RE: Reduce the bandwidth of Gig Interface

From: Larson, Chris (CLarson@usaid.gov)
Date: Tue Dec 10 2002 - 16:51:11 GMT-3


I would have to disagree that "policing TCP traffic would be disastrous." It
is done in many networks and I know Sprint does it in their Internet pops.

 If you have packet drop due to policing (or congestion or for any other
reason), the source of the traffic is going to back off on it's transmition
rate until packets are acknowledged and not being dropped any longer.
Tranmission will then increase in rate again until packets are again
droppped thereby indicating congestion and the cycle repeats or the
reciever sets an acceptable window size. This is desirable and far from
disastrous.
 Windowing mechanism's are used regardless of whether a person uses WRED,
Policing, Traffic shaping or anything else.
 WRED is not a cure for this behaviour, and this behaviour doesn't change
because you use WRED. WRED simply allows a person to set what traffic gets
dropped first. WRED does help in that it will decrease the likelyhood of
global restarts. Where for instance congestion is detected by a whole slew
of sources due to many packet drops thereby all of the sources backing off
at once.

> -----Original Message-----
> From: Robert Slaski [SMTP:robin@atm.com.pl]
> Sent: Tuesday, December 10, 2002 9:29 AM
> To: P729
> Cc: sg p; ccielab@groupstudy.com
> Subject: Re: Reduce the bandwidth of Gig Interface
>
> P729 wrote:
> > "We want to use some type of traffic policing so that the excess traffic
> > that arrives at our port (10Mb) doesnt get dropped due to the
> bandwidth
> > reduction."
> >
> > Herein lies one of the differences between policing and shaping. With
> > policing, you're allowing for a certain amount of bursting such that TCP
> > between the end-nodes polices itself to your desired commited rate.
> > Excessive bursting will lead to dropped (vs. delayed) frames. This is
> > necessary to "close the window." With shaping, you begin delaying
>
> Policing will cause disaster with TCP traffic, especially if you discard
> 99 out of 100 frames (1Gbps->10Mbps). But you can use tail-drop-WRED
> available on gigabit ports of 3550 along with aggregate policing. This
> will cause TCP traffic to be smoothed using its windowing mechanisms,
> and policing will not allow the traffic to excess 10Mb.
>
> mikrobi,
> --
> .
.



This archive was generated by hypermail 2.1.4 : Fri Jan 17 2003 - 17:21:42 GMT-3