[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
RE: Preemption with MAM RE:
Waisum,
Just to be clear, I am not arguing against such a model.
In fact I see it as a hybrid between MAM-as-currently-defined and RDM, so I have a rather "warm" gut-feeling about it (although I personally need to think more about it).
Also, it can be looked at as a superset of MAM-as-currently-defined, since it effectively "degrades" to MAM-as-currently-defined if you configure BC8=SUM(BC0+...+BC7).
But I am saying this is much more than a clarification to current MAM and has implications on many aspects (ie IGP extensions) we need to sort out if we want to go down that model.
Cheers
Francois
>> -----Original Message-----
>> From: Francois Le Faucheur (flefauch)
>> Sent: 26 March 2003 00:37
>> To: Lai, Wai S (Waisum), ALABS; Te-Wg
>> Subject: RE: Preemption with MAM RE:
>>
>>
>> Waisum,
>> Let me try one more time.
>> Adding some implicit bandwidth constraint (or whatever you
>> want to call that) is not a clarification to MAM. It is a
>> different model.
>> - Where MAM currently has a Maximum Number of Bandwidth
>> Constraints of 8 (ie one per CT), this model would have a
>> maximum of 9 BCs (ie one per CT plus one aggregate). BTW,
>> this in turn, would require change to the ISIS/OSPF
>> extensions which can only advertise 8 Bandwidth Constraints.
>> - You would have to add a completely new rule to the definition like:
>> SUM (Reserved (CTb) ) <= BC9, ( b in the range 0 <= b
>> <= (MaxCT - 1))
>> - All the CAC formulas would be different
>> Francois
>>
>>
>> >> -----Original Message-----
>> >> From: Lai, Wai S (Waisum), ALABS [mailto:wlai@att.com]
>> >> Sent: 25 March 2003 23:10
>> >> To: Francois Le Faucheur (flefauch); Geunhyung Kim; Dimitry
>> >> Haskin; Te-Wg
>> >> Subject: RE: Preemption with MAM RE:
>> >>
>> >>
>> >> Francois,
>> >> Preemption across CTs in MAM was presented to the TEWG, at
>> >> least as early as June 2002, when the 00 version of
>> >> <draft-wlai-tewg-bcmodel> was submitted. Dimitry's posting
>> >> also confirmed this mode of operation.
>> >> The MAM definition simply says that Reserved (CTb) <= BCb,
>> >> i.e., the reserved bandwidth of a CT is *either less than or
>> >> equal to* the bandwidth constraint for the CT. Thus, when
>> >> contention resulting in preemption across CTs occurs, the
>> >> reserved bandwidth of a CT can/may be less than the
>> >> bandwidth constraint for the CT, depending on the relative
>> >> preemption priorities of the different CTs involved.
>> >> We can add this clarification to the MAM spec.
>> >> Thanks, Wai Sum
>> >>
>> >> -----Original Message-----
>> >> From: Francois Le Faucheur (flefauch) [mailto:flefauch@cisco.com]
>> >> Sent: Monday, March 24, 2003 5:31 PM
>> >> To: Lai, Wai S (Waisum), ALABS; Geunhyung Kim; Te-Wg
>> >> Subject: Preemption with MAM RE:
>> >>
>> >>
>> >> Waisum,
>> >>
>> >> >> -----Original Message-----
>> >> >> From: Lai, Wai S (Waisum), ALABS [mailto:wlai@att.com]
>> >> >> Sent: 24 March 2003 19:11
>> >> >> To: Geunhyung Kim; Te-Wg
>> >> >> Subject: RE:
>> >> >>
>> >> >>
>> >> >> Geunhyung,
>> >> >> Thanks for your query. If preemption is disabled, then of
>> >> >> course there will be no preemption, whether within a given
>> >> >> CT, or across CTs, regardless of any BC model used.
>> >> >> However, if preemption is enabled, then there can certainly
>> >> >> be preemption across CTs in the MAM model.
>> >> >> Francois, the DS-TE Requirements document does not place
>> >> >> any restrictions on preemption in MAM.
>> >>
>> >> That is correct.
