[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: I-D ACTION:draft-ayyangar-ccamp-lsp-stitching-00.txt



igor:

Igor Bryskin wrote:
Dimitri,

Looks like we are talking different languages.

The discussion was inititiated by your question: is it worth advertisng
stitching segments as FAs.

i was referring to such segment as and edge-to-edge LSP signaled using RFC 3473 - i was not under the impression this document intended to cover any other case -


At the same time you keep complaining that I bring PCE into the discussion. Who in your opinion is the user of such advertising apart from PCE?

what do you mean by "who is the user ?" ... it is imho obvious that the only "user" for a LSP segment being part of the stitching operation is the end-to-end LSP


I am saying that segments provisioned by the management plane could be used
in e2e LSP dynamic provisioning and the only way to make it possible is via
stitching. You are saying that this is not stitching because in your opinion
stitching is concerned only with segments dynamically provisioned, and there
is some other mechanism (which you were not specific about)

i think this is spelled out as part of RFC 3473

that makes possible using the statically provisioned segments.

but i do not call this "LSP stitching" as no one did ever rely on this mechanism to make use of a statically provisioned subnetwork since at the end at the discretion of its owner - the wg is involved into the present discussion because we are assuming intermediate nodes (i.e. part of the segment) being GMPLS (RFC 3473) capable


Finally:


IB>> Just to be clear. You mentioned the use of advertised node
capabilities. My point is that these advertisements are useful for path
computation but are of no use for e2e signaling

and so just to be clear if they are not useful for e2e signaling what is the purpose of advertising them for this purpose ?

Comments like these just knock me off. I always believe that any type of advertising is being done for the purpose of routing and path computation but have zero relevance for signaling. Am I not right ? Consider, for instance, the case when all EROs are fully specified by the user. Can I signal such LSP? If yes, do I need TED in this case?

obviously this is not the comment i made here - the comment is "if not useful for end-to-end multi-domain LSP what is the purpose of this advertisement"


Igor


----- Original Message ----- From: "dimitri papadimitriou" <dpapadimitriou@psg.com>
To: "Igor Bryskin" <ibryskin@movaz.com>
Cc: <dimitri.papadimitriou@alcatel.be>; "Arthi Ayyangar"
<arthi@juniper.net>; <ccamp@ops.ietf.org>
Sent: Thursday, February 17, 2005 5:33 PM
Subject: Re: I-D ACTION:draft-ayyangar-ccamp-lsp-stitching-00.txt




igor - see in-line:

[snip]


Your point/question is very simple, which is what does it give to
advertise stitching segments as FAs, is that correct?

My answers are very simple either:

1) How else are you going to use statically provisioned segments as

part

of dynamically provisioned end-to-end LSPs ?

you ask a different question from the initial issue because advertize a TE link was possible today without using stitching, however this issue also depends what do you mean by statical provisioning ?

IB>> I have segments that are provisioned/owned by the management plane

exactly - there is no *LSP stitching* anymore - we discuss about control plane operations (the intermediate segment is an entity owned by the control plane)

note: by the way, operation you are referring to is supported today
(without the procedures described as part of this document because this
has nothing to do with stitching)


2) How else you can accomplish a transition strategy when some of the
nodes that are going to be involved in end-to-end dynamically

provisioned LSPs

do not have proper signaling software?

and i did ask what is the point in applying LSP segment signaling on segment that "do not have proper signaling software"

IB>> The point is that transit nodes of the segments do not participate

and

hence do not have to support e2e signaling

well i think we should then determine the working assumptions here, intermediate nodes part of an LSP segment are GMPLS nodes as any other, otherwise the segment could itself not be setup (stitching is just a way to leave control to the head-end of the segment but there are no specific assumptions made in terms of signaling capabilities for the intermediate nodes, they are assumed to be GMPLS signaling capable nodes as any other)


The specific example is using LSRs that
do not support p2mp signaling as transit nodes in p2mp LSPs. Your point
about advertising of node capabilities is relevant only for routing -
the PCE computing p2mp tree may take in consideration the advertised

node

capabilities and assume that stitching segments may be created
dynamically if needed.

just to be clear the PCE discussion is completely outside of the present discussion - afaik the working group has (still) to provide requirements for the PCE wg -

IB>> Just to be clear. You mentioned the use of advertised node capabilities. My point is that these advertisements are useful for path computation but are of no use for e2e signaling

and so just to be clear if they are not useful for e2e signaling what is the purpose of advertising them for this purpose ?


What if I don't want them to be dynamically created and rather
have them pre-provisioned to spead up the process of p2mp tunnel setup
and decrease blocking probability?

not sure that compared to the global complexity of computing the tree this is going to be a significant improvement, but if you have the possibility to determine the bandwidth on the corr. segments i don't see here how you decrease the blocking probability (afaik propagation of segment related information is equivalent to the corresponding links advertisement)

IB>> Dynamic establishment of the stitching segments may fail

not sure to catch your point, if you think a p2p LSP establishment may fail not sure it is wise to start thinking about p2mp LSP establishment - on the other side pre-provisioning of the segment if not guaranteed within certain limits can not guarantee that you would be able to stitch it afterwards (refresh, etc.)


