[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Comments on draft-kompella-tewg-bw-acct-00.txt



Kireeti and DS-TEers,

A few comments/thoughts:

1) Oversubscription
As Jim already clarified, other DS-TE approaches do support 
oversubscription. They assume that tunnel bandwidth can effectively 
directly be configured as "normalised bandwidth" ie an oversubscription 
factor can be applied when selecting the "bandwidth" parameter of the 
tunnel. Then nothing further has to be done to support oversubscription.

The one functional difference I can see between the two oversubscription 
approaches is that your approach allows to have differnet oversubscription 
ratios on different links (ie a given LSP may use 200Mb/s of normalised 
bandwidth on a link and use 300Mb/s on another link). Some of us discussed 
this a while back and we felt that while it may be a nice-to-have in some 
cases , it was not a strong requirement which justified the extra 
complexity of dealing with oversubcription ratios & normalised bandwidth on 
every link configuration and in every advertisement.

Do you see a strong requirement for being able to fine-tune 
oversubscription on a per-link basis for the same given LSP?
Anyone from Service Provider community could give us input on this?

Also, to size the bandwidth of a TE tunnel, I believe Service providers 
often first collect traffic measurement and then massage this according to 
their own black magic (ie look at peak or average or percentile or mixture, 
compute those over different time windows,...). All this exercise can 
pretty much be viewed as "how do I summarise all this information about 
traffic into a single representative parameter -which is bandwidth- ?". So, 
I look at all of this as already somehow trying to come up with a 
"normalised bandwidth". Since "oversubscription" is one of the many black 
magic parameters and part of the normalisation, it seems fairly natural to 
factor it as part of this exercise of deciding the size of the tunnel in 
the first place. No?


2) Migration from existing TE
It would be very useful if you could describe how you see migration from a 
network running existing TE (with existing TE advertisement) into running 
DS-TE as per your proposal.
Since you are proposing to deprecate the existing TE extensions, are you 
assuming a flag day or are you assuming that a "new" implementations 
support both extensions (and then what do I advertise in the old extensions)?


3) Bandwidth Constraints

As described in 2.2 of the Requirement draft, there are a combination of 
requirements which lead us to define "Russian Dolls" bandwidth constraints 
(ie constraint on CT2, then on CT2+CT1, then on CT2+CT1+CT0) rather than 
simple bandwidth constraint per class/Class-Types..
If I take a typical 3 class (EF/Voice, AF/Data-Premium , BE/Best-Effort) 
network, then I think it would be reasonable for QoS motivations to want to:
	- limit EF traffic on its own (eg to 60%.)
	- limit AF traffic DEPENDING on how much EF traffic is set-up: eg. limit 
AF to 80% if there is no EF traffic , but obviously limit AF below if there 
is some AF traffic (eg. limit EF+AF to 80%)
	- limit Best Effort traffic DEPENDING on how much EF and AF is set-up: eg. 
limit BE to 100% if there is no EF&AF, but limit BE below if there is EF 
and AF traffic.
(the above discussion assume we're talking in "normalised" bandwidth).
No?

I could not see how one could achieve the above with your proposal. Do you 
see a way to do this?

In other words, I think with your proposal I could limit normalised EF to 
40% , normalise AF to 40% and normalised BE to 20%. But if there is no EF 
currently, I cannot accept more AF than 40% which seems wasteful.

However, I think your proposal could be modified to achieve this without 
changes in the advertisements, simply by changing the admission control 
formulas to reflect the "Russian Doll" relationship of bandwidth constraints.


4)Multiple Computations on Head-End
I believe this is what Jim has already pointed out in one of his comments:
Not being able to predict bandwidth means a Head-End doing multiple 
computations simply computes as accurately [ or as inaccurately '^) ] as 
any other Head-End (when it could have computed more accurately).


5) Cross-Class Preemption
Just to avoid confusion I believe by "Cross-Class Preemption" you mean the 
ability to have multiple classes/class-types use the same preemption level 
(as opposed to "Preemption across Class-Types" as defined in 2.5.2 of DS-TE 
Requirements).
I agree that there is currently no identified requirement for "Cross-Class 
Preemption". As a matter of fact, it's not identified as a requirement in 
the current DS-TE Requirements document.
Input from SP on this is welcome.
Just want to clarify that DS-TE proposal happens to support it naturally if 
one wants to do it, but obviously it does not have to be used.


Cheers

Francois

PS: I'll be out of email from now on until the IETF, so I'll only be able 
to resume discussions over there. Someone will probably have come up with a 
fifth proposal by then.
_______________________
Francois Le Faucheur
Development Engineer, IOS Layer 3 Services
Cisco Systems
Office Phone:          +33 4 97 23 26 19
Mobile :               +33 6 89 108 159
Fax:                   +33 4 97 23 26 26
Email:                  flefauch@cisco.com
_________________________________________________________________
Cisco Systems Europe
Domaine Green Side
400, Avenue de Roumanille
Bātiment T3
06 410   BIOT
SOPHIA ANTIPOLIS
FRANCE
_________________________________________________________________