The MPLS WG Archive

[Matthew Meyer <mrm@gblx.net>]

Greetings,

I'd like to try to propel the Soft Preemption RESV RRO Flag
vs PERR discussion to a close.  The draft authors are in
agreement and Cisco, Juniper and Avici are all now favoring 
to implement the RESV RRO Flag solution.
    
Clearly they are not the only voices that matter so I would
like present some key points/benefits that in my eyes would
seem to make RESV RRO use more beneficial than PERR.

 1) Nodes upstream of the points of soft preemption can   
    perform preemption selection taking into account LSPs 
    already soft preempted by downstream nodes, resulting 
    in a more efficient scheme.
 2) With RRO the HE can receive an entire list in one message, not
    so with PERR. (up to hop-count minus one messages per LSP)
 3) Kicking the RRO up to the processor to look at the RESV RRO
    protection desired bit may be necessary anyway.
 4) Waiting an arbitrary amount of time to 'collect' a bunch of
    uncoordinated tinigrams is not required. When the RESV RRO
    arrives at the HE, the option of executing immediately is 
    available (with assuredness that all hops have been collected).
    With PERR you must wait.                                       
 5) Removes the need to store which soft preempted hops have already
    arrived for a given soft preemption triggered reroute.

I was asked off-line to expand on the relevance/frequency of
item (2) above. There are a few common network topologies that are
specific catalysts to a chain of preemptions along the length
of a single LSP.  String-of-pearls and ring network configurations 
are common building blocks for sections of provider networks. Being 
that a ring of point to point connected nodes in failure can be 
viewed as string-of-pearls, the two topologies possess the same  
exposure.
 
Imagine, then, a great number of LSPs transiting North America    
whose sources & destinations are inter-continental (Europe,Asia)
or intra-continental but coast to coast.  They all transit an
North American intra-continental shortest path from New York to
San Francisco that is ultimately many intermediate segments
in serial (string-of-pearls).  As the demands are primarily transit
coast to coast, the network would be sized in a similar
denomination all along the series of links according to that
aggregate demand.  Were this not the case and there was one
significant choke point, preemption would be limited to, or
at least disproportionally occurring at, the one choke point. 
Anyway, this set of transiting LSPs would be prone to (soft) 
preemption at multiple points for the same preemption inducing
event.  Using the RESV RRO (rather than the PERR) could reduce
the number of required messages in this common network construct.

Matthew