Tuesday, March 20, 2007

Network Architectures for High End Science and Research

[Bill Johnston who is the ESnet Department Head and Senior Scientist at
Lawrence Berkeley National Laboratory recently gave an excellent presentation at the ON*Vector workshop on the future demands of high end science and research on networks. His conclusions and predictions for future network architectures very much echo our own experiences in that we are seeing traffic flows from a relatively small number of research projects dwarf the general IP traffic that emanates from our universities and research sites. The CANARIE network architecture and that of other high end networks like ESnet are now increasingly being optimized to handle these flows. The question that is being asked is whether these new traffic patterns reflect the particular needs of the high end science and research community, or will they also be a precursor to what we will eventually see in the larger global Internet? Some excerpts from his slide deck- BSA]

http://dsd.lbl.gov/~wej/ESnet/files/LHC Networking(Photonics-2007-02-26).v2.ppt

On ESnet more than 50% of the traffic is now generated by the top 100 sites — large scale science increasingly dominates all ESnet traffic. While the total traffic is increasing exponentially, the peak flow and the number of large flows is increasing.

On Esnet most large data transfers are now done by parallel / Grid data
• In June, 2006 72% of the hosts generating the top 1000 flows were involved in parallel data movers (Grid applications) • This is the most significant traffic pattern change in the history of ESnet • This has implications for the network architecture that favor path multiplicity and route diversity

Increasingly these large flows are exhibiting circuit like behaviour. Over 1 year, this work flow / “circuit” duration is about 3 months.

[As a consequence, the new Science Data Network being deployed by ESnet..]

Science Data Network (SDN) core for
– Provisioned, guaranteed bandwidth circuits to support large, high-speed science data flows – Less expensive router/switches – Initial configuration targeted at LHC, which is also the first step to the general configuration that will address all SC requirements – Can meet other unknown bandwidth requirements by adding lambdas

Probably no more than one lambda will ever be needed to carry production IP traffic