Re: P2P Modeling

[Marcus G. Daniels]

>>>>> "OD" == Owen Densmore <address@hidden> writes:

OD> Short Form: I'm looking for agent models (Swarm, RePast, Ascape)
OD> looking at bottom-up, decentralized, Small Worlds networks.  Could
OD> anyone provide a pointer or two?

Catherine Dibble has done a lot of research (and software development)
in this area.  For example, check out these slides:

   http://www.econ.ucsb.edu/~cath/SwarmFest

OD> DHCP, for example, renders most Peers "nameless" due to transient
OD> name to address binding.  They are often behind NAT gateways
OD> (Network Address Translation which lets several hosts use a single
OD> address .. but at the cost of being again without identity) or
OD> within limited ISP domains not supporting full internet
OD> connectivity (multicast, for example, is typically not available)
OD> or worse, behind a corporate firewall.

Yeah, a `Swarm' is like a private IP space that routes all outgoing
traffic through a single public IP.  Swarm `Actions' are like packets
carried on such a network topology.  And, unlike some agent-based simulation
systems, a `Swarm' has the same nice properties of a private IP space --
complex patterns of behaviors that don't temporally impact outside
`Swarms' are never seen by outsiders.

For example, suppose that a private Myrinet-connected cluster is doing
some complex finite-element simulation and it periodically makes bulk
changes to a shared landscape on a colocated file server, and that the
data on that file server gets modified at low-frequency agents on some
DSL-connected system across the country.  Further, suppose this is all
one big Swarm simulation run from some host at a third location that,
for the sake of the argument, is connected by a 33kb modem.  Swarm has
the property that time-indexed Actions local to the finite-element
simulation don't need to be sent out over the internet, except to the
extent they overlap with Actions of the low-frequency agent
simulation.  Thus, there is no central data structure that records all
events, only those events at boundary conditions.  

You can see how this impacts scalability if you think about what would
happen if everything single action in the simulation had to go through
the poor 33kb modem.

This is all a metaphor, of course.  I'm not implying that Swarm comes
equipped with the nuts-and-bolts to do this kind of distribution.  I'm
just saying that the scheduling data structures were designed with
these issues in mind..


                  ==================================
   Swarm-Modelling is for discussion of Simulation and Modelling techniques
   esp. using Swarm.  For list administration needs (esp. [un]subscribing),
   please send a message to <address@hidden> with "help" in the
   body of the message.
                  ==================================