freebsd stable - Jan 2014 - forwarding performance with releng7 vs 8 vs 9 vs 10 on an Alix box

We have a couple of sites where we found it was possible to live lock /
crash an alix box if there were too many pps flowing across the unit in
RELENG_8.  The same device with releng_7 did not show the same
behaviour.  As these boxes are remote and the watchdog would kick in, I
was not able to debug things very much. Even back in the lab, it wasnt
always easy to trigger the crash synthetically, but I could.

Has anyone done any network performance comparisons between the various
branches for speed and stability for single core embedded platforms like
the ALIX or Soekris ?

	---Mike
-- 
-------------------
Mike Tancsa, tel +1 519 651 3400
Sentex Communications, mike at sentex.net
Providing Internet services since 1994 www.sentex.net
Cambridge, Ontario Canada   http://www.tancsa.com/

On Fri, Jan 17, 2014 at 11:54 AM, Mike Tancsa <mike at sentex.net> wrote:
> We have a couple of sites where we found it was possible to live lock /
> crash an alix box if there were too many pps flowing across the unit in
> RELENG_8.  The same device with releng_7 did not show the same
> behaviour.  As these boxes are remote and the watchdog would kick in, I
> was not able to debug things very much. Even back in the lab, it wasnt
> always easy to trigger the crash synthetically, but I could.
>
> Has anyone done any network performance comparisons between the various
> branches for speed and stability for single core embedded platforms like
> the ALIX or Soekris ?
>
>         ---Mike
>
FreeBSD 8 introduced default support for TSO. You might try turning it off
(-tso) if it is enabled on the interfaces involved.

-- 
R. Kevin Oberman, Network Engineer, Retired
E-mail: rkoberman at gmail.com

On 17 January 2014 11:54, Mike Tancsa <mike at sentex.net>
wrote:
> We have a couple of sites where we found it was possible to live lock /
> crash an alix box if there were too many pps flowing across the unit in
> RELENG_8.  The same device with releng_7 did not show the same
> behaviour.  As these boxes are remote and the watchdog would kick in, I
> was not able to debug things very much. Even back in the lab, it wasnt
> always easy to trigger the crash synthetically, but I could.
>
> Has anyone done any network performance comparisons between the various
> branches for speed and stability for single core embedded platforms like
> the ALIX or Soekris ?
No, but it would be nice to actually do that, then figure out why it's
livelocking.

I can introduce livelock in a single core of a much, much more
powerful box by doing certain bad things. It's almost like the
scheduler needs to have some kind of "fuck it, I'm going FIFO!"
mode.


-a