Source-linked AI summary
Impact of the COVID-19 pandemic on the Internet latency: a large-scale study
Massimo Candela, Valerio Luconi, Alessio Vecchio
TL;DR
The paper examines how pandemic-driven online activity affected Internet latency, a network property important to applications and perceived network health. Using large-scale measurements, it finds increased latency variability, especially in the evening, with impacts that differ across countries.
Problem
The pandemic prompted increased online activity, but available statistics largely reflected individual operators and traffic volume, leaving a broader view of latency effects limited.
Source: §1 · pp. 1–2Method
Results
The impact was relevant especially as higher latency variability, with major changes in the evening and different patterns across countries.
Source: §1 · p. 2, §8 · p. 17Takeaways & Limitations
Latency provides an indication of network health, and the measured impact varied across the European countries considered.
Source: §1 · p. 2, §8 · p. 17The analysis is limited to a portion of the Internet.
Source: §8 · p. 17
Abstract
from arXiv · showhide
The COVID-19 pandemic dramatically changed the way of living of billions of people in a very short time frame. In this paper, we evaluate the impact on the Internet latency caused by the increased amount of human activities that are carried out on-line. The study focuses on Italy, which experienced significant restrictions imposed by local authorities, but results about Spain, France, Germany, Sweden, and the whole of Europe are also included. The analysis of a large set of measurements shows that the impact on the network can be significant, especially in terms of increased variability of latency. In Italy we observed that the standard deviation of the average additional delay -- the additional time with respect to the minimum delay of the paths in the region -- during lockdown is ~3-4 times as much as the value before the pandemic. Similarly, in Italy, packet loss is ~2-3 times as much as before the pandemic. The impact is not negligible also for the other countries and for the whole of Europe, but with different levels and distinct patterns.
1. Introduction
The paper examines how pandemic-driven changes in online activity affected Internet latency, using a large measurement set with Italy as its focus and comparisons across other European countries. It reports increased latency variability in Italy, especially in the evening, with differing impacts elsewhere.
1. Introduction
- The analysis uses a large measurement set to examine latency alongside packet loss, path changes, and HTTP-based latency.It decomposes delay by time of day, target type, and IP version.Source: §1 · pp. 1–2
- For Italy, the baseline is February 11–17, 2020, and the post-lockdown week is March 10–16, following the introduction of restrictions in stages.The paper labels these periods W1 and W2, respectively.Source: §1 · pp. 1–2
- The study addresses the lack of a global, latency-focused view beyond operator-specific traffic statistics.The authors also note that operator reports were limited to their organizations’ boundaries.Source: §1 · p. 2
- Measurements from one platform support comparisons across Italy, Spain, France, Germany, Sweden, and Europe without bias from heterogeneous systems.Restrictions varied by country, and the study shifted comparison periods for countries according to major lockdown events.Source: §1 · p. 2
- Italy’s lockdown impact was most evident as increased latency variability, especially in the evening; effects were milder in Germany and France and similar to Italy in Spain and Sweden.The evening pattern suggests recreational activity affected latency more than remote work or distance learning.Source: §1 · p. 2
2. Related work
Prior Internet-disruption studies mostly examined reachability, routing, or traffic, while pandemic reporting focused on traffic within individual organizations. This paper situates its latency analysis in that gap and in the unusually broad scale of the pandemic.
2. Related work
- Existing scientific work on COVID-19’s Internet impact was limited, with one cited study examining an Italian university campus network and its distance-learning system.That study reported changes in incoming and outgoing traffic and used passive measurements and application logs.Source: §2 · pp. 2, 4
- Earlier natural-disaster studies examined effects including routing changes, outages, and latency increases, often finding impacts limited in location or duration.The cited studies include earthquakes, hurricanes, and severe weather.Source: §2 · p. 3
- Few studies of Internet disruptions examine latency, although latency can indicate end users’ perceived quality of service; COVID-19 also had broader spatial and temporal reach than earlier events.Earlier work primarily considered reachability, routing, or traffic.Source: §2 · p. 3
- Operator and provider reports documented traffic increases but generally did not report latency, and their measurements were bounded by the reporting organizations.The paper notes that traffic alone cannot estimate perceived Internet performance without information about capacity.Source: §2 · pp. 3–4
3. Collection of data
The study draws on RIPE Atlas ping measurements from both automatically scheduled and user-defined measurement sets. It filters these data for sustained observations and European targets before constructing two large datasets.
