Subject: Re: Strange network problem with buffer space, huge lag, and so on
To: None <current-users@netbsd.org>
From: M Graff <explorer@flame.org>
List: current-users
Date: 10/26/2006 13:13:43
-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Peter Postma wrote:

> What seems to happen is that the queue gets full and ALTQ starts to drop
> packets. That would also explain the slow performance (busy queue).
> You can verify this with 'pfctl -sq -v'. An extra -v will show the
> packet/byte rate for the queues. If there are indeed packets dropped,
> then you might want to increase the queue limit with the 'qlimit' keyword.

I have what I consider to be a very typical setup.  I will play with the
qlimit keyword, but none of the examples I've seen indicate that this
should be required for a single TCP stream copying out data.  I have
VERY little UDP activity (DNS lookups, NTP syncs, etc) so needing to
change a default for a single TCP stream seems that the default is a bad
choice.

BTW, with the default qlimit of 50, I get the problems I described.
Increasing this to 300 STILL causes the same problems.  I don't think
it's a queue full situation...

> not# pfctl -sq -v
> queue root_vlan9 bandwidth 1.10Mb priority 0 cbq( wrr root ) {def_vlan9}
>   [ pkts:       1460  bytes:     691173  dropped pkts:      0 bytes:      0 ]
>   [ qlength:   0/ 50  borrows:      0  suspends:      0 ]
> queue  def_vlan9 bandwidth 198Kb qlimit 300 cbq( red borrow default )
>   [ pkts:       1460  bytes:     691173  dropped pkts:      0 bytes:      0 ]
>   [ qlength:  26/300  borrows:    417  suspends:     71 ]


Running this command several times never shows a queue length above 36.

- --Michael

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.2.2 (MingW32)

iD8DBQFFQPrXuzMQWQwZDN0RAoCbAKCaD74qSvSi5KaOSx3eXbL9Z8/63ACfbL5k
DBKtdYGRh4T6CCd1sh200k8=
=qub1
-----END PGP SIGNATURE-----