* A mathematical model for assessing packet loss
Last time, we commented on the growing importance of data replication and the fact that it was significantly different than other data applications. In this newsletter, we’ll discuss the amazing impact that small amounts of packet loss has on goodput – the amount of data that is successfully transmitted.
Whenever we resort to using mathematical models we feel the need to point out that their purpose is insight and not numbers. By that we mean that any mathematical model of a complex system makes numerous assumptions. As such, the models are not accurate to two or three decimal places. The value of these models is that they provide insight into the bigger picture. In the case of data replication, they can be used to provide insight as to what happens to goodput if packet loss were to increase from 0.01% to 0.1%. Does the goodput remain basically the same? Is it cut in half?
Many models of TCP performance have been developed over the years. What is needed is not to develop yet another model (YAM), but to make use of the ones we have. One such model was developed by Mathis, et. al. The model states that goodput is less than or equal to (MSS/RTT)*(1 / sqrt{p}), where MSS stands for maximum segment size, RTT is the round trip time and p is the packet loss rate.
We strongly encourage our readers to not get too hung up on the preceding equation. We do, however, want to draw your attention to one simple fact that is highlighted in that equation: as the packet loss increases, the goodput decreases. To exemplify the impact of packet loss, assume that MSS is 1,420 bytes, RTT is 100 milliseconds, and p is 0.01%. Based on the formula, the maximum throughput is 1,420Kbps. If however, the loss were to increase to 0.1%, the maximum throughput drops to 449Kbps. That is almost a 70% reduction in goodput.
As previously noted, all mathematical models make assumptions and so the results of the model may be off by a small amount. However, even if the goodput is only reduced by 60% the point is the same – small amounts of packet loss can significantly impact an IT organization’s ability to do data replication.
In the next newsletter we’ll talk about ways to overcome the impact of packet loss, discuss the impact on goodput on out of order packets, and also visit a topic that generates a lot of heat among our readers: functionalities that belong in a router. In the meantime, more information on this topic can be found at Webtorials. Also, check out Network World’s IT Buyer’s Guide on data backup and replication, and data replication and mirroring.
