Hi
I share my bandwith (adsl 512/128) between 12 users. I set up simple qos script for incoming (IMQ) and outgoing traffic using htb in root, 4 classess and esfq qdisc at leafs.
If you are NATing then you need to patch/select options for IMQ, so that esfq will work properly with local addresses.
Interactive traffic goeas to class1 , http,mail etc. to class2, p2p, ftp to class3, and rest to class4. Classess divide link in proportion 20% (prio 1), 40% (prio 2), 20% (prio 3) ,20% (prio 4)
If you really want latency to come first using HTB you need to give class one almost all of the rate and make the rest borrow.
This works very well when users surfing internet (www,mail, ssh), but is insufficient when i really need low, stable latency for example for games. Games traffic is in class1.
Without QoS, while uploading some files via ftp i have pings (in my favourite game) 1000+. With my qos script my ping lower to 150 - 300, but is very unstable.
On just upload you should be able to get between your baseline ping and about 90ms with 128k and MTU 1500.
There is a tweak to HTB you can use to help accuracy -
Set HTB_HYSTERESIS 0 in net/sched/sch_htb.c
If I had 128k up I would mss clamp to say 576 any traffic classes that send alot of data. I would do this for bittorrent anyway as it it hard to shape (ingress) and reducing the packet size helps.
Is there any way to configure htb, to have good, __STABLE__ pings ( 40 - 100) while other people exploring the internet.
This can be tricky, shaping traffic from the wrong end of the bottleneck is not as easy as upstream.
I have 512 down and can stop browsing hurting latency by having a class with ceil half my rate (about 80% of link speed). Using connbytes the first 80KB of new connections go here and it has a short (16 packet) queue - so that packets get dropped and get the new connections out of slowstart before they get too fast.
The trouble is that it only works well if there is no other traffic using the bandwidth. If you share and let bulk traffic have any spare then there will be a brief latency rise when people browse big pages.
P2P and bittorrent are harder to shape on ingress than "real" servers and need lower % of bandwidth and clamping to keep the packets small.
Is use default htb parameters, maybe i must change some of them? Any suggestions?
Set burst on bulk classes low and quantums to MTU / what you clamp to.
Unless you are going to try Ed Wildgooses' DSL patch, set ingress and egress to 80% of advertised bandwidth.
Andy.
_______________________________________________ LARTC mailing list / LARTC@xxxxxxxxxxxxxxx http://mailman.ds9a.nl/mailman/listinfo/lartc HOWTO: http://lartc.org/
