Re: How multithreaded is networking in FreeBSD 15.1?

On Sun, 16 Aug 2026 at 18:25, Brett Glass <[email protected]> wrote:
>
> At 03:00 PM 8/16/2026, David Christensen wrote:
>
> >My theory for running mixed sequential/ parallel programs is that
> >there is a speedup vs. efficiency sweet spot when the number of
> >threads is numerically equal to Amdahl's maximum speedup
> >value.  If I run the program with more threads, I will get more
> >speedup but less efficiency. If I run the program with less
> >threads, I will get less speedup but more efficiency.
>
> David, thank you for the ideas. However, what I am asking is which
> choice of processors (more simple cores, lower clock speed vs.
> fewer fast cores, higher clock speed) will maximize the networking
> and routing performance of FreeBSD 15.1, given that we will be
> using lagg(4) heavily and IPFW as the firewall. The choice of
> hardware will depend on, among other things, how well
> multithreading works within the network stack and whether the
> threads are likely to get in one another's way. In older versions
> of FreeBSD it paid to use fewer, faster CPUs to avoid contention;
> I'm not sure if that's still the case with the latest kernel.

We should fix any contention issues that you see in 2026.

You should likely target something with say 4 to 8 cores for parallel network
TX/RX and then some other cores to run control/management plane stuff.

Ideally the ethernet input path into lagg should be broken out into parallel
RX rings based on 2, 3, 4 tuple (ethernet, ip, tcp/udp header) contents.

TX will need a bit more careful analysis because yeah, RSS, lagg and
TX queue / NIC selection. But we can fix that.

> >All that said, have you considered FreeNAS (now TrueNAS)?
>
> We're not interested in a "canned" solution. We want to handcraft
> these routers for the purpose, building them from the ground up on
> standard releases of FreeBSD.

If you want to chat offline about it then let me know. It would be good to
get a "reference" platform together to explore what the performance issues
are and just iterate on fixing them.



-adrian