[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: CBQ vs. TBF

On Sun, 27 Feb 2000, jamal wrote:

> > 	With correct bandwidth, weight, rate, maxburst, avpkt etc
> > things seem to go correctly.  (maxburst out of sync with iface rate
> > and wrong avpkt screws the calculations)
> > 
> > 	Another helpful item when doing small allocs (16Kbit -
> > 256Kbit) on fast links (10 - 100 FD ethernet) is cranking the HZ value
> > from 100 on X86 to 1024.  IIRC the QOS stuff does its control every
> > HZ, by default, and every 10ms is not very often on a high rate
> > interface so small allocs tend to go over.  HZ=1024 is every ~ 1ms
> > (just under) instead of every 10ms.  I find this useful when I am
> > doing service based controls, ie having say 1:4 being a 512Kbit
> > class with different services going into 1:4 with a policed u32
> > filter.
> > 
> A rule of thumb is as follows:
> max achievable rate = size * maxburst * 8 * Hz
> for a maxburst of 20; size of 1000 bytes; Hz 100 (10ms timer)
> achievable rate =   1000*20*8*100 which equals 16Mbps;
> increasing Hz to 1000 makes it 160Mbps

> You can also up the maxburst; but what will happen is that you'll most
> likely start seeing on-off burst as CBQ penalizes the flow for exceeding
> its bandwith and on to allow it to send after some timeout.
> Infact as you keep increasing maxburst you should see that you get bursts
> of data going out at wire speed, then penalization by CBQ, then another
> burst at wire speed etc
> This should be a good exercise. Any volunteers?

	Already done.  When playing with low bit rate flows I kept
watching traffic levels bounce, wreaking havoc with transfers until I
got maxburst taken care of.  The machine was bursting 5 - 10 seconds
of traffic at a time over a 100Mbit iface.

> Increasing Hz to 1Khz is probably Ok for machines >= pentium pros.

	Works well, most of our Linux routers and BW managers are K6-3
450Mhz w 1MB L2 or Athlon 500Mhz - 700Mhz

> Increasing the packet size in the estimate will result in an
> over-estimation if small packet sizes are the average seen. Vice-versa if
> large packets are seen.


> Note that in Linux however, the accuracy of the bandwidth measurement can
> be improved by using a different clock source.

	Yes, this is the reason I crank HZ up, at very low and very
high flow rates 10ms does not cut it.

As folks might have suspected, not much survives except roaches, 
and they don't carry large enough packets fast enough...
        --About the Internet and nuclear war.