The IP Over NBMA (ION) Archive

[Jon Crowcroft <J.Crowcroft@cs.ucl.ac.uk>]

 >                Multicast Inscalability over Large Cloud


as a comp sci exponent, i'd like to ask for a proper complexity
analysis here

1/ scaling is to do with growth in cost w.r.t some parameter 

looking at address resolution, and routing and signaling as three
problem areas, w.r.t int-serv IP over NBMA nets, we have several
potential costs

network centric:
a) control messages (e,g, NHRP traffic, MARS traffic, MCS traffic, I-PNNI
traffic, RSVP to q.2931 translation stages etc)

b) memory/state (e.g. address mapping tables, route entries, temporary
NHRP entry table entries, MARS entries, topology maps etc)

user view:
c) latency for traffic (e.g. delay before users traffic flowsm,
suboptimal delay paths etc)

d) throughput,

and so on.....

now, for something to simply "not scale", it has to be exponential
(i.e. non polynomial etc etc.....) - i don't believe that is the case
here - what we have is some set of feastures that maybe O(n**2), or
other nasty growth, but we need a formal analysis of each and
everyone...


by the way, the national or world wide TV analogy is  probably a bad one
since either
i) there is a small number of laege national and world wide stations,
using dense mode - and a PVC net could be used in a large cloud
 or
ii) there is a large number of smal lstations, and a sparse mode
approach could be used, and there arte a lot of ways to map a sparse
mode PIM tree (or CBT) onto an MCS efficiently

and TV doesn't have really low latency requirements, so suboptimal
placement of the RP or core is not a problem - virtual cores within
the cloud can be placed easily.....to assign the 'center' to a
MCS.....or a hierarchical approach such as HPIM or CGBT over a
hierarchy of MCSs could be used

the real problem areas are in VR and multiplayer games......(DIS...)

cheers
j.