Re: [rtcweb] Pictures of congestion control on the Internet - which is more realistic?

["Karl Stahl" <[e-mail]>]

-----Ursprungligt meddelande-----
Från: rtcweb [mailto:[e-mail]] För Harald Alvestrand
Skickat: den 23 april 2014 14:31
Till: [e-mail]
Ämne: [rtcweb] Pictures of congestion control on the Internet - which is
more realistic?

I quote Karl's piece below, because I want other's opinion of whether it's a
realistic picture.

My impression is that it is not - in particular that:

- Lots of interactive traffic goes over the Internet today without layer 2
separation

[Karl] I am NOT for more level 2 separation (was describing it though) - we
are on the level 3 Internet! We can do equally well with level 3 QoS
(diffserve and reservation, but for that we need the traffic info in each
packet, unchanged end-to-end, so network elements can act upon it.)

[Karl] Today we mostly only have level 4 QoS over the Internet. We need
level 3 QoS for WebRTC. 

- All UDP endpoints that use the open Internet have some form of backoff on
congestion - using delay, packet drop or (most likely) both to detect
congestion

[Karl] Unfortunately that is not very effective. With level 4 QoS, UDP
backing off means that TCP flows take that bandwidth instead (today's
Internet traffic is very TCP-flow intensive), and the congestion is back
again. (Had we level 3 QoS, and starting to fill the whole pipe with
priority data, then priority data backing off would be effective. Otherwise,
the improvement is small, if any.) 


I'm writing the transport document based on those assumptions - that we
should specify something that will work as well as it can over an Internet
that does not care, and try to allow for better things to come along later.

[Karl] Two things to address for achieving that:

1. Doing file transfer in the data channel bundled with real-time traffic in
the same flow/5-tuple, destroys the level 4 QoS we already have over the
Internet. Getting file transfer out of the bundle restores that. SCTP's prio
mechanism cannot restore that when the congestion is in the network (only
when in the endpoints) 

2). To "try to allow for better things to come along" we need the traffic
info (bandwidth requirement and traffic type) in every priority packet,
readable at the network level, unchanged end-to-end. Then network elements
can be programmed to act upon that traffic info to give us level 3 QoS.
(When routing, the priority packets are simply put first in the out queue,
i.e. never dropped, but TCP file transfer flows will have packets discarded
and be backed off. Reservation type of networks will have to do some
calculations also.) It is easy to put in that traffic info now - very
difficult to do it later.


Am I wrong to write it like this?

Harald



------------------------------------
[Karl] I wrote this sometime before, regarding how QoS over the internet
works in general - may be useful for understanding

For better understanding of these QoS/QoE problems, and methods for
improving (not specifically discussing the policies of sharing real-time
traffic space), I would like to explain:

Over an IP pipe only used for real-traffic (no TCP data traffic), it is
sufficient that the pipe is wide enough for good QoS. That is often used and
implemented by separating IP pipes at a lower level using e.g. Ethernet
VLAN, MPLS, Ethernet over ATM (std for ADSL modems). TURN servers can
enforce real-traffic into such pipes and QoS is achieved. Let’s call this
Level 2 QoS (Network level)

Over an IP pipe shared between quality requiring real-time traffic and less
demanding data or streaming (usually TCP) traffic, we have the sharing
between these two traffic classes to consider. The main method (and the
mechanism making today’s real-time QoE as good as it often is) is that TCP
endpoints back-off and share their bandwidth usage. Real-time traffic using
UDP transport do not back-off, the endpoints using UDP occupy the bandwidth
needed. When an IP pipe gets filled, it is all endpoint’s TCP bandwidth
usage that is back-off and shared between them, leaving room for the UDP
traffic. This is mechanism we experience everyday over the Internet, using
our “Surf IP pipes”. Let’s call this Level 4 QoS (Transport level)

However, this Level 4 QoS is based on that at congestion times (which happen
every time we click – setting up a TCP flow and transferring some amount of
data as quick as possible) the router handling the most narrow part of the
pipe (the congestion point) drop packets. It is this packet dropping that
(i) signals to TCP endpoints to reduce their bandwidth (via TCP’s error
correction/retransmission mechanism) and (ii) destroys the QoE of real-time
traffic. (Both TCP and UDP packets are dropped in this process that is
triggered by flow intensive TCP traffic.)

We need (i) but don’t want (ii) and to improve on this we can e.g. use
diffserve, DSCP bits in IP packets to instruct routers to always forward the
real-time traffic before any unmarked TCP traffic (which usually fills most
of the pipe). Then QoS is then achieved for real-time traffic. This is Level
3 QoS (IP level). The method used is “traffic shaping”: backing off data
traffic, leaving real-time traffic free passage without packet loss.

_______________________________________________
rtcweb mailing list
[e-mail]
https://www.ietf.org/mailman/listinfo/rtcweb