar patch earlier in this release.
> > > > https://inbox.dpdk.org/dev/CAJFAV8z-5amvEnr3mazkTqH-7SZX_C6EqCua6UdMXXHgrcmT6g@mail.gmail.com/
> > > > 
> > > > So repeating the same: do you actually _need_ more than 128 lcores in
> > > > a single DPDK application?
> > 
> > We did not receive an answer to this question.
> > 
> > > Yes I forgot this previous discussion concluding that we should not increase
> > > more than 128 threads.
> > 
> > We had a discussion yesterday in techboard meeting.
> > The consensus is that we didn't hear for real need of more than 128 threads,
> > except for configuration usability convenience.
> > 
> > Now looking again at the code, this is how it is defined:
> > 
> > option('max_lcores', type: 'string', value: 'default', description:
> >        'Set maximum number of cores/threads supported by EAL;
> >        "default" is different per-arch, "detect" detects the number of cores on the build machine.')
> > config/x86/meson.build: dpdk_conf.set('RTE_MAX_LCORE', 128)
> > config/ppc/meson.build: dpdk_conf.set('RTE_MAX_LCORE', 128)
> > config/arm/meson.build: it goes from 4 to 1280!
> > 
> > So I feel it is not fair to reject this AMD patch if we allow Arm to go beyond.
> > Techboard, let's have a quick decision please for 21.11-rc4.
> > 
> I would support increasing the default value for x86 in this release.

This patch is not increasing the default for all x86,
only for some CPUs as given at compilation time.
I think it is the same logic as Arm CPU-specific compilation.

> I believe Dave H. had some patches to decrease the memory footprint
> overhead of such a change. I don't believe that they were merged, and while
> it's a bit late for 21.11 now, those should be considered for 22.03 release
> and then maybe for backport.