3. Collection of data
- RIPE Atlas supplies the study’s ping measurements through both Anchoring Measurements (AMs) and User-Defined Measurements (UDMs).AMs use predefined targets; UDMs allow users to select targets and schedules, and both produce RTT measurements using ping.Source: §3 · p. 4
3.1. AM-derived dataset
- The AM-derived dataset retains European source-target pairs with successful results on at least 40 days during February 11–March 26, 2020.The stable AM set provides more than 11.7 billion RTT values from 303,603 source-target pairs.Source: §3.1 · pp. 4–5
3.2. UDM-derived dataset
- The UDM-derived dataset uses periodic measurements with at least three RTT values per ping, sustained results on 40 days, and targets geolocated in Europe.When target counts were too large, the study randomly selected 10,000; the resulting dataset contains more than 1 billion RTT values from 671,292 source-target pairs.Source: §3.2 · p. 5
4. Method
The method estimates a path’s fixed delay from its minimum observed delay, then measures additional delay relative to that baseline. It aggregates these additional-delay values across regional source-target pairs in 30-minute buckets.
4. Method
- The method uses the minimum observed path delay to estimate transmission and propagation delay, isolating the queueing and processing component that can be affected by increased traffic.More samples improve the chance of observing lightly loaded conditions.Source: §4 · p. 5
- The method computes minimum, average, and maximum RTT values per ping execution and subtracts the pair’s baseline to estimate additional delay.These values are computed for each pair and time.Source: §4 · pp. 5–6
- For each source-target pair, the global minimum across the observation period estimates the best path’s transmission and propagation delay if paths change.This minimum becomes the baseline for additional-delay measures.Source: §4 · p. 6
- The additional-delay values are averaged across regional source-target pairs in 30-minute buckets.The resulting regional averages describe delay relative to the best path observed over the full observation period.Source: §4 · p. 6
5. Characterization of the dataset and preliminary data analysis
The section characterizes Italian and European path distances and minimum RTTs, then uses processed Italian measurements to reveal increased delay variability after lockdown. The subsequent analysis focuses mainly on the variable latency component, while an analogous European or UDMD discussion is omitted for brevity.
5. Characterization of the dataset and preliminary data analysis
- Figure 1 characterizes path distances and minimum RTTs: distances span a few kilometers to nearly 1,000 km in Italy, while minimum RTTs reach ~40 ms in Italy and ~80 ms in Europe.European distances cover up to 500 targets, using all sources; city-level locations were available for approximately 83–98% of considered pairs.Source: §5 · p. 6, Figure 1 · p. 7
- Processed Italian measurements show increased variability after lockdown, clearest in the 95th percentile of qmin.Raw values do not show a clear trend, while the five-minute moving average shows increased variability; 30-minute buckets reduce spikes in smin.Source: §5 · p. 7
- No analogous discussion is provided for the European dataset and/or UDMD, for brevity.Source: §5 · p. 7
- The analysis mainly uses r to track latency’s variable component, except in Section 6.1, which also includes s.The analysis examines how the measures evolve over time in relation to lockdown measures.Source: §5 · p. 7
6. Impact of lockdown on the Italian Internet
The Italian analysis examines lockdown-related latency across measurement types, times of day, IP versions, YouTube servers, and AS-level paths. Across these perspectives, latency variability generally increased, with effects varying by traffic type and period.