3) What if I want to attract certain LSPs  to use pre-provisioned
segments even if the overall cost of such LSPs would be higher than if

they would

take paths computed without consideration of these segments, but it is
desirable for the LSPs to use the segments anyway. Exanple: the

segments

are provisioned to satisfy wavelength continuity constraint and I don't

have

PCE that can perform path computation with such contraint on path

segments

between points of OEO.

as said above it would desirable to leave PCE aspects outside of the picture, but this application is really interesting because on one side i thought we had the label set to solve this issue (note: that even if a path is pre-computed explicit label control can be used to provision the segment between converting points - don't PCEs for this)

IB>> First, this is just an example of the situation when I want pre-provisioned segments to be used in e2e LSP provisioning. Second, satisfying the wavelength continuity constraint takes much more than just using the label set. You need to have: a) information about all lambdas in all isles of transparency;

this may be the case but once known explicit label control on each link is enough to assume continuity


b) very complex PCE that is capable of computing optimal path(s) that

use

the same wavelength on all links within each OEO-OEO segments satisfying

all

other optical constraints.

PCE, again PCE, but the computation of such path was possible well before even the term did exist - you mix the computation with the access to the computation result -


It is much simpler to pre-provision such OEO-OEO segments and advertise

them

as TE links, than a simple PCE would do the job.

it is the term "pre-provision" that causes the main issue here, one dimension of the problem is simplified but how many are getting complexified ? and where is the evaluation of this trade-off ?


Igor



on the other hand is how much possible pre-provisioned segments - so
overhead - are now going to be advertised if the traffic demand is
unknown ? and on the other hand, if all traffic matrix is known
beforehand is there any rationale to create these segments instead of
directly provisioning

it is interesting to see that you are populating this discussion with
PCE application - i would be more in favor in leaving the possibility of
advertising segments once a clear set is being identified rather than
the other way around



Igor





----- Original Message ----- From: "dimitri papadimitriou" <dpapadimitriou@psg.com>
To: "Igor Bryskin" <ibryskin@movaz.com>
Cc: <dimitri.papadimitriou@alcatel.be>; "Arthi Ayyangar"
<arthi@juniper.net>; <ccamp@ops.ietf.org>
Sent: Thursday, February 17, 2005 1:19 PM
Subject: Re: I-D ACTION:draft-ayyangar-ccamp-lsp-stitching-00.txt






Igor Bryskin wrote:




dimitri, see in line.

----- Original Message ----- From: "dimitri papadimitriou" <dpapadimitriou@psg.com>
To: "Igor Bryskin" <ibryskin@movaz.com>
Cc: <dimitri.papadimitriou@alcatel.be>; "Arthi Ayyangar"
<arthi@juniper.net>; <ccamp@ops.ietf.org>
Sent: Thursday, February 17, 2005 11:40 AM
Subject: Re: I-D ACTION:draft-ayyangar-ccamp-lsp-stitching-00.txt







igor - see in-line





Please see my replies (AA--->) inline.







---------> An LSP segment may be created either by

configuration

or



due to arrival of an e2e LSP setup request itself. Similar to

an

FA,



an




LSP segment may or may not be advertised as a TE link. But, if
pre-created, it could be advertised, in which case other nodes

may


compute a path over it.

Why would you want to or not want to advertise an FA ?

i understand the point on pre-created <-> advertised but this

knowledge




may be useful for nodes part of the same area (not for nodes

external



to this area)

AA -------> Absolutely ...this definitely cannot be advertised

outside



the area (domain). I think this has been explicitly mentioned.

so in case a node for inst. advertises three terminating





links with PSC-2 (one of these being the LSP segment) then a

another



node (part of the same area) receiving an incoming multi-area

PSC-2


LSP




request may start making use of this segment to join the next

border,



therefore advertisement of the LSP segment may create a

multi-hop

condition, but now once used relevance of the existence of the

segment




is not a useful information (for the area) as there is no

possibility



to make re-use of it except when the end-to-end LSP is torn

down

AA----------> I understand your point that once an LSP has been

admitted




into an LSP segment it is no longer usable by other nodes in

that

area.




But would you rather stop advertising the link at this point, if

you


were previously advertising it ? Don't you think that is a big

hammer



? E.g.




how would a head end which has indeed computed a path over that

LSP


segment differentiate this event from an LSP segment down event

where



the link is deleted from the database ? So, all the document

says

today is




that you set the unreserved bw on the LSP segment to zero. The

idea


is



to still let other nodes know that the link exists but is

unusable.


It



is




not different from a FA-LSP being consumed...in that case we

don't

stop




advertising the FA (if we were doing so previously), right ?

