>the technical benefits i see are reduced signalling
>bandwidth and processing and a reduced per lsp memory
>budget.
Signaling will be reduced a bit (but signaling BW is not an issue in MPLS),
but it creates another problem, which is that you need to configure your
EXP->PHB mappings (which are now more than 8) in each router. Remember, the
purpose of signaling was to eliminate the EXP->PHB configuration.
Yes, I meant that originally. Indeed, I was speaking as configuration management duties as opposed to classical signalling tasks as we know from the B-ISDN world. In this perspective, I guess that the power of configuration flexibility (optional) does not induce really a control plane heavy weight.
Sergio
--------------------------------------------------------------------
Sergio Catanzariti
Senior Project Manager, Technology Integration
France Telecom R&D
1000 Marina Boulevard Suite 300
Brisbane CA 94005
Tel. 650-875-1526
Fax. 650-875-1505
email:sergio.catanzariti@rd.francetelecom.com
--------------------------------------------------------------------
-----Original Message-----
From: Shahram Davari [SMTP:Shahram_Davari@pmc-sierra.com]
Sent: Tuesday, June 06, 2000 8:54 AM
To: 'tom worster'; curtis@avici.com
Cc: 'CATANZARITI Sergio FTR&D/TI'; mpls@UU.NET
Subject: RE: MPLS Diffserv Extensions related questions/comments
Hi,
>-----Original Message-----
>From: tom worster [mailto:tom@ennovatenetworks.com]
>Sent: Tuesday, June 06, 2000 11:05 AM
>To: curtis@avici.com
>Cc: 'CATANZARITI Sergio FTR&D/TI'; mpls@UU.NET
>Subject: RE: MPLS Diffserv Extensions related questions/comments
>
>
>From: Curtis Villamizar [mailto:curtis@avici.com]
>>
>> In message
>> <000801bfcfb5$fd44adb0$6640a8c0@tst.ennovatenetworks.com>, "tom wors
>> ter" writes:
>> >
>> > > I do not understand the "hardware cost friendly"
>> > > problem.
>> >
>> > the cost i'm referring to is that a system needs to
>> > provide memory to store an exp -> phb mapping for
>> > every lsp. in a fast implementation that may come
>> > at a noticeable marginal hw memory cost.
>>
>> If the network provider is reasonably sane, then you will probalby
>> need to map a medium to large number of LSPs into a small number of
>> EXP to PHB mappings.
>
>that's what i imagined too. which is why i question the
>design of a protocol that provides unnecessary resolution
>with a non-trivial cost attached to that resolution.
>
In the diff-ext draft, the EXP->PHB signaling is optional. If an operator
does not want to use more than 8 PHBs in his network, and wants to use
E-LSP, then the pre-configured EXP->PHB mapping can be used.
>
>> OTOH .. at least a few people built full mesh IGPs on ATM overlays so
>> assumptions of ISP sanity don't hold up universally. :-(
>
>if the requirement for a different exp -> phb mapping for
>each of a medium to large number of lsps is considered
>unreasonable then that should be taken into consideration
>in the diff-ext protocol design.
>
>i thought about a separation of ds semantics from mpls
>signalling using an "exp -> phb mapping index" which would
>be signalled in the new tlv instead of the mapping itself.
>a set of mappings would be configured and referred to in
>the by the index in the signalling messages.
>
>the technical benefits i see are reduced signalling
>bandwidth and processing and a reduced per lsp memory
>budget.
Signaling will be reduced a bit (but signaling BW is not an issue in MPLS),
but it creates another problem, which is that you need to configure your
EXP->PHB mappings (which are now more than 8) in each router. Remember, the
purpose of signaling was to eliminate the EXP->PHB configuration.
Regarding the reduction of per-LSP memory; well it all depends on what you
are comparing it to. One solution which will reduce the per-LSP memory and
still uses the signaled EXP->PHB mapping is to build a few general purpose
EXP->PHB mapping tables (in each LSR), from the first few LSP setup
messages, and use them as an index in per-LSP ILM/NHLFE tables. Using this
method, the amount of per-LSP memory is less than or equal to your
proposal.
>
>but there are other advantages. the diff-ext draft could
>be radically simplified since the distinction of e-lsp
>and l-lsp disappears -- they are just different classes
>of mappings. conceivable, the mappings could go into
>experimental rfcs and a range of indexes could be reserved
>for standardisation. such a separation of mpls signalling
>and (perhaps more controversial) ds-related semantics could
>be expeditious.
>
Diff-ext-04 has been in circulation for quite a while. And we are releasing
diff-ext-05 by Thursday. Therefore I not only I don't think such a
seperation could be expeditious, but also that would delay the whole process
by months if not a year.
Regards,
Shahram
>c u
>tom
>