The MPLS WG Archive

[Curtis Villamizar <curtis@workhorse.fictitious.org>]

In message <39469E08BD83D411A3D900204840EC55FB70F7@vie-msgusr-01.dc.fore.com>, 
"Naidu, Venkata" writes:
> Curtis,
> 
> -> Without hierarchical LSPs, RSVP/TE does not have multipath unless 1)
> -> more than one LSP is configured to the same destination with equal
> -> cost and MP is enabled, or 2) more than one LSP is on a path to
> -> different egress and the total cost is the same, and this form of MP
> -> is both supported and enabled.  This form of MP is easy to deal with
> -> from an OAM standpoint because the only branch is at ingress.
> 
>   Very delightful to read such a good technical explanation. But...
>   what ever you said above that current MPLS signaling can't support
>   ECMP at non-ingress branch point is a limitation. From MPLS-TE
>   point of view, all the necessary information to split the traffic
>   at non-ingress can be calculated easily. Unfortunately, signaling
>   doesn't support that.

The ingress can set up two LSPs that take an identical path up to the
intended branch point.  The traffic is forwarded onto one of those two
LSP at the ingress.  Also note that the traffic split need not be
equal.  I don't see this as very limiting.

>   If given a chance, such an ECMP split can be computed by ingress
>   CSPF by finding all augmented "equal min-cost max-flow" paths to 
>   the destination. Chosing any of least/farthest such common ancestor 
>   split can be made as part of singaling decision.

Currently implementations (that I'm aware of) require the operator to
configure more than one LSP.  If there is one path with lots of
available bandwidth, both LSPs will take identical paths (obviously
though if the operator also asks that they be disjoint, they will take
very different paths, but the purpose of specifying disjoint is
different).  If there are two paths and neither can support both LSPs,
then the LSP will take separate paths.

I don't know of any implementation that creates extra LSP if it
notices equal cost paths.  A worst case (often used in testing with
the aid of simulationed topology) would be getting from one corner to
the other of a square grid.  There are an enormous number of equal
cost paths as the grid gets larger.

The case where split is completely automatic is where the total cost
to reach a given destination using two LSPs with different egress is
equal.  This is an IP only situation.  It can't occur other services.
Some implementations don't support this at all.  Some require that it
be enabled (or allow it to be disabled).

> -> QoS based on EXP can be enabled and ECMP can also be 
> -> enabled.  If each
> -> branch of the path honors the EXP bits, QoS still works and exists in
> -> the presense of ECMP, over LDP or RSVP/TE.  There is very clear
> -> existance proof of this.
> 
>   I view ECMP as a TE mechanism than a QoS workhorse. TE and QoS are
>   infact orthogonal.

That was the point I was getting at (or very close to it).  ECMP are
orthogonal.  As you point out TE and QoS are also orthogonal.  Maybe
in the long message the main point got lost.  Thanks for summarizing.

> Venkata.

Curtis