[brite-users] modeling tcp connections
Thu Jul 24 17:53:00 2003
I have to solidify arguments and find references regarding modeling of point
to point connections on top of large synthetic internet topologies. I am
working on a simulator where part of the overall problem is to model tcp
connections on relatively large topologies 1x10^6 nodes +.
Obviously in these simulations, tcp low level details are irrelevant. The
main focus is to approximate physical links used when two end-hosts establish
tcp connection between themselves. BRITE generated top-down hiearchy is used
for physical links representation. I am still trying to figure out how to
use "the real internet traces".
Due to the Internet's policy-based routing and other various complexities
I thought that the best way to approximate links used in point to point
connections is a variation of the shortest path algorithm. I say variation
because the search for the shortest path from node A to node B should not
involve all the possible edges encoutered in the search. Assuming that
node A sets up connection with node B, A and B belonging to different
ASes, connection simulation setup involves:
1) node A first finds path to closest edge router having inter AS links,
i.e links to other ASes.
2) executing shortest path algorithm traverse only AS-AS edges , i.e
border router nodes until border router node from AS where node B belongs
3) reverse of 1)
This approach has one problem. There is no guarantee that path searched in
step 2 exists. Even though AS-level topology is a strongly connected
graph, set of all edge routers belonging to different ASes do not
represent a strongly connected graph. To bypass this problem I inserted
special intra-AS links between all edge routers belonging to a certain AS.
I hope this is not far-fetched approximation. It basically disregards the
details of links used within transit ASes. Transit ASes are, if you will,
shrank into one hop.
Do you see any major weakness in this approach? Do you know if anyone has
wrestled similar problems? Any references are highly appreciated.