>> >>
>> >> >> As stated in Section 4.4 of the DS-TE Requirements
>> >> >> document, "if LSP1 contends with LSP2 for resources, LSP1
>> >> >> may preempt LSP2 if LSP1 has a higher set-up preemption
>> >> >> priority (i.e. lower numerical priority value) than LSP2's
>> >> >> holding preemption priority regardless of LSP1's OA/CT and
>> >> >> LSP2's OA/CT."
>> >>
>> >> That statement is correct. But note that it starts with "
>> >> *IF* LSP1 contends with LSP2". With MAM as it has been
>> >> defined for a long time in the TEWG, LSP1 can only be
>> >> contending with LSP2 if they are in the same CT. If they are
>> >> in different CT, they are each subject to independent
>> >> bandwidth constraints and therefore simply can not be in a
>> >> situation where they contend. So while preemption across CTs
>> >> is not "forbidden", it is simply not applicable.
>> >>
>> >> >>Different methods can be used to meet this
>> >> >> requirement. The MAM model takes care of it simply by an
>> >> >> implicit aggregate constraint.
>> >>
>> >> There is no such thing as "implicit constraint". There are
>> >> Bandwidth Constraints which are all pretty explicit. The
>> >> reason we've written down definitions for these models is so
>> >> that everybody knows what are the bandwidth constraints.
>> >>
>> >> MAM definition says:
>> >> "
>> >> o Maximum Number of Bandwidth Constraints
>> >> (MaxBC)= Maximum
>> >> Number of Class-Types (MaxCT) = 8
>> >> o for each value of b in the range 0 <= b <=
>> >> (MaxCT - 1):
>> >> Reserved (CTb) <= BCb,
>> >> "
>> >> MAM involves just one Bandwidth Constraint per CT, and those
>> >> are independent.
>> >>
>> >> It may be that another model which has independent bandwidth
>> >> constraint + aggregate constraint (this would basically be a
>> >> hybrid between MAM and RDM) is a reasonable idea; but this
>> >> is not MAM as it is currently defined.
>> >>
>> >> Cheers
>> >>
>> >> Francois
>> >>
>> >>
>> >> >> Other models may use an
>> >> >> explicit constraint.
>> >> >> Geunhyung, your example below reflects the correct
>> >> >> operation in MAM with preemption across CTs (assuming that
>> >> >> CT0 has the highest set-up preemption priority, and CT3 is a
>> >> >> typo of CT2).
>> >> >> Thanks, Wai Sum
>> >> >>
>> >> >> -----Original Message-----
>> >> >> From: Geunhyung Kim [mailto:geunkim@postech.ac.kr]
>> >> >> Sent: Wednesday, March 19, 2003 11:57 PM
>> >> >> To: Te-Wg
>> >> >> Subject: RE:
>> >> >>
>> >> >>
>> >> >> Hi all,
>> >> >>
>> >> >> At first, I would like to ask one question.
>> >> >> Is there preemption across CTs in the MAM model ?
>> >> >>
>> >> >> If there is preemption across CTs, there is aggregate limit
>> >> >> implicitly in MAM with preemption.
>> >> >>
>> >> >> In the Faucheur's example (link of 100 and BC0=50, BC1=50,
>> >> >> BC2=50), when the link is congested, there are only CT0 and
>> >> >> CT1 traffics.
>> >> >>
>> >> >> in the first phase, there is only CT2 traffic with 50.
>> >> >> after that, there are requests of CT0 traffic demand with 40
>> >> >> and CT1 traffic demand 30.
>> >> >> if these requests are accepted, the link capacity is divided
>> >> >> into CT0(40), CT1(30), CT3(30), because there is preemption
>> >> >> across CTs.
>> >> >>
>> >> >> However, if there is not preemption across CTs, there is
>> >> >> aggregate limit in MAM with preepmtion(neither explicit
>> >> nor implicit)
>> >> >>
>> >> >> Regards,
>> >> >>
>> >> >> Geunhyung
>> >> >>
>> >> >> None of us is as smart as all of us
>> >> >> ==========================================
>> >> >> Geunhyung Kim
>> >> >>
>> >> >> E-mail: geunkim@postech.edu
>> >> >>
>> >> >> Tel: +82-54-279-5655
>> >> >> Fax: +82-54-279-5699
>> >> >>
>> >> >> Networking & Distributed Systems Lab.
