The MPLS WG Archive

["Sasha Vainshtein" <sasha@iprad.co.il>]

______________________________________________
With best regards,
Sasha Vainshtein
mailto: sasha@iprad.co.il
phone: +972-3-7659993 (office)
fax:      +972-3-6487779 (office)


>>-----Original Message-----
>>From: dnb@zurich.ibm.com [mailto:dnb@zurich.ibm.com]
>>Sent: Monday, November 27, 2000 3:47 PM
>>To: Sasha Vainshtein
>>Cc: mpls@UU. NET
>>Subject: Re: E-LSP or L-LSP
>>
>>
>>> The examples I have in mind are based on the naive understanding of the
>>> "weighted" per E-LSP admission rules.
>>> Example 1: Bandwidth on the link is pre-allocated like you have
>>said (i.e.,5
>>> MBit/s for EF, 10 Mbit/s for AF1 etc.).  However, "all" the  5
>>Mbit/s of the
>>> EF traffic in this node as well as, say 1 Mbit/s of AF1 and  1
>>Mbit/s of AF3
>>> go to the same destination. Such a flow could be satisfied by a 7 Mbit/s
>>> E-LSP with appropriate admission rules - but these rules
>>cannot, to the best
>>> of my understanding,  be derived from the overall distribution
>>of capacity
>>> of the egress link.
>>
>>Of course, there are other (and maybe better) ways to conduct
>>admission control
>>for E-LSPs. If you have another solution that works for both
>>L-LSPs and E-LSPs,
>>then I'm really eager to know how it works.
>>However, I still think that my proposal makes some sense. The
>>distribution of
>>pre-allocated resources reflects the distribution of the total expected
>>traffic. It is therefore also reasonable to request that the distribution
>>of traffic inside an E-LSP corresponds to this global
>>distribution. The system
>>would then work as follows:
>>1.) For each PSC, the amount of reservable bandwidth is maintained.
>>2.) If an L-LSP is admitted, only a single PSC applies and
>>admission control
>>    is easy. There must be some sort of policing at the edge of
>>the network,
>>    such that no more than the requested bandwidth is used. This
>>might means a
>>    re-mapping of the DiffServ Code Point.
>>3.) If an E-LSP is admitted, multiple PSC apply. Admission
>>control uses the
>>    global distribution, as described earlier, to compute the
>>requested bandwidth
>>    per OA and thus can check whether enough bandwidth for the
>>corresponding PSC
>>    is available. Also, there must be some sort of policing at
>>the edge of the
>>    network. In fact, the same global distribution is used to
>>break down the
>>    total reserved amount of bandwidth. Policing of E-LSP traffic is then
>>    be carried out per OA.
>>
>>To come back to your example: Given the global distribution and
>>the 7 Mbit/s
>>of total bandwidth for that E-LSP, then policing will kick in.
>>The distribution:
>>EF - 5, AF1 - 1, AF3 -1 would not be possible. If such a
>>distribution is required,
>>then three L-LSPs need to be set up.
Of course 3 separate L-LSPs will solve the problem. The key underlying
problem (from my point of view) is:
Can we define meaningful applications for E-LSPs? And if yes, then what are
these applications?
The problem with implcit redistibution of bandwidth between several BAs
within the same E-LSP is exactly that: it is implicit.
That means that the network operator can obtain unexpected results and then
be stuck with them. Your interpretation of example 2 illustrates this
situation perfectly: the network operator has to guess (how? and based upon
which information) that the combination of two E-LSPs is illegal and to
choose anothre tool (e.g., L-LSP). From my point of view this means that the
bandwidth redistribution rules have to become part of the standard in order
to avoid interoperability problems and ambiguity.
My gut feeling (and I may be quite wrong, I have just started working in
this area) is that E-LSPs can be used for BAs which do not imply bandwidth
reservation, e.g., for the CS (say, with a trivial CS-->EXP mapping, you
happen to have exactly the right number of bits). I have strong doubts
regarding meaningful application of E-LSPs for combining BAs with
independent bandwidth reservation, like EF and AFx.
>>
>>> Example 2: Two 25 Mbit/s E-LSPs are requested, one carrying EF
>>and AF3, the
>>> other one carrying EF and AF4. According to the "relative
>>weights" logic, up
>>> to 5 Mbit/s would be admitted to each of the two E-LSPs, with the total
>>> exceeding the overall 5% limit on EF.
>>
>>Since admission control does not work in parallel, the following
>>would happen.
>>Global distribution: EF gets 5%, AF1.x gets 10%, AF2.x gets 20%,
>>AF3.x gets 20%,
>>AF4.x gets 20%. Total link capacity: 100Mbit/s. Reservable bw for
>>EF: 5 Mbit/s.
>>First E-LSP: 25 Mbit/s -> 5 Mbit/s for EF, 20 Mbit/s for AF3.
>>This can be admitted.
>>Now, there is no reservable bandwidth for EF left.
>>Second E-LSP: 25 Mbit/s -> 5 Mbit/s for EF,20 Mbit/s for AF4.
>>Since no bw is left
>>for EF, the second E-LSP will not get admitted.
The situation when the results depend upon the order of processing (i.e., if
we start with the 2nd LSP then the 1st one would not be admitted) look to me
highly suspicious.
>>
>>Best regards,
>>
>>-Daniel
>>