Identifies potentially uncalibrated M-Lab servers by comparing measurement distributions across nearby sites: if a server is well calibrated, some other server should give the same result for at least one client ISP in the region. Low scores are best, high scores indicate problems.
This version is not suitable for general use because it is prone to both false positive and false negative results.
The algorithm scans the selected sites for triplets — a target site, a benchmark site, and a client ISP — where both sites give similar measures of that ISP. Target sites with high scores (ratio ≫ 1.0 or KS distance ≫ 0.0) have no matching site and are suspect for calibration problems. Poor results always need to be checked with other methods (e.g. Regional Details); since we have no ground truth, the judgement is somewhat subjective.
For information about Differential Histograms and how they expose anomalies in Internet mid-paths see the project overview.
See complete The M-Lab Calibration Dashboard documentation.