>> >> >> CSE
>> >> >> POSTECH
>> >> >> ===========================================
>> >> >>
>> >> >> -----Original Message-----
>> >> >> From: owner-te-wg@ops.ietf.org
>> >> >> [mailto:owner-te-wg@ops.ietf.org]On Behalf Of Francois Le
>> >> >> Faucheur (flefauch)
>> >> >> Sent: Thursday, March 20, 2003 12:32 PM
>> >> >> To: Lai, Wai S (Waisum), ALABS; LE ROUX Jean-Louis FTRD/DAC/LAN
>> >> >> Cc: te-wg@ops.ietf.org
>> >> >> Subject: RE:
>> >> >>
>> >> >>
>> >> >>
>> >> >> Waisum,
>> >> >> As Jean-Louis said, there is no aggregate limit with MAM
>> >> >> (neither explicit nor implicit).
>> >> >> If you have a link of 100 and BC0=50, BC1=50 and BC2=50,
>> >> >> then you may very well endup with a load of up to 150 across
>> >> >> the three CTs.
>> >> >> Hence, with MAM:
>> >> >> - you may have preemption within a CT (ie an LSP of CTx
>> >> >> may need to preempt another LSP of same CTx)
>> >> >> - you will not have preemption across CTs (ie an LSP of
>> >> >> CTx will not preempt another LSP of Cty, since those don't
>> >> >> contend for bandwidth).
>> >> >> Cheers
>> >> >> Francois
>> >> >>
>> >> >>
>> >> >> >> -----Original Message-----
>> >> >> >> From: Lai, Wai S (Waisum), ALABS [mailto:wlai@att.com]
>> >> >> >> Sent: 20 March 2003 04:13
>> >> >> >> To: LE ROUX Jean-Louis FTRD/DAC/LAN
>> >> >> >> Cc: te-wg@ops.ietf.org
>> >> >> >> Subject: RE:
>> >> >> >>
>> >> >> >>
>> >> >> >> JL,
>> >> >> >> Do you mean that there is no *explicit* constraint for
>> >> >> >> the aggregate bandwidth reserved from different
>> classes? As
>> >> >> >> I described in my previous reply below, when the
>> constraints
>> >> >> >> sum up to link capacity, there is "total isolation" with no
>> >> >> >> preemption among classes (which is not necessary of
>> course).
>> >> >> >> When this is not the case, then the link capacity will act
>> >> >> >> implicitly as the aggregate constraint. This is a natural
>> >> >> >> aggregate constraint (or an appropriately scaled aggregate
>> >> >> >> constraint in the case of overbooking) that does
>> not need to
>> >> >> >> be explicitly spelled out, right? When this aggregate
>> >> >> >> constraint is to be exceeded, then preemption among classes
>> >> >> >> will act in accordance with the definition, which says that
>> >> >> >> Reserved (CTb) <= BCb, i.e., the reserved bandwidth of a
>> >> >> >> class is *either less than or equal to* the bandwidth
>> >> >> >> constraint for the class, depending on the relative
>> >> >> >> preemption priorities of the different classes involved.
>> >> >> >> Thanks, Wai Sum
>> >> >> >>
>> >> >> >> -----Original Message-----
>> >> >> >> From: LE ROUX Jean-Louis FTRD/DAC/LAN
>> >> >> >> [mailto:jeanlouis.leroux@rd.francetelecom.com]
>> >> >> >> Sent: Wednesday, March 19, 2003 8:25 PM
>> >> >> >> To: Lai, Wai S (Waisum), ALABS
>> >> >> >> Cc: te-wg@ops.ietf.org
>> >> >> >> Subject: RE:
>> >> >> >>
>> >> >> >>
>> >> >> >> Wai Sum,
>> >> >> >>
>> >> >> >> Preemption among classes can definitively not occur with
>> >> >> >> current MAM defintion, whatever the preemtion priorities,
>> >> >> >> because there is no constraint for the aggregate bandwidth
>> >> >> >> reserved from different classes
>> >> >> >>
>> >> >> >> If you define BC2= 5M, BC1= 7M, BC0= 15M, then you can
>> >> >> >> reserve simultaneously 5M of CT2 LSPs and 7M of CT1 LSP and
>> >> >> >> 15M of CT0 LSPs.
