[afnog] QoS/TE

[Jim Forster]

> When discussing this I made the same points as have been raised  
> here. The
> response was that even relatively uncongested links, around 20%  
> utilisation
> upwards, can result in measurable increases in latency and jitter.  
> So if you
> have a network link with moderate utilisation (say 50%) and a  
> relatively low
> proportion of high-priority traffic (say 20% of that), the  
> performance for
> the high priority traffic can be made noticeably better.

Depends on the link speeds and distances.   Think of this way: a 1000- 
bit packet on a 1Gigabit link has a service time of 1 microsecond.    
Speed of light time for 1000 km is about 3msec.   Link utilization of  
90% leads to an average queue depth of about 10 (engineering-level  
queueing theory accuracy); 95% average utilization has an average  
queue depth of 20.     So packets sit on queue for 0-20 microseconds  
or so for the 90% case, then spend 3 msec on a 1000km link.   
Obviously this is very rough and variable sized packets add to the  
jitter, but when the service time is at least two orders of magnitude  
less than the propagation time, well, then queuing-induced jitter is  
not so important.,

So the example you cited of the large UK ISP might be true, but it's  
relevance to other situations depends on the speeds and distances.

   -- Jim