Photo of a woman waiting at a platorms edge as a train goes past

On the Constancy of Latency at the Internet’s Edge

Picture of Aditya Bhat
Guest Author | BITS Pilani, Goa
Categories:
Twitter logo
LinkedIn logo
Facebook logo
July 24, 2025
In short
  • A recent study characterized the stability of latency between end users and the Content delivery network edge servers (CDNs), which are critical infrastructure for delivering content to most end users.
  • The findings indicate that our network infrastructure is typically well-optimized for offering network connections with little to no variations.
  • Determining the exact root causes of the latency spikes and variations remains a key focus of our future work.

Latency or round-trip-time (RTT) between two end hosts (for example, a user’s laptop or tablet and a web server) is a key determinant of user experience. Establishing a low-latency connection between any two communicating endpoints is, therefore, crucial for maximizing application performance.

Infographic explaining how CDNs distribute content to end users
Figure 1— Content delivery networks (CDNs) reduce the latency experienced by end users (or their applications) by delivering content from servers that are in “close proximity” to the users. The reduction in latency substantially improves the performance of transport protocols, improving applications’ performance and end-users’ quality of experience.

Content delivery networks (CDNs) have become critical for reducing latency by serving content from “edge” servers that are in “close proximity” to end users. This task has required CDNs to deploy hundreds of thousands of servers in diverse networks and geographic locations, and they can dynamically determine which server a given end-user’s request can be served from based on performance metrics (latency) and other factors (load). As such, most of the content consumed by end users on the Internet is currently served by CDNs, making them a valuable reference for understanding the overall performance of the Internet from an end user’s perspective.

In a recent study, my colleagues and I at BITS Pilani, Goa, and Vrije Universiteit Amsterdam sought to characterize the latencies of the paths between end users and the CDN edge servers. 

Our interest lay in understanding the “stability” of latency, since it has substantial performance implications for various applications.

World map showing the location of vantage points and targets used in the study.
Figure 2 —  The geographically diverse footprint of the vantage points and targets spans thirty-four and fifty-six countries, respectively.

Even the Best Systems See Variation

Our work focused on the latencies’ “constancy”—to what extent the observed latencies are free from “substantial” variations.

To measure constancy, we used 69 RIPE Atlas anchors as our vantage points and one hundred edge servers (see Figure 2) belonging to five widely used CDNs as targets. (Please refer to our paper for more details on selecting our vantage points and targets.)  We measured the latencies between our vantage points and targets every two hours for 10 consecutive days. We then collated the latencies between each vantage point-target pair into a “timeline.” We characterized the constancy of latency observations using three well-known approaches: 

  • Mathematical Constancy —Using three statistical algorithms, we measure where the observed latencies exhibit a “substantial” deviation from past values in a timeline. Essentially, we ignore minor deviations in the latency signal and only consider substantive changes, often called “level shifts.” The instances where we observe such level shifts are called “changepoints” (see Figure 3), and the region between any two changepoints is the “change-free region (CFR)” wherein the latency (signal) is deemed constant. The larger the duration of the CFRs, the more stable (or constant) the latency timeline.
  • Operational Constancy — We define a sequence of latency observations as constant if these latencies are within certain predefined thresholds (say, 25-50 ms).
  • Predictive Constancy — If, based on historical data, we can accurately predict the following latency observation, then the latency variations are deemed not substantial—the latency timeline is (per the predictive notion) constant.

The findings indicate that our network infrastructure is typically well-optimized for offering network connections with little to no variations (or jitter). 

We observed that the maximum CFR regions—the maximum period during which the latencies in a timeline remained mathematically constant—are at least three times larger than those reported in recent prior work.

We also noticed that, even from an operational perspective, most timelines were constant: 95% have a maximum CFR of one hundred hours or longer, a factor of four more than that shown in prior work. Even with a stricter constancy threshold of 25 ms, we observed that the median of the maximum CFR of all timelines was more than 230 hours (96% of the duration of our study).

We observed, nevertheless, that about 4% of all our observations manifest latency ‘spikes’, which is at least two orders of magnitude larger than that reported in recent prior work. Though not typical, the network paths provided by our current Internet infrastructure experience non-trivial degradations in latency, which in turn can result in poor performance of end-user applications.

Time series line graph showing shifts (change points) in latency constancy.
Figure 3 — A plot showing a synthetic RTT timeline (in blue) with time on the x-axis and RTT on the y-axis. Here, the mathematical notion of constancy is used to detect level shifts, changepoints, and change-free regions.

Determining the exact root causes of the latency spikes and variations remains a key focus of our future work. Our current work also focused only on latency measurements between IPv4-capable end hosts, so we plan on investigating latencies between IPv6-capable hosts next. We also want to check the correlation between the IPv4 and IPv6 latency measurements between a given pair of (dual-stacked) end hosts. These explorations will likely offer rich insights into the capabilities of our network infrastructure.

Contributors: Aniket Shaha, Vaibhav Ganatra, Balakrishnan Chandrasekaran, and Vinayak Naik

Aditya Bhat is a Computer Science undergrad at BITS Pilani, Goa, interested in network measurements, big-data analyses, and deep learning.

The views expressed by the authors of this blog are their own and do not necessarily reflect the views of the Internet Society.


Image by wal_172619 from Pixabay