is anybody here working on the udp performance problem in freebsd6?

Geert Jan de Groot GeertJan.deGroot at xs4all.nl
Fri Sep 15 21:52:27 UTC 2006


On Fri, 15 Sep 2006 20:54:19 +0000  Paul Vixie wrote:
> > > http://unix.derkeiler.com/Mailing-Lists/FreeBSD/net/2006-09/msg00129.html
> > 
> > My impression of this (e-mail) thread was that it's more likely a network
> > driver issue, and I'm not very much enthusiastic about tracking it further
> > right now.
> it's been seen on both em and bge interfaces.  definitely a freebsd kernel
> networking problem of some kind.

What isn't clear for me from the thread is whether there is saturation
if one throws more than 15K packets at the box, or it's just
a matter of the send-respond path a bit slower and hence 
the test code being locked, waiting for responses from the 
test box.

If it's the latter, do note that fxp, em and bge all have the ability
to coalesce interrupts. Instead of older cards what would interrupt
on receipt of a single packet, they wait a little longer until
either a hardware-timeout happens, or more packets have been received,
the idea being that a single interrupt service run would service
them all. We all know that interrupts are expensive, so the less
interrupts, the better.

If it's the latter case, then I think the test is irrelevant:
I'm not delaying my DNS requests till you got your responses,
and the real test would be how many requests one can throw at the box
before it saturates.

The coalescing stuff is quite sensitive and only available with
NDA's in place. Unfortunately, I no longer have access to these
documents.

Geert Jan



More information about the bind-workers mailing list