Other factors that exert an effect on throughput are the latency and window sizes
employed by the TCP protocol. The end-to-end acknowledgment model of TCP entails
an inverse dependency between throughput and latency. This dependency, by extension,
superimposes an inverse relationship envelope over the achievable throughput, as can be
seen from the figure below.
Figure 2: StorTrends Transfer Profile vs. TCP Transfer Profile
Based on the data above, it is clear that for short distances, where latencies are in the sub-
ms range (for example, in LAN environments), the throughput is dictated entirely by the
link bandwidth. However, the situation is quite the opposite as the latencies increase.
Generally speaking, as bandwidth speed increases, the effect of latency becomes more
pronounced. For example, a 1GB data transfer via satellites using 45mbps bandwidth
could take more than 24 hours, as opposed to just the few minutes it would have taken if
the latencies were minimal. Due to the nature of the TCP protocol, replication over long
distances is essentially bottlenecked by latencies, and higher bandwidth transports do
very little to help. The result of this bottleneck effect is a gross under-utilization of
available bandwidth.
Appendix J : WDS and Deduplication in StorTrends iTX 257