Evaluating Predictors of Congestion Collapse in Communication Networks
Christopher E. Dabrowski, Kevin L. Mills
Researchers model congestion in communication networks using a percolation process, where congestion spreads minimally before a critical load and expands rapidly afterwards. Some studies identify precursor signals arising near critical load, but none attempt to predict congestion collapse. We investigate whether such precursor signals could be used to predict onset of rapidly expanding congestion, and potentially to alert network managers to take mitigating actions to avoid congestion collapse. Specifically, we consider signals that might arise with changes in time series of router queue lengths. Using simulated networks, we specify and evaluate five predictors: autocorrelation, variance, threshold, growth persistence, and growth rate. Although most previous studies of spreading congestion use abstract network simulation models, we include both realistic and abstract models. We measure predictor performance under two scenarios: increasing and steady load. Under increasing load, we compare predictors based on consistent- prediction rate, latency, and persistence. Under both scenarios, we compare the rates and types of errors made by each predictor. We find that predictor performance is influenced by model realism. We also find that autocorrelation and variance predictors perform poorly under steady load in realistic network simulation models. For the most realistic model, the threshold predictor yields best accuracy.