DEISA-Teragrid performance

PTS Case 8

Synopsis

DEISA contacted the PERT via DANTE to request help with the poor TCP performance they were experiencing in data transfers between their site in Garching and the San Diego Supercomputer Centre (SDSC), which were being run as part of the SC2005 Teragrid tests. Between Garching (connected with a GigE to DFN's network) and the San Diego supercomputing centre the performance measured by IPerf showed that UDP could easily achieve 800Mbps (with only 0.006% packet loss) but TCP only reached around 30Mbps. The network path between Garching and SDSC, which had a Round Trip Time of approximately 180ms, was as follows:
  • Garching computer centre was connected with a Gigabit Ethernet to DFN's network. This interface was dedicated to this purpose.
  • DFN routed the IPv4 traffic to the GÉANT router, de1.de.geant.net
  • LSPs in GÉANT sent the traffic via the Amsterdam-NY link
  • ny1.ny.geant.net was connected to MANLAN (Cisco 6513) with a 10GE. VLAN 520 was assigned for this demo.
  • Internet2 (I2) carried this VLAN between NY and Chicago by using a CCC tunnel.
  • I2 was connected via 10GE to a Teragrid router in Chicago.
  • Teragrid routed the traffic to the San Diego supercomputing centre.

By increasing "tcp_sendspace" and "tcp_recvspace" DEISA were able to get some performance improvement, but the demo finished before fully satisfactory results could be achieved.

-- TobyRodwell - 31 Jan 2007

Topic revision: r1 - 2007-01-31 - TobyRodwell
 
This site is powered by the TWiki collaboration platform Powered by PerlCopyright © 2004-2009 by the contributing authors. All material on this collaboration platform is the property of the contributing authors.