Re: I-D ACTION:draft-kompella-tewg-bw-acct-00.txt

[Kireeti Kompella <kireeti@juniper.net>]

Hi Amir,

(I am moving this discussion from the MPLS WG to the TE WG.)

> A few comments regarding your proposal:

Thanks for your comments, and for your careful reading!

> 1) it seems there is a typo in your draft, section 7.3. The last two

Good catch!  Even with the current "bandwidth glut", we can't afford
not to give the bandwidth back :-)

> 2) The current approach supports configuration changes without "losing
> the math". However, in your proposal, if the administrator changes the
> subscription factor after some LSPs were set up at that priority, it
> would mean that when one of these LSPs were removed you can't properly
> calculate which B' to use in the math.

This is an excellent point -- thanks for bringing it up.  Let me work
out the whole issue of config changes and repost to the list.

> 3) With your proposal, it is possible to model the current approach with
> over-subscription (with necessary modifications to TE extensions and
> CSPF code). However, under-subscription can't be modeled (due to the
> restriction in 7.1). What are the reasons for this decision?

One could of course remove the restriction.  However, undersubscription
is merely stating that you can't use the entire link bandwidth; the max
bandwidth field gives you that control directly.  If you can see a
scenario where max bandwidth doesn't provide enough flexibility, please
let me know.

> 4) It seems that this proposal (as well as others before it) doesn't
> fully address considerations for (Diffserv) inter-class(type)
> pre-emption. Since B' is computed using subscr[p], and is used at lower
> priorities (q > p) for pre-emption, IMHO an explanation of how this is
> sensible for pre-empting between classes should be accompanied.

There are two orthogonal points here.  First, how does one compare
bandwidths at different priorities, since each priority has different
oversubscription factors?  (The answer is normalization, which is not
a real answer; however, a real answer would involve probability
functions, and I've long since forgotten what those are :-)  Besides,
the notion of oversubscription is based on pragmatics rather than
mathematics; normalization just follows that trend.)

Second, does preemption make sense?  This is up to the provider;
preemption allows higher priority LSPs to pick better paths at the
expense of some network churn.

The bottom line is, a 100Mbps link can only support 100Mbps worth of
normalized bandwidth.  If you're over that limit, some LSP needs to
go.

(I'm not sure I answered your question; if not, ask again.)

> 5) Your "Current Approach" explanation is IMO the clearest and most
> concise out there.

Thanks!  I try.

Kireeti.