[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: Diff Serv TE
Kireeti Kompella wrote:
> 2) In my opinion, the notion of preemption *across* Classes/Class
> Types should be discussed further. Jim's and my proposals
> suggest that each class be given a definite priority, so one
> explicitly configures whether best effort trumps voice traffic
> or not, by assigning BE a higher or lower priority wrt voice.
actually, mine suggest that classes be given "different" values. An LSP
setup into priority 0 may actually have no impact on the advertised
bandwidth at another priority. In yours, one can assume remotely, that
an LSP at p=0 will take bandwidth at p=7. One never would have to
assume that, as it falls out in the observed and predicted IGP
announcements. So a more likely implementation would not "search" for
bandwidth in this manner.
> 3) The notion of "oversubscription" has been thoroughly ignored.
> Do Service Providers really want their best effort traffic and
> the voice traffic to be oversubscribed at the same level?
> Intuitively, I would think that one would want BE to be rather
> more oversubscribed (~300%) than voice (~10%)
Maybe because that can be addressed with current approaches. The
methods are less explicit, but they provide more flexibility. For
instance, the oversubscription of a link might be based on the link
speed (e.g. oversubscribe fast links more) and that would be encoded
into the maximum reservable bandwidth (and thus unreserved per
priority). Likewise, the bw requested per LSP can be manipulated. One
might understand best efforts by 33% (above), or voice (90%) of observed
utilization on the LSP. Or the ratio might be a function of bandwidth.
Along the lines of predicting IGP LSAs, I have a two questions for you.
In your draft, you mention a node A which is doing CAC on a link (L?).
Remote from this is node Z which is setting up multiple LSPs over L and
is keeping its own internal view of what is likely available by priority
on L so as to determine whether or not it should try to setup another
LSP on L (or use other links or fail, etc...). There is also the
recommendation that nodes like A holdoff on advertisements for BW
related issues by multiple seconds, even minutes. How does Z what
impact on link L some other node, say X is having?
Also, in the requirements draft, I believe there was a scenario where a
network administrator might wish to divy up percentages of the network
to classes and have those percentages remain as traffic loads the
network. Is this possible with your approach?