|
Asymmetry in throughput using bbftp
|
Our first test transfered data from a SLAC machine, evagore.slac.stanford.edu, to a Stanford machine, loggy.slac.stanford.edu, using the standard bbftp binary.
The SLAC machine is a Sun Ultra 5/10 with a single processor UltraSPARC-IIi 333MHz, with network connection via a 100 mbps/full duplex Ethernet link. The Stanford machine has also a single processor Pentium II (Klamath) 300MHz. The operating system of the two machines is:
The parameters of the transfer are the defaults for bbftp:
number of streams = 1
window size = 256K
The results obtained indicated an asymmetry:
| Stanford to SLAC (mbps) | SLAC to Stanford (mbps) |
| 42.4 +/- 6.7 | 21.2 +/- 2.6 |
As a second test, we transfered data on the link Stanford
University - Darsbury Lab, over the Internet2 backbone, where the
marked packets can be observed.
The Stanford machine is the same of the previous test, while the
machine in the UK is:
| Stanford to UK (256 KB) mbps | Stanford to UK (64 KB) mbps | UK to Stanford (256 KB) mbps | UK to Stanford (64 KB) mbps |
| 3.23 +/- 0.08 | 3.23 +/- 0.07 | 3.29 +/- 0.006 | 3.31 +/-0.16 |
| Stanford to UK (256 KB) mbps | Stanford to UK (64 KB) mbps | UK to Stanford (256 KB) mbps | UK to Stanford (64 KB) mbps |
| 17.11 +/-4.21 | 15.23 +/- 5.24 | 20.92 +/- 5.14 | 19.46 +/- 5.7 |
| Stanford to UK (256 KB) mbps | Stanford to UK (64 KB) mbps | UK to Stanford (256 KB) mbps | UK to Stanford (64 KB) mbps |
| 19.57 +/- 1.46 | 16.54 +/- 2.57 | 22.58 +/- 2.39 | 18.93 +/- 4.16 |
As a third test, we transfered data on the link from SLAC
- Darsbury Lab, in order to verify that the differences we have
observed are indeed due to the different operating systems.
The machine used to trasfer data are the same as in the previous
test, evagore and rtlin1.
The results of data transfer with different windowsizes and different
number of streams are obtained using bbftp, clearly show how the
window size change does not have any impact in the transfer rate to a
LINUX machine, while there are some variations with the Sun
operating system at the receiving end.
We checked looking at the SYN packets ecahnged between the two
systems, that the windowsizes (send and receive) on the Linux box
cannot be modified from the default of 32KB.
On the other end, the receiving windowsize on the Sun machine can be
modified and this leads to different perfomances expecially with one
single stream.
Revised
URL:
http://www-iepm.slac.stanford.edu/
Comments to
iepm-l@slac.stanford.edu