Packet loss rate and propagation delay can vary significantly in a
wide-area network depending on the physical span and the operating
conditions of the network. We investigated the impact of such factors
to file transfers by experimenting with round-trip times of about 1and 75 ms, and with packet loss rates about 
and 10%,
respectively, using Dummynet. In Figure 13, we measure
the download time and server miss ratio when transferring a 512MB file
over T1 and T3 links from the same server. When packet loss of 10%
and delay of 75ms are combined in out-of-order transfers, download
time over T3 links increases by an order of magnitude approaching the
level of sequential transfers. This ten-fold increase from the base
case can be attributed to the mechanism used by the congestion
avoidance algorithm to recover the congestion window at the sender.
Longer round-trip delays make the recovery to take more time and increase accordingly the wasted network bandwidth. This can be explained by the TCP operation; packet losses lead to triple duplicate acknowledgments (rather than timeouts), and the congestion window increases by at most one data segment every round-trip time [41,29]. Individual sequential transfers have already low throughput due to the disk bottleneck, and are not further affected at low load. However, raising the system load from 10% to 30% doubles the time of T3 sequential transfers, while leaving almost unchanged the out-of-order transfer time. When combining delay and loss with out-of-order transfers, disk throughput drops because data retransmissions hit into the buffer cache thus reducing disk accesses. We don't observe similar effects in sequential transfers, which provides additional evidence about the poor disk access locality of this policy.
[]
[]
 |
[]
[]
 |