Re: Question on GMPLS contentions resolution

[Guangzhi Li <gli@research.att.com>]

Hi, all:

I posted the contention resolution question 10 days ago. After some fruitful
discussions with John Drake, Eric Gray, Fong Liaw, Hang Liu, I propose the
following three rules to resolve any contention in GMPLS although there is no
signaling correctness issue. The following rules should applied in sequence to
both bi-directional and uni-directional LSPs:

(1) Higher priority LSP wins lower priority LSP.
(2) LSP with Reply (Resv in RSVP) message wins LSP with Request
(Path in RSVP) message.
(3) LSP with the higher node ID assigning the label wins the contention
if race conditions happen between two LSPs with Request messages.

Since race conditions won't happen between two LSPs with Reply messages,
the above three rules solve all possible contention scenarios.

Proposed modification of the GMPLS functional specification, text from
section 4.2:
---------------------------------------------------------------------------

By allowing bi-directional LSP setup and suggested label in the GMPLS
framework,  contention for labels may occur between two bidirectional
LSP setup requests, or between an uni-directional and a bi-directional
LSP, traveling in opposite directions.  This contention occurs when both sides
allocate the same resources (labels) at effectively the same time. If there is no
restriction on the labels that can be used for bidirectional LSPs and if
there are alternate resources, downstream label assignment will apply
and there is no contention. However, if there is a restriction on the
labels that can be used for the bidirectional LSPs (for example, if they
must be physically coupled on a single I/O card), or if there are no
more resources available, then the contention must be resolved by other
means. Although contention resolution is not a mandatory requirement for
signaling correctness, contention resolution is strongly recommended
for performance optimization. To resolve contention, the following three rules
SHOULD be applied in sequence:
(1) A higher priority LSP wins over a lower priority LSP.
(2) LSP with a Reply message wins over a LSP with a Request
message.
(3) LSP with higher node ID assigning the label wins the contention if
race conditions happen between two LSPs with Request messages.
........

The rest of the text from this section should be updated accordingly.

Thanks,

Guangzhi


>  Dear GMPLS authors and all experts:
>
> During GMPLS last call, I posted the same question on the mailing list
> without response. Please somebody spend a little time and check with the
> following example? Something seems not clear when the current GMPLS
> contention resution schemes are applied on a framework with both
> bi-directional and uni-directional LSPs. Your clarification is very much
> appreciated.
>
> Sincerely,
>
> Guangzhi
>
> -------------------------------------------------------------------------------
>
> The issue arises because contention is resolved between bi-directional
> LSPs by the node with the higher node index while for uni-directional
> LSPs, contention is resolved by the downstream node. Consider the
> following example of two nodes with paired, bi-directional interfaces
> (i.e., a transmitter/receiver pair of ports).  Node 1 with ID=100 and
> node 2 with ID = 50.  Node 1 uses label 1 for the transmitter port and
> label 2 for the receiver port; node 2 uses label 4 for the transmitter
> port and label 3 for the receiver port.   We assume that a
> bi-directional LSP requires a single I/O interface.
>
> We consider two LSPs - a uni-directional LSP (LSP A) and a
> bi-directional LSP (LSP B). Both LSPs are going from node 1
> to node 2, with the uni-directional LSP setup request arriving marginally before
>
> the bi-directional LSP.  LSP A does not use a suggested label, and thus
> is assigned a label (port) by node 2.  Label 3 is assigned,
> corresponding to label 1 at node 1.  At the same time, for LSP B (the
> bi-directional LSP) node 1 assigns label 2, with suggested label 1.
> Because node 1 has the higher node ID, node 2 will assume (due to the
> contention resolution rule for bi-directional LSPs) that LSP B wins the
> contention and thus label 3 is assigned to LSP B.  Thus label 1 at node
> 1 (label 3 at node 2) has been assigned to two different LSPs.  Both
> LSPs have "won" the contention.