Grid Datafarm Bandwidth Challenge at SC2002

Grid Datafarm team (AIST, KEK, Titech, Univ. Tokyo, Indiana Univ., and San Diego Supercomputer Center) participated in the high-performance bandwidth challenge at the SC2002 international conference. In this challenge, the network performance of Gfarm file replication was measured using 7 PC clusters in US and Japan.

At the result, we achieved 2.286 Gbps for file replication using a 12-node cluster at SC2002 booth, which is equivalent to the performance to transfer a CD-ROM data in 2.4 seconds. Between US and Japan, a distance of more than 10,000 kilometers, we achieved 741 Mbps, a record speed over the transpacific networks. The round trip time of packets is about 0.2 seconds.

For details, refer to the article [ PDF ] in the Spring issue of AIST Today in 2003.

News

InfoWorld: Japan's grid system sets speed record sending data to U.S.
http://www.infoworld.com/articles/hn/xml/02/11/26/021126hnrecordspeed.xml?s=

GRIDtoday: JAPAN'S GRID SYSTEM SENDS RECORD-SPEED DATA TO US
http://www.gridtoday.com/02/1209/100843.html

TransPAC assists with high-performance computing experiment at SuperComputing 2002 (SC2002)
http://www.transpac.org/

Network Configuration

Result

SC2002 High-performance Bandwidth Challenge

The following graph shows the 10-second average network bandwidth of Gfarm file replication from a cluster in US to a cluster in Japan. We achieved 741 Mbps out of the theoretical peak bandwidth 893 Mbps.

10-sec average bandwidth for transpacific traffic (from US to Japan)

SC2002 poster

Paper

• Osamu Tatebe, Youhei Morita, Satoshi Matsuoka, Noriyuki Soda, Satoshi Sekiguchi, "Worldwide Fast File Replication on Grid Datafarm", 2003 Conference for Computing in High Energy and Nuclear Physics, March, 2003 [ PDF Letter | A4 | Presentation slide ]