Bayesian approach to correlating network traffic congestion to performance metrics
View Patent ↗For each of a number of network performance metrics, an associated value r pm that represents a difference between a first correlation coefficient r 1 computed for a baseline data set and a second correlation coefficient r 2 computed for a second data set that includes the baseline data set and other data points classified as duration outliers is computed. The first and second correlation coefficients for each network performance metric represent correlation between that network performance metric and durations of network connections. The network performance metric that has a largest associated r pm value of all statistically significant r pm values computed is selected as representing the probable root cause of the duration outliers. Statistical significance is measured through comparison of an r pm value with a statistical property of a set of Bayesian correlation coefficients computed for each performance metric.
1. A method for determining a probable root cause of congestion within a network, comprising:
collecting data representing a network performance metric;
computing for the collected data an associated value r pm that represents a difference between a first correlation coefficient r 1 computed for a baseline set of the collected data and a second correlation coefficient r 2 computed for a second set of the collected data that includes the baseline set and other data points classified as duration outliers, the first and second correlation coefficients representing correlations between the network performance metric and durations of network connections;
generating a value representing statistical significance of the value r pm by comparing r pm to a statistical property of a set of Bayesian correlation coefficients computed using the baseline set of the collected data;
eliminating the network performance metric from further consideration if the value r pm is less than the statistical property of the set of Bayesian correlation coefficients, or otherwise retaining the network performance metric for further consideration;
repeating the computing and eliminating or retaining processes for other network performance metrics to develop a set of retained network performance metrics and associated r pm values for each retained network performance metric; and
determining the probable root cause of congestion within the network based on selecting that retained network performance metric that has a largest associated r pm value of all r pm values associated with the retained network performance metrics, wherein the largest r pm value indicates congestion based on anomalous deviations in the durations of network connections.
2. The method of claim 1 , wherein the statistical property of the set of Bayesian correlation coefficients comprises at least one of standard deviation or root mean square deviation.
3. The method of claim 2 wherein the Bayesian correlation coefficients are each computed by selecting M random data points from the baseline set of data collected for an associated performance metric, M being equal to the difference between the number of data points in the second set of data and the baseline set of data for the associated performance metric, adding the selected M random data points to those data points in the baseline set of data to produce a Bayesian data set for the associated performance metric, and computing a correlation coefficient for the associated performance metric and duration using that Bayesian data set.
4. The method of claim 1 wherein the network performance metric comprises one of round trip time, server response time, packet loss or file size.
5. The method of claim 1 wherein the duration outliers are classified using one or more of Grubb's Test, Rosner's Test and Walsh's Test.
6. The method of claim 1 wherein the duration outliers are classified using information from the network performance metric.
7. A method for finding a cause of congestion within a network, the method comprising:
computing performance metrics for a network;
evaluating correlation coefficients, for each of the performance metrics, between the performance metric and connection duration within the network; and
determining a probable cause of congestion within the network based on selecting that network performance metric having a highest difference between an outlier correlation coefficient and a baseline coefficient which indicates congestion based on anomalous deviations in the durations of network connections.
8. The method of claim 7 wherein the performance metrics are chosen from the list comprising round trip time, file size, server response time, and packet loss.
9. The method of claim 7 wherein the difference (r pm ) between an outlier correlation coefficient and a baseline correlation coefficient for a particular performance metric is available for selection only if that r pm compares favorably to a statistical property of a set of Bayesian correlation coefficients computed for the associated performance metric and connection duration.
10. The method of claim 9 wherein each of the Bayesian correlation coefficients is computed by selecting M random data points from a data pool used to compute the baseline correlation coefficient for the associated performance metric, M being equal to the difference between the number of data points used in computing the outlier correlation coefficient for the associated performance metric and the baseline correlation coefficient for the associated performance metric, adding the selected M random data points to those data points in the data pool to produce a Bayesian data set for the associated performance metric, and computing a correlation coefficient for the associated performance metric and duration using that Bayesian data set.
11. The method of claim 9 wherein the statistical property of the set of Bayesian correlation coefficients comprises one of standard deviation or root mean square deviation.
12. The method of claim 11 wherein to compare favorably that r pm must be greater than the statistical property of the set of Bayesian correlation coefficients.