Revisiting the Constancy of Internet Path Delays
DescriptionInvestigating internet path properties such as delay, loss, and throughput have been of interest to the network measurement community for decades. The 2001 work of Zhang et al. ("On the constancy of Internet path properties", IMW 2001) [ZPRS01] identified three notions of constancy – mathematical, operational, and predictive.
This poster presents early-stage work that reexamines ZPDS01 in today's internet. At this stage, we are focusing on internet path delays, due to the significant impact latency has on performance of real-time interactive apps. We hypothesize that changes to Internet connectivity structure in the past decades have directly impact path delays constancy.
Our efforts utilize the RIPE Atlas platform. Despite good data availability, there are a number of methodological challenges. First, Atlas anchors and "ordinary" probes provide a broader set of vantage points than the original NPD infrastructure, but it is unclear how sensitive the results of any analysis of delay constancy may be to the specific endpoints selected. Second, CP_bootstrap and CP_rank, described in ZPDS01, may not be as robust as recently developed techniques. Third, not all data are noise-free, thus we need to separate good and bad measurements.
Our current results highlight the challenges above. We found that simple methods of filtering measurements (e.g., removing values beyond some threshold) can eliminate high-latency events that are likely not due to noise in measurement. Change point methods in ZPDS01 can perform poorly for some traces, and while recent methods perform somewhat better, none robustly and consistently identifies mathematically constant time segments.