Bulk Throughput Measurements - University of Florida

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

SLAC to UFL

The measurements were made between pharlap.slac.stanford.edu and iperf1.nslabs.ufl.edu at the University of Florida in Gainesville, Florida. Pharlap is a Sun E4500 with 6*336MHz cpus and maximum buffer size of 8Mbytes running Solaris 5.8. Iperf1 is an 800MHz P3 running Linux 2.4, with 4 MByte buffers and a GE interface to the campus network. On 9/19/01 the campus link to Internet 2 was at OC3. The pipechar from SLAC to the University of Florida indicates that, as expected, the route is via Internet 2 with some OC3 links.

Measurements made on September 19 '01, are shown to the right. They indicate that the throughput maxima are around 100 Mbits/s. It is apparent that with a single stream one can get 80% of the optimum throughput with windows of 512KBytes. The RTT from SLAC to UFL is min/avg/max (std)=65.1/66/220 (3.9) msec. This corresponds to a window size of about 650KBytes.

Bbcp throughput from SLAC to UFL

We used the bbcp progress reporting feature to report the incremental and cumulative throughputs from SLAC to UFL with a window size of 512KBytes and 8 streams. To save having to create and save large files the data source was /dev/zero and the data was written to /dev/null. It is seen to the right that the cumulative throughput (the magenta line) creeps up to just under 80KBytes/s in about 100 seconds(~ 60Mbits/s), and then again increases to about 87KBytes/s. If one accumulates the throughput in the intervals 16- 109 seconds (light blue line) and beyond 119 seconds (green line) separately then two distinct cumulative throughputs of about 80 KBytes/s (16-109s) and 100KBytes/s (> 119s) are observed. This may be caused by external (to the current transfer) activities, such as changes in the loading on the computers at either end or changes in the cross-traffic on the bottleneck link between SLAC and UFL. It also appears that the incremental throughput (aside from possible external activity effects) reaches a maximum after about 15 seconds.

Bbcp throughput from UFL to SLAC

We installed the bbcp executable on iperf1 (unfortunately iperf1 did not have AFS to access the bbcp program from the slac.stanford.edu cell). We also set up ssh keys so we could run bbcp automatically without requiring a password for the remote bbcp peer each time it executed. We made measurements transferring a 500MByte file with different window sizes and numbers of streams. Measurement durations varied from 63 to 2119 seconds. The measurements shown below indicate that the maxima (i.e. the top 10% throughputs) of the bbcp throughputs are about 67Mbits/s compared to about 95Mbits/s for iperf. It is also apparent that the optimum window size is 256KBytes or greater.

Sample Bulk Throughput Measurements

BACK TO MAIN PAGE

Summary of Bulk Throughput Measurements

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