CAR

Fri Apr 19 00:05:44 UTC 2002

At 05:09 PM 4/18/2002 -0500, Ken Yeo wrote:
>Thanks for the explaination. Based on your email, if CAR is applied at
>router B to rate limit inbound and outbound traffics at 128kbps but the
>realtime video takes 512kbps, the traffics will transverse from
>upstream-->backbone and CAR will drop packets at 384kbps in router B. At
>that time, I guess the RTSP or H.323 application will error out and stop
>requesting traffics?

Yes, either the user will stop the application or it will time out.

>  So in order to use CAR to rate limit traffics to
>customers, we need to apply CAR at ingress and egress routers. How is
>everyone else doing that?

Most traffic ends up getting dropped at the lowest bandwidth link -- in 
your case, the link from router B to the customer. While applying CAR on 
this link will give you more control, the link queueing discpline would 
normally take care of implementing a drop policy that tries to be fair in 
some way. For links 2MB and under, WFQ works nicely on Cisco boxes and will 
at least be kinder to TCP sessions.

The reason that this is normally not a problem (and why CAR on Router A is 
probably a lot of trouble for nothing) is that TCP flows will automatically 
back off to the maximum end-to-end bandwidth they can get, and UDP apps 
that try to send too much traffic get shut down (because they don't work) 
or give up. Any other large traffic flow that persists for a long time and 
overwhelms the customer's link is probably a DoS attack.

Regards,

Mathew

----- Original Message -----
>From: "Mathew Lodge" <mathew at cplane.com>
>To: "Ken Yeo" <kenyeo at on-linecorp.com>; <nanog at merit.edu>
>Sent: Thursday, April 18, 2002 12:23 PM
>Subject: Re: CAR
>
>
> > At 11:20 AM 4/18/2002 -0500, Ken Yeo wrote:
> > >-For UDP based audio/video trafffics, if the applications use RTSP and
> > >H.323, RTCP/H.245 will signal the sender to slowdown the transmission if
>the
> > >receiver lost packets.
> >
> > No, that won't happen. Much real-time voice and video traffic is constant
> > bit-rate, though there are CODECs that offer variable bit rates within an
> > envelope of min/max bit rates. When packet loss is encountered, the
> > equipment will usually try error concealment to deal with the lost data --
> > how sophisticated this is depends on the equipment. Re-sending the lost
> > data is pointless because it will arrive too late.
> >
> > More importantly, the sender will not slow down its transmission. It can't
> > -- at any instant, there's a fixed amount of bandwidth required to carry
> > the real-time voice and video, and there's no way to magically use less
> > bandwidth. Buffering options are limited because the stream is real time
> > and end-to-end latency must be bounded. And if the required bandwidth of
> > the stream is greater than the available bandwidth over a long period of
> > time (long in this case means seconds), no amount of buffering can help
> > you. You need to abandon real-time and switch to a store-and-forward
>paradigm.
> >
> > For example, a G.729a encoded real-time voice call requires 8Kbit/sec
> > constant bit rate (excluding IP and RTP packet headers). If it can't get
> > 8Kbit/sec, the user gets crappy voice quality.
> >
> > Some streaming systems (non real-time -- the content is stored for
> > on-demand viewing) encode the video/audio at several different data rates
> > and try to guess the available bandwidth for a customer's connection. This
> > process typically happens at the start of the streaming session, However,
> > automatically switching to a lower rate later is hard to do, so it's rare
> > to see it implemented. But none of this can be done for real-time voice or
> > video.
> >
> > >Did I miss anything? How about UDP traffics that are not using
>RTSP/H.323?
> >
> > H.323 is all about call signalling. It doesn't have anything to do with
> > congestion management. Also, I think you mean H.323 (optionally including
> > H.245) paired up with RTP -- this is the usual combination for VoIP calls.
> > RTSP is typically used for streamed content.
> >
> > Other real-time UDP protocols (protocols used by games such as Half Life,
> > for example) typically use the same kind of concepts and techniques that
> > are in RTSP to detect packet loss so that they can conceal it or react in
> > some other way. But if they're carrying a real-time service that has
> > minimum bandwidth requirements, then there's no way for the sender to slow
> > down.
> >
> > Cheers,
> >
> > Mathew
> >
> >
> >
> > | Mathew Lodge                 | mathew at cplane.com     |
> > | Director, Product Management | Ph: +1 408 789 4068   |
> > | CPLANE, Inc.                 | http://www.cplane.com |

| Mathew Lodge                 | mathew at cplane.com     |
| Director, Product Management | Ph: +1 408 789 4068   |
| CPLANE, Inc.                 | http://www.cplane.com | 