Automated Detection of Load Changes in Large-Scale Networks

Performance evaluation of an online load change detection algorithm

“…The algorithm was presented in [1], and uses a p-variate normal distribution to model the Internet traffic load within a day (p = 16). The measurements are input to the algorithm in a per day basis, after a preprocessing step where among other tasks abnormal data is removed (e.g.…”

“…To make this sample more manageable, we average such values in 16 disjoint intervals of 90 minutes, starting at midnight, and in addition we remove holidays to circumvent potential abnormal data. The reasons to choose 90 minutes as the averaging period are reported in [1], being the main one that the Internet traffic can be assumed Gaussian when there is enough time aggregation of the measurements ( [7], [8]). Consequently, the load model for a day is a 16-variate normal distribution.…”

“…When we find two suitable clusters, we apply the BF procedure after testing for normality. The BF procedure addresses the statistical problem of testing whether the means of two normally distributed populations (X (1) , X (2) ) 1 are the same (null hypothesis H 0 ), for the 1 we use superscripts between parenthesis to designate populations and subscripts for vector components of the random vectors case of unknown covariance matrices. The assumptions are that…”

“…1 summarizes the algorithm. Interested readers are referred to [1] for a more detailed description of the change load detection algorithm.…”

“…This algorithm applies clustering techniques to locate potential change free regions, and statistically sound methodologies to discover whether there is a remarkable change point between them. We described deeply our algorithm and applied it to real network traces showing good performance in a previous work ( [1]), but we found it necessary to verify its performance against synthetic data before deployment. Therefore, we have generated different synthetic datasets with the objective of verifying the different properties of our algorithm.…”