The MPLS WG Archive

[Mark Duffy <mduffy@quarrytech.com>]

 > >                       This information is sufficient for a set of LSRs
> > >    along the path followed by an LSP to discover either the exact MTU
> > >    for that LSP, or an approximation which is no worse than could be
> > >    generated with local information on the ingress LSR.
> >
> > [md] How do we define "no worse"?  I'd suggest that it should mean the
> > procedure is not more likely to discover an LSP MTU value that is higher
> > than the actual value.
>
>Basically, the path MTU discovery problem is to take the min of a set
>of values (those values being the MTUs of the involved links).  If some
>of those values are missing, the best you can do ("no worse") is to
>take the min of the values you *do* have.  As you point out, this can
>be too large.  (See below also.)

[snip]

Hi Kireeti, and thanks for responding.  Sorry about the delay in getting 
back to you.

I think the basic question is, what is the right thing to do when some of 
the Hop MTUs are missing (e.g. because the LSR upstream of the hop does not 
support the MTU signalling extensions).

Just to summarize the lengthy section I snipped...   As written, (as you 
point out above) the scheme calculates and uses the min of the values it 
does have.   This is per the [U=1, F=1] mode of LDP.  The alternative would 
be to use a [U=1, F=0] LDP behavior.

The benefit here of F=1 is that you can get some use out of the scheme even 
though not all LSRs in the path are supporting it.  This is especially 
important for a feature such as this that is being added late in the game.

The benefit of F=0 is that when the scheme is missing necessary 
information, it tells the affected upstream LSRs that it failed, rather 
than telling them a possiblly incorrect value.  This would alert the 
upstream LSRs to use a different strategy; perhaps using a 
known-conservative MTU value.

I suppose I agree with you that given a choice between the two, the first 
way (F=1) is better.  I think is is possible to get the best of BOTH by 
using another, separate TLV that is originated by the egress LSR only, and 
passed with U=1, F=0, that tests the continuity of support for the MTU 
discovery.  Then LSRs would learn an MTU for the downstream portion of the 
LSP, as well as learning whether that MTU was based on all Hop MTUs or just 
a subset of them. What do you think?  Worthwhile or not worth the extra 
complexity?

Mark