IEPM

Bulk Throughput Measurements - ORNL


Bulk Throughput Measurements | Bulk Throughput Simulation | Windows vs. streams | Effect of load on RTT and loss | Bulk file transfer measurements

SLAC to ORNL

SLAC to ORNL bulk thruput The measurements were made between pharlap.slac.stanford.edu and sunbird.ccs.ornl.gov at Oak Ridge National laboratory (ORNL) Tennessee. Pharlap is a Sun E4500 with 6*336MHz cpus and maximum buffer size of 8Mbytes running Solaris 5.8. Sunbird is a 2*400 MHz Intel processor running Linux 2.2.18, with 100Mbps Ethernet interface and a GE interface to the campus network. the GE interface was used for all outbound traffic, and inbound traffic to sunbird used the 100Mbps ethernet, to sunbirdj it used the GE interface. The SMP feature was turned off to improve network performance. On 9/26/01 the ORNL link to ESnet was at OC12. The pipechar from SLAC to ORNL (sunbirdj) indicates that, as expected, the route is via ESnet and the bottleneck appears to be the last hop to sunbirdj.

Measurements made on September 26 '01, from pharlap to sunbirdj (Gbps interface) are shown to the right. They indicate that the throughput maxima are around 90Mbps (the top 10% throughputs measured are above 87 Mbits/s. It is apparent that with a single stream at best one gets less than 50% of the maxima. The RTT from SLAC to ORNL was min/avg/max (std)=87.5/88.5/98.8 (3.21) msec. This corresponds to a window size of about 1100KBytes for a bottleneck of 100Mbits/s. Similar measurements from pharlap to sunbird (the 100Mbits/s interface) show that 10% of the results were over 82.9 Mbits/s and the maximum throughput measured was 85.5 Mbits/s. Thus the Gbit/s interface does not help much (e.g. maxima increased by < 10%). Tom Dunigan of ORNL writes: I think the last hop to sunbirdj is screwed up somehow. TCP slow-start from sunbirdj outbound is often very "extended". (We usually do our throughput testing to the AIX box stingray for that reason).

Pharlap gained access to ESnet at OC12 speeds (622Mbps) on November 6, '01. We therefore remeasured the throughput. The results are shown below. It can be seen that the maxima increased by about a factor of 2 to over 205Mbits/s.


Created September 19, 2001, last update November 7, 2001.
Comments to iepm-l@slac.stanford.edu