Probing for Loss: The Case Against Probe Trains
Packet loss is a fundamental measure of performance in computer networks. Loss and delay are the two sources of raw information available to end-to-end measurement by probing. In contrast to delay, loss data is scarce as loss probabilities are typically low and loss yields coarser binary information than delay. It is therefore crucial to use probing methods and statistical estimators which make the most of every recorded loss, and which are robust to details of how they occur. The authors investigate the best sampling strategy for loss measurement, i.e., how to choose probe sending times. Previous work concluded that trains are superior because they better estimate statistics such as the duration of lossy episodes.