Stratification Strategies for Sampling - based Non - intrusive Measurements of One - way Delay

be transferred and processed increases proportionally to the number of packets measured. High packet rates elevate the resource consumption and with this the costs for the measurement process. Since measurement costs should be only a small fraction of the service provisioning costs itself, there is a strong incentive for limiting the needed measurement resources and lowering the associated costs. Sampling provides adequate techniques for cost reduction by reducing the number of measured packets. Nevertheless, the achievable estimation accuracy for random sampling is limited by the applied sample size. With random sampling a higher accuracy can only be achieved if more packets are sampled, and with this higher costs emerge. In this paper we show how the estimation accuracy can be improved without increasing the sample size by using stratified sampling techniques. We investigated different stratification strategies with regard to the attainable reduction of the amount of sampled packets. It is shown that the sample size can be significantly reduced if packets are stratified in accordance their sizes. The influence of the setting of the stratification boundaries to the stratification gain is investigated. Furthermore, adaptive schemes with a dynamic adjustment of the stratification boundaries are compared to schemes with fixed boundaries.