Measuring conditional web throughput
G. Mardente, Marco Mellia, Andrea G. Bianco · 2005
We focus on the analysis of the performance of HTTP transaction over TCP connections as derived from a measured data set collected at the ingress/egress point of Politecnico di Torino. We argue that looking at global average performance figures may hide some interesting trends that researchers and operators may want to detect. Thus, we partition the original measured set by looking at conditional densities and averages; considered conditional parameters include, among others, TCP connections RTTs, length of data set, adopted loss recovery algorithm. We show that particular care must be taken when analyzing data, since the measured set whose size is reduced by the partitioning scheme could be not large enough to ensure significance. A simple method to evaluate the significance of the presented measures is used. Ingenuity is required to understand behaviors not in line with the classical intuition driven by TCP knowledge.