Exactly. I've been saying this to vendors for the last few years, as they
try to push their QOS mechanisms on me.
I'm not in the business of trying to engineer some optimum packet discard
strategy. I would rather spend my time and money trying to minimize the
drop percentage. (I haven't been tested yet with the task of trying to
minimize or standardize latency for traffic like VOIP - might change my
tune if I was dealing with that.)
I haven't been tested yet with the task of trying to minimize or
standardize latency for traffic like VOIP - might change my tune
if I was dealing with that.
The place where QOS always made sense to me is not where they want to
put it (and it's very apropos to the current conversation.)
Specifically, at the customer edge, as opposed to the provider core.
Some folks are always going to have higher bandwidth needs than money.
We can complain all we want about how, in a perfect world, everybody
should buy more bandwidth before queue delay and congestion are a
problem, but in the end, Generic Office X's communication budget is
probably for a single T1, which can either run VOIP or be channelized
into nxDS0 for phone and 24-nxDS0 for data. In either case, it
doesn't take much legitimate traffic before you start seeing problems,
especially if you're doing VOIP, video- conferencing, web serving, and
Bob is sitting in his office listening streaming video and browsing
And sure, it would be nice if people were responsible and didn't use
bandwidth frivolously, but we're in the real world, and when people
won't behave, the system should have the option of trying to do the
right thing. So, at the edge, you prioritize voice over streaming
media over web over news over mail, and everything behaves acceptably.
*That* is where QOS really shines. For some reason, though, nobody
wants to sell it there. Must be more profitable to sell 'em a bigger
That is exactly right. The whole point of QoS is to optimize the network.
It does not do much for the customer if we let him saturate a low speed link
with nonconforming traffic and then drop it at the core. It is best to drop
the nonconforming traffic at the customer end or at very least the service
edge router. The sooner you get the nonconforming traffic off the network,
better the entire network performs.
One problem with QoS policy enforcement is that it kind of gives the
a false sense of capacity. I have seen customers that prioritize certain
and then complain because of high latency and drops on the other traffic.
is something to be said for engineering capacity to provide full performance
all traffic types. I personally engineer backbones to provide no drops and
latency for all traffic period.
QoS helps a customer control his usage but the backbone should be able to
good performance for all traffic classes otherwise you are punishing the
customer that pay for pipes big enough for all of their needs.