Date: Fri, 13 Jul 2001 16:06:59 +0200 (CEST) From: Martin Karsten <Martin.Karsten@KOM.tu-darmstadt.de> To: Mike Silbersack <silby@silby.com> Cc: freebsd-net@freebsd.org Subject: Re: UDP packet loss on FreeBSD 4.x Message-ID: <200107131406.QAA03047@KOM.tu-darmstadt.de> In-Reply-To: <20010711153113.L2408-100000@achilles.silby.com> from Mike Silbersack at "Jul 11, 2001 03:33:24 pm"
next in thread | previous in thread | raw e-mail | index | archive | help
Thanks for this hint, but it doesn't seem to help. Here's the output of
netstat -m after a test with packet losses:
386/608/8192 mbufs in use (current/peak/max):
385 mbufs allocated to data
1 mbufs allocated to packet headers
384/494/2048 mbuf clusters in use (current/peak/max)
1140 Kbytes allocated to network (18% of mb_map in use)
0 requests for memory denied
0 requests for memory delayed
0 calls to protocol drain routines
IMO the problem must have been introduced due to a change between versions
3.4 and 4.x. I'm not experienced with driver programming, the only obvious
difference (at least with respect to the xl driver) is that it uses the mii
driver now.
Thanks anyway,
Martin
> > Greetings,
> >
> > I have observed the following behaviour on FreeBSD 4.x platforms (4.0 till
> > 4.3 seem to be affected).
> >
> > When receiving a sufficiently fast stream of UDP packets (the borderline
> > seems to be around 3,500 packets/sec for e.g. the 'xl' driver on a 450MHz
> > Pentium), an application on the receiving host does not receive all packets
> > anymore, depending on which nic driver is used. A fraction of the lost
> > packets is reported in net.inet.ip.intr_queue_drops, but not all.
> >
> > Tests have shown that the losses occur for the 'xl' or 'ti' drivers, whereas
> > the 'fxp' and 'de' driver don't seem to be affected.
>
> Check the output of netstat -m after doing the test. I know that the xl
> driver puts all incoming packets into mbuf clusters, while the dc driver
> goes straight into mbufs if possible. That might not explain a slowdown
> in overall throughput, but it does mean that dc cards have a 4x larger
> incoming packet queue than xl cards.
>
> Back to netstat -m. If you see that your peak number of clusters is
> hitting the max, that's partially your answer, and upping the # of mbuf
> clusters could help.
>
> Mike "Silby" Silbersack
>
To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-net" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?200107131406.QAA03047>