grout use-cases for host traffic

[Christoph]

Robin Jarry:
> We have a lot on our plates these days, but I agree it would help to
> have something easy to consume to help newcomers. I will see if I can
> put something in place so that people can enrich it as they go.
> 
>> Yes we are looking at high bandwidth use cases (>10Gbps).
> 
> Then, do not use grout as direct termination (for now).

Ok sad to hear that grout can not help solve our use-case but thanks for 
the clear recommendation to not use grout in our case.

>> Since frr (bgp)/grout is expected to decide where to route the traffic
>> (which grout interface), this can't be done using a separate non-grout
>> interface on the same system, correct? So we would use the TUN/TAP setup
>> I guess or am I missing something? :)
> 
> You can do it by spitting your physical NIC in multiple VFs and exposing
> only one to the outside world. Hiding the others behind for local Linux
> use. See
> https://hosted-files.sched.co/dpdksummit2026/75/grout-dpdk-summit-2026.pdf
> (slide 11) for more details.
> 
> Grout was designed to be good at *forwarding* traffic. Local TCP/UDP
> termination was added for low bandwidth use-cases (BGP/OSPF/ISIS).
> 
> Being better than Linux at *terminating* traffic (especially TCP) is
> hard. The challenge is matching the level of batching and tweaks that
> were added in the Linux TCP stack (end-to-end TSO mainly).
> 
> What are you planning to use grout for? Could you describe your topology
> and use-case?

We run servers doing mainly two things (both on the same HW and OS - 
mainly for power efficiency reasons):
- eBGP routing with BGP full tables to multiple transit provider
- TCP proxy servcies (tor)

~50% of the traffic is terminated locally for the TCP proxy services, 
the rest is forwarded traffic to other servers in our AS.

So as far as I understood for 50% of the traffic - the pure forwarding 
part - grout would outperform Linux kernel based routing but the rest, 
the locally terminated traffic the performance with using grout would be 
worse than useing Linux.

When documentation is available we might give it a try to be able to 
compare it to VPP and Linux kernel based routing/termination.

best regards,
Christoph