[e2e] how far it queues?
Alexandre L. Grojsgold
algold at rnp.br
Fri Oct 4 08:39:47 PDT 2002
I´ve got a few answers to my first posting, that can be summed up to:
- QoS benefits are questionable;
- it is only noticeable and/or useful when congestion arises.
I did not really mean to launch a discussion on QoS - I know the subject
is highly controversial, and largely discussed before.
So, lets put QoS apart for a while, lets assume that no way exists to
actively manage router queues, and that the only way that packets can be
forwarded is by sending them in a first come first served basis. And that
no effort is done to differently serve individual flows.
Lets assume also that we run a "standard" backbone network, made out ou
routers of a popular brand and of hi capacity deterministic links. And
that looking at traffic statistics we see that instataneous mean traffic
(the traffic reported by the router itself over a 5 minutes period) never
exceeds 80% of the maximum bandwidth, event at most busy hours.
Looks like a healthy network, doesn´t it? Quite probably web browsing
users will feel happy, and have no reason to blame the core of the
Then come people with some voice and video applications and tell us that
they have some special needs to run their stuff:
- the packet transmission delay shall not exceed much the time that light
takes to cover the distance between the sender and the receiver;
- the jitter must be low, i.e., the delay must not vary much, nor very
- very few packet discards.
Ok, given these needs, what can one expect from the above mentioned happy
and healthy net? Will it fit the needs? Most of the time? Half the time?
I know that many of you will tell me "it depends" but definetly this is
not an acceptable answer.
Well, to be honest, in this area the questions are not acceptable neither.
Most of those quality demanding applications will decline from quantifying
their needs: how much of packet loss, jitter and delay will they live
But, event if they were to express precise needs ... would a network
operator be able to anticipate success of failure ?
It would be much easier to think abut this problem if we had a deeper
insight on what goes on with router queues.
Again, I am supposing an *uncongested* and sane network, where only fifo
queueing is used.
According to what many people say, even on such a network short traffic
jams can occur, and that are accomodated on queues.
But ... how often? And how big does a queue grow when it happens?
More information about the end2end-interest