[ovs-dev] [PATCH] Improve MPLSoGRE performance by reducing EMC hash collisions.

Nitin Katiyar nitin.katiyar at ericsson.com
Wed Aug 14 08:33:23 UTC 2019



> -----Original Message-----
> From: Ilya Maximets [mailto:i.maximets at samsung.com]
> Sent: Wednesday, August 14, 2019 1:42 PM
> To: Nitin Katiyar <nitin.katiyar at ericsson.com>; ovs-dev at openvswitch.org
> Cc: Stokes, Ian <ian.stokes at intel.com>
> Subject: Re: [ovs-dev] [PATCH] Improve MPLSoGRE performance by reducing
> EMC hash collisions.
> 
> On 13.08.2019 20:49, Nitin Katiyar wrote:
> > When a packet is received, the RSS hash is calculated if it is not
> > already available. The Exact Match Cache (EMC) entry is then looked up
> > using this RSS hash.
> >
> > When a MPLSoGRE encapsulated packet is received, the GRE header is
> > popped, the RSS hash is invalidated and the packet is recirculated for
> > further processing. When the recirculated packet is processed, the
> > MPLS header is popped and the packet is recirculated again. Since the
> > RSS hash has not been invalidated here, the EMC lookup will hit the
> > same entry as that after the first recirculation. This degrades
> > performance severely.
> 
Hi Ilya,
> Hmm. Function 'dpif_netdev_packet_get_rss_hash()' accounts recirculation
> depth into the hash to avoid such cases. Why this doesn't work for you?
This will not very for multiple inner flows. If there are multiple flows are sent across same tunnel (with same MPLS label) then they will all lead to same hash as external tuple and recirculation id remains same for them. Let me know if I am wrong.

Regards,
Nitin
> 
> Best regards, Ilya Maximets.


More information about the dev mailing list