Andy Furniss wrote:
It's a big script - I only looked briefly and may have missed things.
Shaping inbound traffic from the wrong end of the bottleneck is hard, HTB would need to be predictive to do it better.
So what should I change?
For SFQ you can change queue length (SFQ_DEPTH) in net/sched/sch_sfq.c or you could use esfq and choose length with options.
I made this. SFQ_DEPTH was changed from 128 to 16.
How many users do you have?
I have 58 users and it will be more...
Hmm - if I had to do this many users on 2meg I would consider a different approach.
Assuming I read the script correcty, you are putting interactive traffic whithin each users share so there will be times when you delay while others get their rate.
The way I do it is to have a class with a high rate and high prio for interactive - but I do not do per user fairness on it - only on bulk traffic.
Doing this with lots of users will mean you really have to be careful about not letting bulk into interactive class. It also means that a user could get more than there share of the link by having bulk and interactive traffic. As a policy the Idea that on a highly contended link interactive should get > than fair share at any time is attractive to me as a gamer. It means that users that 24/7 download / browsers can't leave me with so little bandwidth that gaming is impossible.
If you want to do things so that interactive takes from bulk per user and keeps low latency you will need to look at HFSC.
If you have 2mbit link then I think you need to reduce the ceils or you won't have a queue to shape with.
I tested real speed - without packet queue (minimum ping, maximum transfer).
I have 2mbit link but 2Mbit for download and 256kbit for upload. 2000kbits without queue (even more:P)... and 220 upload without queue.
You need to back off a bit from the inbound link speed so that a queue builds up. There are other tweaks I would do but will post later when I know what you do/use.
What sort of DSL do you use (pppoa or e etc) and can you get a cell count out of the modem?
I assume a miss is traffic squid gets from the net, if unshaped this will make you go overlimits - so is not neccessarily just a secondary problem.
Squid is switched off. I will not use, as long as I will not make filters.
If you are NATing and have IMQ hooking after NAT in PREROUTING then traffic headed for squid will still have your real IP and forwarded local IP. You need to use u32 to seperate it.
Can You tell me how?
I think it should work OK as you do in the script with u32 for local IPs.
Thanks for answer. I thought, that nobody will write. _______________________________________________ LARTC mailing list / LARTC@xxxxxxxxxxxxxxx http://mailman.ds9a.nl/mailman/listinfo/lartc HOWTO: http://lartc.org/
_______________________________________________ LARTC mailing list / LARTC@xxxxxxxxxxxxxxx http://mailman.ds9a.nl/mailman/listinfo/lartc HOWTO: http://lartc.org/