>> >> >> >>
>> >> >> >> To allow preemtion between classes you need constraint on
>> >> >> >> the cumulated bandwidth reserved from diffrerent classes,
>> >> >> >> which does not exist in MAM.
>> >> >> >>
>> >> >> >>
>> >> >> >> Regards
>> >> >> >>
>> >> >> >> JL
>> >> >> >>
>> >> >> >> -----Message d'origine-----
>> >> >> >> De : Lai, Wai S (Waisum), ALABS [mailto:wlai@att.com]
>> >> >> >> Envoyé : mercredi 19 mars 2003 02:29
>> >> >> >> À : LE ROUX Jean-Louis FTRD/DAC/LAN
>> >> >> >> Cc : te-wg@ops.ietf.org
>> >> >> >> Objet : RE:
>> >> >> >>
>> >> >> >>
>> >> >> >> Jean-Louis,
>> >> >> >> Associated with each class-type (or simply
>> referred to as
>> >> >> >> a class in my draft, as stated in the 2nd paragraph of
>> >> >> >> Section A.3) there is a preemption priority, which together
>> >> >> >> form a TE-class. This enables preemption among
>> class-types.
>> >> >> >> "Total isolation between classes" is provided in
>> MAM only
>> >> >> >> when the bandwidth constraints for different classes add up
>> >> >> >> exactly to the link capacity. When this is not the case
>> >> >> >> (e.g., with overbooking), there will be interference among
>> >> >> >> classes. As shown in my draft, the degree of this
>> >> >> >> interference depends on the degree of bandwidth sharing,
>> >> >> >> whether preemption is used or not, and the relative
>> >> >> >> preemptin priority. This is a general property for any BC
>> >> >> >> models: the higher the degree of sharing, the less robust
>> >> >> >> the service isolation.
>> >> >> >> My view of overbooking is concerned with dimensioning a
>> >> >> >> link to carry the different classes of traffic
>> offered while
>> >> >> >> meeting service objectives. I have not explicitly used a
>> >> >> >> multiplier to scale the bandwidth of might appear to be
>> >> >> >> available and advertised, if that's what you are referring
>> >> >> >> to. But I think I have done that implicitly, so as to show
>> >> >> >> the performance impacts, and the need for a
>> judicious choice
>> >> >> >> of overbooking multipliers. Thus, my example of twice the
>> >> >> >> normal traffic (while discussed in the context of overload)
>> >> >> >> is effectively scaling with a factor of 2.
>> >> >> >> Thanks, Wai Sum
>> >> >> >>
>> >> >> >> -----Original Message-----
>> >> >> >> From: LE ROUX Jean-Louis FTRD/DAC/LAN
>> >> >> >> >> [mailto:jeanlouis.leroux@rd.francetelecom.com]
>> >> >> >> Sent: Monday,
>> >> >> >>
>> >> >> >> March 17, 2003 8:59 PM
>> >> >> >> To: Lai, Wai S (Waisum), ALABS
>> >> >> >> Cc: te-wg@ops.ietf.org
>> >> >> >> Subject:
>> >> >> >>
>> >> >> >>
>> >> >> >> Hi Wai Sum and all
>> >> >> >> I have a question regarding draft-wlai-tewg-bcmodel-01
>> >> >> >> Section A.3 :
>> >> >> >> "Preemption is enabled so that, when necessary, class 1 can
>> >> >> >> preempt class 2...."
>> >> >> >> How can you apply this to MAM ??
>> >> >> >> If I refer to 3.0 definition,MAM ensures total isolation
>> >> >> >> between classes, preemption can occurs only inside a class,
>> >> >> >> but not between classes
>> >> >> >>
>> >> >> >>
>> >> >> >> "Overbooking is allowed as it is to be described below..."
>> >> >> >> How do you define overbooking here ?
>> >> >> >> Overbooking is definitively not allowed in your RDM
>> >> >> example (BC0=15)
>> >> >> >> Regards
>> >> >> >> JL
>> >> >> >>
>> >> >> >>
>> >> >>
>> >> >>
>> >>
>>
>>