From mboxrd@z Thu Jan 1 00:00:00 1970 From: Andy Furniss Date: Tue, 25 Apr 2006 23:16:29 +0000 Subject: Re: RES: [LARTC] Backlog with less rate than defined Message-Id: <444EADCD.4050505@dsl.pipex.com> List-Id: References: <002001c6689d$49b439a0$0900fe0a@LucianoNotebook> In-Reply-To: <002001c6689d$49b439a0$0900fe0a@LucianoNotebook> MIME-Version: 1.0 Content-Type: text/plain; charset="iso-8859-1" Content-Transfer-Encoding: quoted-printable To: lartc@vger.kernel.org Luciano wrote: > Hi Andy, >=20 > I=B4m not sure if I understood what you told about arp packets. > I use htb default but the problem occurs even when the default queue > rate is low (it is almost always low in rate and pps). It's still not ideal even if it's not the cause - sfq default length is=20 128 packets so if they were mtu size when it's full thats 1.5sec delay +=20 drops - and the stats show drops. class htb 1:efff parent 1:1 leaf efff: prio 1 rate 1Mbit ceil 1Mbit=20 burst 2909b cburst 2909b Sent 1113213839 bytes 9059857 pkts (dropped 61529, overlimits 0) rate 1130bps 13pps lended: 9059857 borrowed: 0 giants: 0 I would not use default on eth. Also 100mbit eth is not 100mbit at ip=20 level, which is almost what htb sees (ip+14), so 1:1 needs to be less -=20 but if children don't add up to that then it won't hurt. You could just send all unmatched ip to 1:efff with a low prio filter - tc filter add dev eth0 protocol ip parent 1:0 prio 99 u32 match u32 0 0=20 flowid 1:efff then arp will not get shaped. I notice you use handle on filters maybe OK but I usually only see it=20 with hashing or fw. >=20 > The attached files are: > Rc.local - criation of the basic queues including default > Regras.inc - criation of each queue when the user login > Queues - statistics of the basic queues Have you measured the rate another way? Andy. _______________________________________________ LARTC mailing list LARTC@mailman.ds9a.nl http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc