Bulk Throughput Measurements - NERSC
Bulk Throughput Measurements | Bulk Throughput Simulation | Windows vs. streams | Effect of load on RTT and loss | Bulk file transfer measurements
|
Bulk Throughput Measurements - NERSCBulk Throughput Measurements | Bulk Throughput Simulation | Windows vs. streams | Effect of load on RTT and loss | Bulk file transfer measurements |
The window buffer sizes on gigatcp1 were as shown below:
ndd /dev/tcp tcp_max_buf = 4194304
;ndd /dev/tcp tcp_cwnd_max = 2097152
;ndd /dev/tcp tcp_xmit_hiwat = 16384
;ndd /dev/tcp tcp_recv_hiwat = 24576
The performances (seen below) are asymmetric, throughput being ~ 50%
better
from SLAC to NERSC than from NERSC to SLAC. This is interesting since the
machines at both ends are very similar with similar window configurations. The
measurements were repeated at different times with similar results. We used top
to identify that both machines were unloaded (typically top and iperf were the
top processes and were using < 1% of the cpu). The maxima (i.e. the top 10%
of measured throughputs) from SLAC to NERSC is about 94Mbps, and from NERSC to
SLAC is 58Mbps.
On November 6th 2001 Pharlap received access to ESnet via an OC12 (622Mbps) link. We therefore remeasured the iperf throughputs on November 7, '01. the results are shown below. The maxima are seen to have increased roughly by a factor of 4.
