Re: [LibreQoS] libreqos vs a vs paraqum

[Stephen Hemminger <stephen@networkplumber.org>]

On Sat, 12 Nov 2022 07:54:02 -0800
Dave Taht <dave.taht@gmail.com> wrote:

> I (personally) have zero interest in dpdk and vpp. I don't want to
> give up all the other nifty things a linux box can do by using it, so
> I followed along on the xdp work - but I concede that these
> technologies are probably always going to be faster than xdp, and
> there are a ton of products deployed using it. I'd hoped someone would
> fund open sourcing a fq_codel or cake implementation for it. (Same
> goes for freebsd and pfsense which could use a native BQL + fq_codel
> implementation. The BSD packet buffering scheme is really alien to
> me). DPDK and vpp were born of the recognition that with conventional
> processors, the bottleneck is more on the read side than the write
> side, once you crack 10gbit.

Linux and BSD has to solve the general purpose problem (lots of devices, any application, etc)
and therefore has lots of locking an other overhead. All that overhead keeps
it in the 1 to 4 Million packets/second/core forwarding range.  With BPF some
of this can be bypassed but then most of the interesting stuff is gone and still
limited to 10Mpps.

The dedicated stuff gets much higher PPS but at a cost, limited HW suppor,
no generic locking etc.

Doing research before diving into CAKE on DPDK, the real issue is that there
is no supported flow classification model in DPDK. There was a limited/dead library
and doing the full classification like Linux gets harder.