6.1. Overall results
- During lockdown, Italian smin peaks were approximately 30% higher than before lockdown, while smin standard deviation rose 201.5% for AMD and 310.1% for UDMD.For rmin, the W2 average increased 27.7% in AMD and 66.8% in UDMD; its standard deviation increased 203.6% and 280.8%, respectively.Source: §6.1 · pp. 7–8
- Across measurements from Italy to the rest of Europe, variability generally increased during lockdown, while reverse-direction AMD and UDMD patterns differed.The study also notes that operator infrastructure improvements may have lowered some night-time latency minima.Source: §6.1 · pp. 8–9, Figure 4 · p. 10
6.2. Packet loss
- Packet loss increased from 2.8e-3 to 5.9e-3 (+110.3%) for AMD and from 3.1e-3 to 9.4e-3 (+205.7%) for UDMD.Its standard deviation also rose substantially in W2; a roughly two-day AMD spike around February 29–March 1 appeared across measurements to 74% of targets, without a corresponding RTT increase.Source: §6.2 · pp. 9–10, Figure 5 · p. 10
6.3. Circadian rhythms
- The largest time-of-day increase occurred from 16:00 to 22:00 UTC, peaking at 20:00–21:00 UTC; night hours showed no considerable increase.W2 rmin rose substantially in peak hours, while lighter-load periods sometimes showed better performance than before lockdown.Source: §6.3 · pp. 11–12, Figure 6 · p. 11, Figure 7 · p. 11
6.4. IPv4 and IPv6
- In peak hours, average rmin increased 121.7% for IPv4 and 11.2% for IPv6; variability rose 103.8% for IPv4 and 128.1% for IPv6.IPv6 showed generally higher variability than IPv4, but its peak-hour latency was less affected; UDMD IPv6 results were not reported because the sample was too limited for statistically sound results.Source: §6.4 · p. 12
6.5. A content delivery example: YouTube
- For measurements to 115 YouTube servers in Italy, average rmin rose 25.2% and its standard deviation rose 86.9% in W2.Night rmin slightly improved during lockdown, while evening values abruptly increased after the first day of complete lockdown; measurements to Facebook and Netflix lacked sufficient coverage for significant results.Source: §6.5 · p. 13
6.6. Path changes
- The average daily fraction of source-target pairs experiencing an AS-path change was 17.9% higher in W2, although daily fractions remained between 0.1 and 0.2.This analysis used AMD pairs because there were not enough matching UDMD traceroutes.Source: §6.6 · p. 14, Figure 12 · p. 14
6.7. HTTP
- HTTP latency increased during lockdown, with average r and its standard deviation 24.9% and 98.5% higher in W2 than W1.HTTP latency includes server response time and TCP retransmission effects, so it is not directly comparable with ICMP ping latency; the results indicate the increase was not limited to diagnostic traffic.Source: §6.7 · p. 15, Figure 13 · p. 15
7. Impact of COVID-19 pandemic in other European countries
Across the countries, lockdown-period changes differed: latency variability rose in several settings, while Europe-wide average latency changed modestly and packet-loss patterns did not consistently track latency.
Spain
- In Spain, rmin average and standard deviation rose 37.4% and 110.8% in AMD, versus 63.2% and 19.3% in UDMD.Packet loss decreased in AMD but increased greatly in UDMD.Source: §7 · p. 16
France
- In France, UDMD rmin average rose 0.5% and standard deviation 49.6%, while AMD averages and standard deviations decreased.Packet loss increased in both AMD and UDMD.Source: §7 · p. 16
Germany
- In Germany, UDMD rmin average and standard deviation rose 34.1% and 88.7%, while AMD rose 2.8% and 37.3%.Latency degradation began more than a week before lockdown; packet loss rose slightly for UDMD and fell for AMD.Source: §7 · p. 16
Sweden
- In Sweden, rmin average and standard deviation rose 44.7% and 531.2% in AMD, and 28.6% and 231.7% in UDMD.Packet loss increased for AMD but decreased for UDMD; the authors offer possible explanations for the progressive latency increase rather than establishing one.Source: §7 · p. 16
Europe
- Across Europe, rmin average rose 8.1% in AMD and 15.4% in UDMD, while rmin standard deviation rose 114.6% and 108.1%, respectively.The authors describe the continent-scale performance changes as noticeable but not dramatic; packet loss decreased for AMD and was almost unchanged for UDMD.Source: §7 · p. 16
Cross-country comparison
- Across countries, increased latency did not consistently coincide with increased packet loss; some cases showed latency rising while packet loss fell, and vice versa.The comparison uses each country's first measurement week and first lockdown week; Germany's periods cover four days, while Sweden and Europe use the first and last observation weeks.Source: §7 · pp. 15–16
8. Conclusion
The study finds that pandemic-era online activity was associated with relevant Internet-latency changes, especially greater variability and evening effects. The authors caution that their measurements cover only a portion of the Internet.
Conclusion
- The large measurement set shows relevant latency impacts, especially higher variability, with the major changes observed in the evening.The authors suggest remote working and distance learning contributed less to additional latency than evening entertainment activity.Source: §8 · p. 17
- Latency matters both for applications sensitive to fluctuations and as an indicator of network health.The paper names online games, video calls, VoIP, and IP geolocation as examples of latency-sensitive applications.Source: §8 · p. 17
- The analysis is limited to a portion of the Internet, so its numbers do not cover the entire network..Source: §8 · p. 17