From: Wes Smith (wesmith@rogers.com)
Date: Wed Mar 24 2004 - 00:04:38 GMT-3
Richard
That sounds like a perfect match to the aggresive policing problem. Or
excessive packet loss. Same result.
I've had long distance GigE Lans that were policed at 100Mbs... that
would only get 15Mbs because the policer was setup wrong.
The policer's burst size needs to be large enough to hold a 'window' of
tcp data. and factor in the latency of the network.
This is an after-effect of TCP's sliding window/rate adapation scheme.
For example .. W2K has a default tcp send window size of approx
17K.Bytes for ethernet.
It will attempt to send 17K before pausing and then waits for an Ack
from the reciever.
It will not transmit again until it gets that ack... effectively adding
'credits' to it's send window.
If a packet is dropped , then the sender never gets the ack.
It waits, times-out and then retransmits the packet(s) ......after
reducing the window size by 1/2
And so on... and so on. If you have a lossey network, you can be
throttled back to a very small trickle.
TCP will also will increase the window if packets are not dropped....
Filling the pipe until packet loss starts again.
Now comes the latency devil....
These LanExtensions can travel a fair distance, and can get upwards of
10ms or more of latency.
So ... if you have 10ms of latency...
.. and you only send one packet every time you get an ack packet back.
.... You end up 'clocking' your outbound packets based on
getting an ACK back every 10ms.
.... Or one window every 10ms
The short answer to my long winded explanation is ..
The burst needs to be at least LinkSpeed*RoundTrip time (aka ping round
trip time)
If you have 4Meg, and 10ms Latency then (4,0000,000 * .010) =
40000bits = 5KBytes
It's usually better to double this value.
Anything less than this, and TCP will start to throttle back, and you
won't reach your 4 meg.
If the provider won't change, then yes, you'll need to shape. But now
you know how to set the shaper parameters.
Hope this makes sense.
(btw. ... this is exactly the same problem that affects high
bandwidth/high latency networks., only the problem here is getting a big
enough window to 'fill' the pipe.)
Richard Dumoulin wrote:
>Thank you Wes.
>
>The symptoms are that the customer is seeing a small bandwidth. "it is going
>slower than my rtc at home" he said. But he also mentioned that he could get
>good bandwidth sometimes. At first I thought there would be something
>misconfigured in the network and finally came to the speed mismatch on the
>Internet link.
>I'll have a look at your tool and will ask the ISP how aggressive is his
>rate limiting,
>
>Regards
>
>--Richard
>
>
>-----Mensaje original-----
>De: Wes Smith [mailto:wesmith@rogers.com]
>Enviado el: miircoles, 24 de marzo de 2004 0:39
>CC: ccielab@groupstudy.com
>Asunto: Re: Traffic shaping on Cat3550
>
>
>Do you need shaping .? .. Maybe. Idealy yes, but maybe not.
>Many people don't shape Lan Extensions, because it's been hard to do
>without putting a router inplace.
>It all depdends on if you want qos (voice) over that link and on how
>much burst they have in that policer.
>
>If they have configured a small burst l, then it'll react very quickly,
>and drop packets to soon.
>TCP flows will see this as congestion, and back off their windows.
>Thruput will drop like a stone. If they have a reasonable burst size,
>maybe it'll be ok.
>But Try it. If you can get close to 3 meg thru it, and you don't need
>qos .. then you won't need a shaper.
>
>As for the 3550. I'm not aware of many switches that do shaping.
>Lots of routers and smaller machines, perhaps even the Netscreen do
>shaping at 10Meg or less.
>Not many do it for fast/GigE . Shaping requires a lot of memory as it
>implies buffering of data.
>At FastE and gig speeds ... that's a awful lot of data for a shaper....
>or a small shaper queue depth ..which sort of defeats the purpose.
>3570Metro was the first 'lower end' box I've seen... but I'm pretty
>cisco oriented. There may be others out there
>All the new switches seem to have it.
>
>btw .. the IPERF tool at http://dast.nlanr.net/Projects/Iperf/ is a
>great tool for testing this type of thing.
>You can set the windows size ... etc.
>
>Wes
>
>Richard Dumoulin wrote:
>
>
>
>>So am I right in thinking that there is a problem here ? And that it is a
>>pain that the 3550 is not able to shape ?
>>
>>Luckily there is a Netscreen in between and if I remember well it is able
>>
>>
>to
>
>
>>shape,
>>
>>Thank you for your reply,
>>
>>--Richard
>>
>>-----Mensaje original-----
>>De: Wes Smith [mailto:wesmith@rogers.com]
>>Enviado el: martes, 23 de marzo de 2004 14:57
>>Para: Richard Dumoulin
>>CC: ccielab@groupstudy.com
>>Asunto: Re: Traffic shaping on Cat3550
>>
>>
>>We had a very similar situation last year.
>>You're correct... 3550 only polices. But at least it polices output,
>>which a Cat6K Sup1/Sup2 can't do.
>>
>>At that time ... you were limited to (with cisco gear) the following
>>devices that could shape.
>>- An IOS Router
>>- Cat4K with Sup IV
>>- Cat6K with Sup720 (with the new line cards)
>>- Cat6K/7600with OSM line card (Big $$$$$)
>>- You can now look at the new 3750 Metro switch.. It's probably their
>>most affordable high speed shaper
>>
>>For 4 Mbs you could use a regular router. A 2621XM might be big
>>enough ..
>>2951/2691/3700 would be better. We tested 7301s shaping up to 150Mbs
>>with out problems.
>>Wes
>>
>>Richard Dumoulin wrote:
>>
>>
>>
>>
>>
>>>Hello group,
>>>
>>>
>>>Suppose we have a cat 3550 configured as a L3 switch connected to an ISP.
>>>The connection is at 100Mbps but the provider is rate limiting at 4 Mbps.
>>>
>>>
>I
>
>
>>>suppose the ISP is dropping all the traffic bursting beyond 4 Mbps. Also I
>>>suppose it is very easy to burst beyond 4 M if the link is configured at
>>>
>>>
>>>
>>>
>>100
>>
>>
>>
>>
>>>M ?
>>>
>>>What could you suggest as a solution. I would like Traffic Shaping but
>>>
>>>
>with
>
>
>>>Cat3550 it's only possible to do Traffic Policing no ?
>>>
>>>Thanks,
>>>
>>>Richard Dumoulin
>>>
>>>_______________________________________________________________________
>>>Please help support GroupStudy by purchasing your study materials from:
>>>http://shop.groupstudy.com
>>>
>>>Subscription information may be found at:
>>>http://www.groupstudy.com/list/CCIELab.html
>>>
>>>
>>>
>>>
>>_______________________________________________________________________
>>Please help support GroupStudy by purchasing your study materials from:
>>http://shop.groupstudy.com
>>
>>Subscription information may be found at:
>>http://www.groupstudy.com/list/CCIELab.html
>>
>>
>
>_______________________________________________________________________
>Please help support GroupStudy by purchasing your study materials from:
>http://shop.groupstudy.com
>
>Subscription information may be found at:
>http://www.groupstudy.com/list/CCIELab.html
>
>_______________________________________________________________________
>Please help support GroupStudy by purchasing your study materials from:
>http://shop.groupstudy.com
>
>Subscription information may be found at:
>http://www.groupstudy.com/list/CCIELab.html
This archive was generated by hypermail 2.1.4 : Thu Apr 01 2004 - 08:15:46 GMT-3