Lawrence Berkeley National Laboratory

Accurately predicting network behavior is beneficial for TCP congestion control, and can help improve routing, allocating network resources, and optimizing network designs. This task is challenging because many factors could affect network traffic, such as the number of network sessions and synthetic reordering. There are also many ways to measure the network state, such as the number of retransmissions per flow and packet duplication. For this work, we use a set of passive TCP flow measurements collected at a major computer center on multiple data transfer nodes (DTN). To assist the operations of the computer network, we propose to detect abnormally slow network transfers in real-time. The proposed system breaks the network monitoring logs into fixed-size chunks and employs a state of art classifier to identify the slow time windows. This method will be validated on real large datasets collected from several DTNs. The proposed method is able to generate models to quickly detect large intervals of low performing network transfers, which require attention from network engineers.