[LARTC] Kernel timer frequency and HTB

[LARTC] Kernel timer frequency and HTB

[lists]

Hello,
i have a linux box which is acting as a lan router towards the internet 
doing traffic shaping.
My link is 10Mbit/s full duplex.

I have set some HTB classes with a rate of 20% (2Mbit/s) and a ceil of 
95% (9.5Mbit/s). Is such an excursion of bandwidth in the HTB classes 
feasible for HTB to control?

What Timer frequency (kernel menuconfig) is the most suitable among 250, 
300 and 1000Hz for HTB? Also, I really don't have any ideas on how the 
frequency could affect the network adapter performance. Is a high 
frequency going to have bad effects on interrupts handling? Has anybody 
any suggestions about this issue?

I have some warnings about HTB quantums being too big, but as i 
understand, those should only affect the precision of the shaping 
without undermining the shaping completely.

Thank you a lot.
jack

_______________________________________________
LARTC mailing list
LARTC@mailman.ds9a.nl
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc

Re: [LARTC] Kernel timer frequency and HTB

[Andy Furniss]

lists@antonello.org wrote:
> Hello,
> i have a linux box which is acting as a lan router towards the internet 
> doing traffic shaping.
> My link is 10Mbit/s full duplex.
> 
> I have set some HTB classes with a rate of 20% (2Mbit/s) and a ceil of 
> 95% (9.5Mbit/s). Is such an excursion of bandwidth in the HTB classes 
> feasible for HTB to control?

I would have thought it's OK, 9.5mbit may be a bit close to the limit 
depending on what yout overheads on the link are.

> 
> What Timer frequency (kernel menuconfig) is the most suitable among 250, 
> 300 and 1000Hz for HTB? Also, I really don't have any ideas on how the 
> frequency could affect the network adapter performance. Is a high 
> frequency going to have bad effects on interrupts handling? Has anybody 
> any suggestions about this issue?

I would say 1000 is best. An individual class needs a burst/cburst size 
big enough to reach its rate/ceil so you can have smaller bursts on classes.

HTB was written when Hz\x100 and unless you specify burst/quantum it will 
choose them from the rates. If you do use 1000 it may be worth having a 
look what it chose with tc -s -d class ls dev ....
> 
> I have some warnings about HTB quantums being too big, but as i 
> understand, those should only affect the precision of the shaping 
> without undermining the shaping completely.

Quantum affects the way excess bandwidth is shared - you can specify it 
on each line that has a rate/ceil if you want - minimum should be 1514 
on eth with mtu 1500. If you don't specify it HTB will use the rates to 
set it, which may not be what you want.

Andy.
_______________________________________________
LARTC mailing list
LARTC@mailman.ds9a.nl
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc