Analysis of Network Performance Test Results

In collaboration with: Brian Bockelman
Meet with Bockelman and me before submitting a proposal on this project

HCC has a database of network performance test results (a few 100M data points). Each test is one of two types: either "one-way delay" (i.e., half of RTT) or single-stream TCP performance. There are on the order of 100 nodes participating in these tests; each endpoint tests against 10-50 other nodes.

Given this dataset, the goal is to identify:

• Significant changes in network performance. Network performance differs on a minute-by-minute basis, but you sometimes see massive changes in performance (routes change and go over a poorly-performing link, line card starts dropping packets and needs to be replaced, etc.) that persist for hours or days. Currently, humans look at the graphs and, hopefully, notice the changes.
• There are other groups that are doing something similar with the same data source; students wouldn't be working alone.
• Same as above, but looking for significant changes in network performance that are correlated across several nodes.
• Some nodes are poorly provisioned and provide nonsensical testing results into the central database. We would like to have a mechanism to identify these "bad apples", or, better yet, determine when a "good node" transitions to a "bad node".

A HCC grad student (Jerrod Dixon) is working on a similar problem, and may have some tools and scripts to help students start out.

Return to the CSCE 478/878 Home Page