Summary of Bulk File Transfer Measurements
Summary
- Bbcp throughputs are similar in behavior, with windows and streams, to those reported by iperf.
- Bbcp maximal throughputs are about 70-80% of iperf's.
- Bbcp overhead (setup etc.) times are proportional to the number of streams.
- Bbcp transfer rates are symmetric from SLAC to IN2P3 and IN2P3 to SLAC.
Other measurements show it is also symmetric between
Stanford and Daresbury Lab UK, but asymmetric between SLAC and Stanford.
- Bbcp can utilize over 70% of a Gbps link.
- For modest transfer rates (< 50MBytes/s), the transfer rate to pharlap
is independent of the receiving file system. At higher rates the source and destination
file systems matter.
- For high speed LAN transfers the cpu utilization by bbcp is of the
order of 1 MHz per Mbit/s. A similar value (0.72 MHz / Mbits/s) is observed for
SLAC to CERN.
- For links with between 40Mbits/s and 500Mbits/s one can use a 100MByte file to
achieve similar throughput as an infinitely long file.
Sample Bulk File Transfer Measurements
CERN | IN2P3 |
INFN |SLAC to SLAC |
Daresbury |SLAC to Stanford
BACK TO MAIN PAGE
Back to top
Created August 25, 2000, last update August 30, 2001.
Comments to iepm-l@slac.stanford.edu
