The MPLS WG Archive

[Alex Zinin <azinin@cisco.com>]

Yangguang,

I'll take the liberty to answer your questions below.

> Jonathan,

> Good work.  I have no major concerns on other issues except the control channel
> management.

> First, it's not only control "channels", it's control "NETWORK". The control
> network has many strong requirement regarding reliability, security, performance
> and scalability. It has to be managed and operated as a "network" not a
> collection of individual control channels.

The notion of the control channel does not substitute the notion
of the control network. In fact, the control channel between two NEs
uses the control network for transport services, i.e. LMP messages
associated with control channels and bundles are delivered in IP
packets.

> Control channel "management" (as mechanisms used by BGP, LDP, etc.) actually
> only provide health check of the control "session" between neighbors. It's not
> as fast as low layer protocols. Again, the control "network" should take care of
> all requirements and can do much better. Control channel(session) "management"
> only serves as an add-on checking.

In cases where the control channel spans more than one link,
the health check and survival of the control network is insured by
IGP protocols (OSPF/ISIS). In cases where LMP CC is between directly
connected nodes, LMP Hello process helps faster detection of
unreachable neighbors.

> Regarding your statement "Each bundled link is identified as described in
> [KRB00] and each bundled link MUST have an associated control channel". Is this
> a product specific requirement?

Which part are you referring to?

> In order for LMP to be a generic protocol, as I
> have said many times, control network should be considered independent of
> transport network. There should be no assumption of any correlation between
> these two networks. Back off a step, would you give me your reasons for this
> requirement? Why node level control session is not enough? A NE can have tens of
> bundles, too many control sessions hurt scalability and are not necessary.

LMP assumes (and it is in the spec, I believe) that more than one
bundle can be assigned to a single control channel, so there may be a
single control channel between two NEs and it can be used to manage
all component links.

Hope this helps.

Alex.


> Regards,

> Yangguang 



>> 
>> *Separated out the functionality into a base core set (Control channel
>> management & Link property correlation procedure) and an extended set of
>> *negotiable* procedures (Test & Fault isolation procedure).  The optional
>> procedures are negotiated as part of the generalized Config message
>> exchange.
>> 
>> *Updated the link finite state machine (FSM) - now separated into an FSM for
>> the control channel and an FSM for the data-bearing links.