[e2e] QoS vs Bandwidth Overprovisioning

[Christian Huitema]

In response to Vernon's point:

> Has anyone (esp. major carriers) tried the simple, obvious approach of
> honoring the old-style IP TOS bits?
> Yes, I remember that "everyone" says that without accounting to
prevent
> cheating, "all users" would soon mark all of their packets low-delay,
> but does that actually happen?  In the absense of cheating, and as
> long as voice-over-IP and other latency sensitive and very loss
> sensitive applications are a minor fraction of the traffic, wouldn't
> old-style IP TOS do fine?  Has any large ISP tried honoring the old
> TOS bits so end users could see what happens?

In this vein, has anyone tried to deploy the "Alternative Best Effort"
proposal of Jean-Yves Le Boudec?
(http://ica1www.epfl.ch/PS_files/abe.htm). His basic idea is that there
is a simple way to prevent cheating, i.e. provide a trade-off: you may
get a lower latency if you mark the packets for "alternate best effort",
but you will also get a lower bandwidth. Indeed, the whole thing can be
combined with ECN marking, to trigger adequate bandwidth adaptation by
the VoIP application, e.g. choice of the right codec and parameters.

In fact, if we observed a real problem with latency of VoIP, we could
even implement ABE without requiring any particular marking. Just
profile the packets that fit the VoIP profile, i.e. short UDP packets,
or maybe just short packets, and fit them into the ABE queue. If I
remember correctly, Sandy Fraser had implemented something like that to
provide adequate latency for interactive traffic in the Datakit, and
reported that it was satisfactory. So, if there is a need, an ISP could
just turn on the service at some choke point, without any particular
coordination with other networks, or any change in the end systems.

On the other hand, in many existing networks, there may just not be any
need for such optimizations...

-- Christian Huitema