Use Internet2 SiteID

Already have an Internet2 SiteID?
Sign in here.

Internet2 SiteID

Statistical prediction models for network traffic performance

Time 01/16/13 11:45AM-12:00PM

We do not host flash-based video files on our servers anymore. Please contact Web Support for further details about netcast videos.

Session Abstract

We present a statistical prediction model for network traffic performance by analyzing network traffic patterns and variation with the network conditions, based on two types of historical network measurement data. A time series model with Seasonal Adjustment is developed to decompose the SNMP data into three components; seasonal, trend and random components. The seasonal component models the periodical pattern in the network traffic. The trend component models the changes in the network traffic without any influences of the randomness and periodic pattern. The component that cannot fit into the seasonal pattern or trend is modeled as the random component. This separation of components enables an effective analysis in prediction, tracing the network traffic and quantifies the variation of the network traffic that flows into multiple outlets. We also present studies from fixed effects, random effects and interactive effects of variables in the Generalized Linear Mixed Model with Netflow data. Our model takes into account both the universal variance caused by randomness and the variance by changes in the network conditions to improve the accuracy of the prediction. We will present prediction models using both SNMP and Netflow data and the preliminary experimental results.


Speaker Kejia Hu Lawrence Berkeley National Laboratory

Presentation Media

Secondary tracks Network Research & Emerging Technologies Network Performance/Utilization

gold Sponsors

silver Sponsors