From: Menga, Justin (Justin.Menga@xxxxxxxxxx)
Date: Wed Nov 21 2001 - 03:02:07 GMT-3
CAR uses excess burst to define a level to which you can 'borrow'
bandwidth. Borrowing implies you must 'pay back' what you have
borrowed.
You only ever need to borrow bandwidth if an application is bursting to
a point that exceeds the burst size over the time measurement interval.
In your example, you have a rate of 8Mbps and a burst of 16000 bytes.
This means that every 16ms you can send up to 16000 bytes (with no
excess burst feature). With excess burst, you can exceed this level.
The algorithm does not work on the simple direct correlation of
borrowing up to the excess burst (24000 bytes). The algorithm rather
skews the calculation by compounding your borrowed bandwidth (summing
the debt you have incurred). Once you have borrowed up to a compound
debt total of 24000 bytes, you cannot exceed the burst (still can send
normal burst) and compound debt is set to zero. Actual debt remains the
same - if you reach the excess burst in actual debt, ALL traffic is
dropped until debt is reduced to zero.
At the end of the day, what it means is that excess burst only affects
instantaneous bandwidth. With no excess burst, in your example the
bandwidth measured to 16ms accuracy will be 8Mbps. With excess burst,
the bandwidth measured to 16ms accuracy could be > 8Mbps (e.g. if you
sent 32000 bytes, instantaneous bandwidth would be 16Mbps). However,
over a longer window (say 100ms), the bandwidth will be only 8Mbps due
to the proviso that you must pay back borrowed bandwidth. Increasing
the excess burst size means you can send at the higher bandwidth rates
for a longer time period, but still your bandwidth rate over time will
never exceed 8Mbps.
Excess burst is important for TCP applications - because TCP slow start
dramatically reduces bandwidth consumption, if you have no excess burst,
a TCP application will send up to 100% of the rate, then detect
congestion and slow to approximately half bandwidth and slowly ramp up
again until 100% is reached again. Because of the lulls, the full
bandwidth of the pipe is never used. However excess burst allows you to
exceed the desired rate (let's say up to 125%), until debt is exhausted
and congestion detected. TCP slow start ramps down, but say to 75% and
then ramps up again. If you average out the bandwidth over a single
slow start cycle, you will get very close to the full desired rate.
Cisco recommends you calculate 1.5s worth of traffic at the desired rate
for burst, and twice the burst for excess burst. Thus in your example,
1.5s of traffic at 8Mbps s 1500000 Bytes. From this your burst would be
1500000 and excess burst 3000000. (I'm not sure if this formula is so
good for higher rates - setting these values means the router may have
to queue large amounts of data and may result in tail drops).
Regards
Justin Menga CCIE#6640 CCDP CCNP+Voice+ATM MCSE+I CCSE
Network Solutions Architect
Wireless & E-Infrastructure
Compaq Computer New Zealand
DDI: +64-9-918-9381 Mobile: +64-21-349-599
mailto: justin.menga@compaq.com
web: http://www.compaq.co.nz
-----Original Message-----
From: tom cheung [mailto:tkc9789@hotmail.com]
Sent: Wednesday, 21 November 2001 3:35 p.m.
To: albert_ccie@yahoo.com; ccielab@groupstudy.com
Subject: Re: CAR Parameters Confirmation
Correct me if I'm wrong. The way I understood it is that excess burst is
the
burst above cir. In your example, anything above the execss, in your
case,
cir + (24000 x 8) bps will get dropped.
>From: "Albert Lu" <albert_ccie@yahoo.com>
>Reply-To: "Albert Lu" <albert_ccie@yahoo.com>
>To: <ccielab@groupstudy.com>
>Subject: CAR Parameters Confirmation
>Date: Wed, 21 Nov 2001 12:22:28 +1100
>
>Hello Group,
>
>Could someone tell me for sure what the 3rd parameter of the rate-limit
>command is? I know it is excessive burst, but is it excessive burst
>above the committed burst (16000) or excessive burst from 0?
>
> rate-limit input 8000000 16000 24000 conform-action set-prec-transmit
>5 exceed-action drop
>
>From frame relay, the Be is on top of the Bc. From the archives, I saw
>a posting that made a good argument that it for CAR, the excessive is
>relative to 0.
>
>http://www.groupstudy.com/archives/ccielab/200111/msg01102.html
>
>Thanks
>
>Albert
>
>
This archive was generated by hypermail 2.1.4 : Fri Jun 21 2002 - 06:45:20 GMT-3