Routing via Campus
Supported by DOE MICS
SLAC has multiple connections to the Internet, and Analysis of Pinger data gathered from the monitoring site at SLAC and the monitoring site at Stanford campus indicates a few candidates for which performance could be significantly improved if routed via campus rather than ESnet.
U. Colorado at Boulder, Colorado State U., U. Texas at Dallas, U. Tennessee at Knoxville
The University of Colorado at Boulder is a BaBar collaborator and considered a key-site in the BaBar computing model. The graph shows the average daily average round trip time (that is, the average ping round trip time reported from a set of ten pings in a PingER probe averaged over a day) from mid-June to the end of July is typically in the region 70-80ms to both nodes monitored from SLAC, but around half that for the site monitored from Stanford. The primary reason for this large difference is The University of Colorado has it's primary connectivity with the vBNS network, and ESnet peers with vbns in Chicago, so packets travel around 3 times farther than if the connection was geographically direct. Stanford's CALREN2 link peers with vBNS in California.
Colorado State University is connected to the Abilene Network, and both SLAC's provider ESnet and Stanfords CALREN2 link peer with Abilene in Berkeley. Given the similarity in routes, it is suprising the round trip times differ so significantly.
The University of Texas at Dallas is another BaBar key-site and the increased traffic since the BaBar start up can be clearly seen effecting the round trip time. The unloaded average round trip time is around 70ms for 100 byte packets, but this exceeds 200ms during times of high congestion, the extra delay being caused by queueing in the routers. Esnet routes to UTDallas via the ESnet hub in Albuquerque and a T1 (1.5Mbps) connection between Los Alamos National Lab (LANL) and the University of Texas at Austin.
The round trip time to between Stanford and Rice is around 50ms and constant. If UTDallas joined Abilene and SLAC peered with Abilene via Campus, this performance could be matched. Stanford connects Rice via Abilene's OC12 (622Mbps) backbone which is typically <1% loaded.
Packet loss shows a similar pattern. At times of high congestion the present SLAC<->UTDallas link suffers 5-25% packet loss. The Stanford - Rice link suffers negligible packet loss.
Note: We are using Campus to Rice only to compare the present SLAC to UTDallas connection with how it might be if we routed through campus AND UTDallas joined Abilene. Rice is NOT one of our candidate sites.
The University of Tennessee at Knoxville benefits from being one hop beyond Oak Ridge, which is connected to ESnet. It is ESnet policy to hold onto packets as long as possible (the opposite of so-called "hot potato" routing common amongst many commercial ISPs) so packets from SLAC to UTK travel largely unhindered. Packets between Stanford and UTK travel via the Southern Cross exchange point at Georgia Tech. The difference is around 30ms faster round trip if campus traffic to UTK were to be routed via SLAC.
Revised 10 August 1999
Comments to email@example.com