SLAC to Daresbury Lab (UK)
We show four measurements from SLAC to Daresbury. The first was made between
08:45 and 10:03 Thursday September 7, 2000, the second was made
between 18:25 and 19:17 Thursday September 7, 2000.
The utilization of the outbound SLAC ATM link to ESnet was 27Mbps
(close to 100% - note the link is actually an ATM T3 link (~45Mbps), however
after removing the ~20% ATM "tax" for headers etc., 3.5Mbps for an ATM
PVC used for a VoIP/QoS pilot and a further 2 Mbps for an IPv6
pilot, the actual available bandwidth is between 27Mbps and 33Mbps) and
9Mbps respectively at these times.
The third graph is for thruput from SLAC
(pharlap.slac.stanford.edu) to Daresbury (rtlin1.dl.ac.uk)
made on November 2, 2000. This was after the SLAC to ESnet link had been
upgraded from the T3 link to an OC3 (155Mbps link). A slight improvement is
seen. Also it is seen that the improvement with increasing window
size is more noticeable.
The fourth graph shows measurements made July 16, 2001.
The pipechar is also available.
At this time the host at DL (rtlin1.dl.ac.uk) was a 200MHz PC running Linux with a
100Mbps Ethernet which
could not accept windows of > 64kBytes.
The SLAC to
ESnet ATM link had been upgraded to OC3 9155Mbps) from T3 (45Mbps)
the week prior to making this measurement. The maximum
iperf window settable for the host at Daresbury was 128kBytes.
See the path characteristics
measured on November 2, 2000 for more details on the route.
The path characteristics identify the bottleneck to be the 10Mbps
ATM PVC
between JAnet and ESnet in New York.
Daresbury Lab to SLAC
The version of iperf running on the Linux operating system at Daresbury did
not support windows over 128Kbytes and also the client set a window
2 times the size requested. We adjusted to this "feature" in our measurements.
The following first graph is for thruput from Daresbury (rtlin1.dl.ac.uk)
to SLAC (pharlap.slac.stanford.edu) and was made on 10/16/00.
The second graph measurements were made on July 25, 2001.
Comparing the two graphs from SLAC to DL (above) and from DL to SLAC
(below) made in July 2001, it can be seen that the maximum throughput
from DL to SLAC is higher than that from SLAC to DL.
The pipechar measured at this time is available.
The third graph
shows the asymmetry in the two throughputs as a function
of the window size * streams, where
Asymmetry = (thru(SLAC>DL)-thru(DL-SLAC))/(thru(SLAC>DL)+thru(DL>SLAC))
The fourth graph shows the frequency histogram of the asymmetry.
We later
discovered that the rtlin1.dl.ac.uk machine running Linux 2.2 was
configured to have a maximum window (/proc/sys/net/core/rmem_max and wmem_max)
size of 64KBytes. So even though Iperf was reporting window sizes of
> 64KB it was not using them.
Sample Bulk Throughput Measurements
Cal Tech | CERN |
Colorado | IN2P3 |
INFN |Daresbury |
Stanford to Daresbury |
ANL |BNL |
LBL |SLAC To Stanford's Campus
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