IB>> Completley agree with Arthi. Besides, several parallel

stitching


segments could be advertised as a single bundle - hence, using

the

advertised link by one LSP does not necessarily take away all

link's


bandwidth.

you don't understand the question, it is do we have to consider as default behavior that a pre-provisioned is to be "advertized"

IB>> My point was that I do not see any difference in this respect

between




the sticthing FA-LSP (the same layer FA-LSP) and FA-LSP created in

the


lower




layer. Besides, what do you mean by the default behaviour? The fact

whether




to advertise//remove FA TE link is a policy driven carefully

thought

through




decision, a dnagerous one that could potentially destabilize the

network.




I'd say that the default behaviour is "NOT ADVERTISE" in either

case.




now beside the fact that there are techniques to do so what would

be

the



purpose of it ? and what it the overhead that such advertisement

would


create - that can be of course decreased by bundling them -

IB>> The purpose is exactly the same as for any other FA-LSP - add flexibility in a particular layer.

which flexibility are we expecting here, this "segment" can

accommodate


exactly one incoming request -


IB>> Disagree - the segment could be a component link within a

bundle.

In



this case stitching FA TE link may accomodate multiple LSPs

additionally only nodes part of the same




area can make use of this advertisement -


IB>> Who said that sticthing segments must necessarily terminate on

domain



borders? There could be multiple reasons why a network operator could
pre-provision (dynamically or statically) LSP segments inside his

network



and advertise them (as bundles or individually) as TE links to be

used

for



specific TE purposes.

it is exactly these purposes that i am looking for




so in fact what it would allow




is the possibility to avoid creation of a segment if the edge node
receiving the request re-orients its request to the head-end for

this

advertisement
                     |
example:      ----------D----------
          |                     |
        - A ---- Segment 1 ---- B -
          |                     |
           ----------C----------
                     |

you would have a segment between A-B that could be reached from C

(the

node receiving the incoming request) decides to make use of this

segment


to reach B (so you would have C-A-B) but if this was the best path

why


not creating directly a segment C-A-B, instead of now having one

segment


C-A, the pre-provisioned A-B and probably one on top of it C-A-B ?


IB>> See my comment above. I might want to use statically provisioned
segments. I might want to use nodes that do not have proper signaling
software.

what does that mean "proper signaling software"




For instance, recall the discussions on P2MP and how we want use

legacy

LSRs



to be part of P2MP tunnels

but in this case there is a flag telling capabilities of the nodes in order to allow for dynamically trigger the segment




in case of classical FA-LSP it makes sense to advertize the FA link
because it represents a lower region LSP (with usually a given ratio

of


unreserved bandwidth that makes worth advertizing the FA link) but

in

case of a segment i do have some difficulties to excatly see where

this


flexibility would deliver ?


IB>> Again, if you imagine that several parallel sticthing segments

are

advertised as as single FA, how it would be different from the

bandwidth


usage point of view compared to advertising lower layer FA ?

issues are different - FAs are used in manner to preferentially

attract

over them because - i am still looking for the reason for attracting
over a bundle




In fact it would be even more useful, because in case of lower layer

FA

you need also



to advertise termination/adaptation capabilities, while in case of

stitching



FA no addaptation is required.

by the way you don't seem to see the issue that i am pointing out, so probably there is a need to go in more detailed examples before

drawing

the above confusing conclusion




Igor




thanks,
- dimitri.





thanks,
- dimitri.






a more technical point is related to the definition of an FA

LSP

which




per LSP-Hierarchy mandates crossing LSP region border: the

head-end


and




tail-end switching capability represent the SC of the resulting

TE


link




while intermediate node terminates the SC corr. to the

switching

type



of





the FA-LSP (e.g. creation of a [PSC-1,PSC-1] link throughout a

PSC-2



capable network with first and last link being [PSC-1,PSC-2]

and

[PSC-2,PSC-1], resp.), while in the LSP segment case we would

have


now




the creation of a [PSC-1,PSC-1] link with first and last link

being


[PSC-1,PSC-1] and [PSC-1,PSC-1], resp. so there is no region

border


crossing anymore - so here the question is about definition and
detailing the triggers

AA--------> As far as trigger for setting up an LSP segment is

concerned,





I agree that there is no longer the notion of "crossing region
boundaries". I realize that the document doesn't discuss this,

especially





given that we are doing other comparisons with FA LSPs. So, I

will

add



this discussion in the next revision. I think in case of LSP

segment


the




trigger for LSP segment setup would come from a) successful

switching



type





and switching capability match and b) some local policy on the

node


which





dictates the setting up of an LSP segment.



IB>> I have a comment here. LSP-Hierarchy is not a Bible and

could

be


challenged in many ways. FA LSP is, generally speaking, created

on

a


layer





boundary rather than on region boundary: nothing prevents

creating

a


VC4




FA





LSP that starts and stops in the middle of TDM region to carry

several



VC12





LSPs. Furthermore, stitching FA is a special case of FA when it

is

used



by





LSPs of the same layer as one where the FA-LSP was created. As

for

triggers,





there could be multiple ones for setting up/tearing down

stitching

FA-LSPs:





configuration, receiving setup request for inter-domain LSP,

other

policies.





Igor









More on a) later.

thanks,
-arthi





.


.


.


.





.




.