[ovs-dev] [PATCH v3 0/6] MFEX Optimizations IPv6 + Hashing

Stokes, Ian ian.stokes at intel.com
Tue Oct 5 14:45:02 UTC 2021


9/21/21 12:23, Kumar Amber wrote:
> > ---
> > v3:
> > - rebase to master.
> > v2:
> > - fix the CI build.
> > - fix check-patch for co-author.
> > ---
> >
> > The patch-set introduces AVX512 optimizations of IPv6
> > traffic profiles and hashing improvements for all AVX512
> > supported traffic profiles for IPv4 and IPv6.
> >
> > Kumar Amber (6):
> >   dpif-netdev/mfex: Add AVX512 basic ipv6 traffic profiles
> >   dpif-netdev/mfex: Add AVX512 vlan ipv6 traffic profiles
> >   dpif-netdev/mfex: Add packet hash check to autovalidator
> >   dpif-netdev/mfex: Add ipv4 profile based hashing
> >   dpif-netdev/mfex: Add ipv6 profile based hashing
> >   dpif-netdev/mfex: Avoid hashing when opt mfex called
> >
> >  NEWS                              |   7 +
> >  lib/automake.mk                   |   1 +
> >  lib/dpif-netdev-avx512.c          |   6 +-
> >  lib/dpif-netdev-extract-avx512.c  | 348 +++++++++++++++++++++++++++++-
> >  lib/dpif-netdev-private-extract.c |  63 +++++-
> >  lib/dpif-netdev-private-extract.h |  12 ++
> >  tests/pcap/mfex_test.pcap         | Bin 416 -> 632 bytes
> >  7 files changed, 432 insertions(+), 5 deletions(-)
> >
> 
> Hi.  A few months ago I was told that it's easy for Intel to set up CI
> to test upstream patches with AVX512 features enabled.  Is there any
> progress on that front?
> 
> My point is that we should refrain from adding new features in this
> area until we have a proper CI.  Especially considering the unit test
> failure you reported yesterday, which is supposedly related to AVX512
> optimizations.
> 

Hi Ilya,

Apologies for the delay in response, I've been on PTO the past 2 weeks so only catching up now.

There is an ongoing effort to deploy an AVX512 CI. Aaron, Michael and myself have started to look at this. I think the initial target was to have something up and running by EOY.

> // Marking this patch-set as deferred for now.

Agree with marking this deferred for now.

However I don’t think this should be dependent on delivery of the CI as that could be some time away yet. I think if we are to run the tests manually on an AVX512 system (We could even run it on the system that has been setup and reserved for the CI in the Intel Lab) and post results in response/part of the review of the series that should help progress the series in the meantime?

Thanks
Ian

> 
> Best regards, Ilya Maximets.


More information about the dev mailing list