Bulk Thruput Measurements-Stanford to Daresbury

Bulk Thruput Measurements | Bulk throughput simulation | Windows vs. streams | Effect of load on RTT and loss | Bulk file transfer measurements

Stanford University to Daresbury Lab

Stanford has a different route to Daresbury and back than for SLAC to DL and back, since Stanford uses Internet 2 rather than ESnet as its carrier. This route is very interesting since we hope to try out the QBone Scavenger Service (QBSS) on it. We made measurements from 2 hosts at Stanford. Elaine was a Sun Sparcv9 single 360MHz processor with a 100Mbps Ethernet interface running Solaris 5.7. Loggy was an Intel processor running Linux 2.2 We also measured the pipechar on July 27, 2001 from loggy.stanford.edu to rtlin1.dl.ac.uk. The main bottleneck appears to be in the last hop entering rtlin1. Before that there is a bottleneck of about 45Mbps in the UK. This is similar to the pipecahr from pharlap.slac.stanford.edu to rtlin1.dl.ac.uk measured around the same period. The iperf measurements are shown to the right. The iperf throughput behaviors and maxima are seen to be similar to one another and to those seen from DL to SLAC (see above).

Daresbury to Stanford University

We measured the iperf TCP throughput for various windows (8k, 16k, 32k, 64k and 128k) and streams from rtlin1.dl.ac.uk to loggy.stanford.edu and elaine21.stanford.edu. Rtlin1 and loggy were running Linux and elaine21 was running Solaris 2.7. As mentioned above the window sizes we could choose from were limited to be <= 128KBytes by iperf running on Linux. The results are shown to the right. The maxima reached are higher than for Stanford to DL (see above). It also appears that in both the Stanford to DL and DL to Stanford cases that the maxima and behaviors are similar for Solaris and Linux hosts.

Comparing the various iperf TCP throuhputs between SLAC or Stanford, and Daresbury Lab we get the table to the right. In this table we use various statistical measures to characterize the ditributions. Stdev stands for the standard deviation, the rows labelled with a % are the percentiles, and IQR is the Inter Quartile Range (or the distance between the 25% and the 75%). The column headings indicate the source and destinations, the source is before the > sign and the destination after. SU = elaine.stanford.edu, SUL = loggy.stanford.edu, SLAC = pharlap.slac.stanford.edu, DL= rtlin1.dl.ac.uk The column labelled "Avg DL > US" refers to the averages of the measures made from DL to either Stanford or SLAC. The column labelled "Sce Linux / Solaris" is the ratio of "SUL>D2" over "SU/DL", i.e. the ratio of throughputs from a Stanford Linux client to DL over a Stanford Solaris client to DL. The column labelled "Dst Linux/Solaris" is the ratio of "DL>SUL" over"DL>SU", i.e. the ratio of throughputs from the DL to Stanford Linux server over from DL to a Stanford Solaris server. It can be seen that throughput from DL to the US (Stanford or SLAC) is better than the inverse, especially for the higher percentiles, i.e. the higher measured throughputs.

Sample Bulk Throughput Measurements

BACK TO MAIN PAGE

Summary of Bulk Throughput Measurements

Created August 25, 2000, last update August 29, 2001.
Comments to iepm-l@slac.stanford.edu