Breaking Through the Clouds: Performance Insights into

Starlink’s Latency and Packet Loss

Robert Richter

HPI, University of Potsdam

Potsdam, Germany

Vasilis Ververis

HPI, University of Potsdam

Potsdam, Germany

Vaibhav Bajpai

HPI, University of Potsdam

Potsdam, Germany

Abstract—Our modern era is experiencing a rapid evolution in

satellite Internet access. However, it is unclear how well these systems

perform and what we can expect from Internet access via satellites.

Previous research has studied the performance and resilience of

such systems, uncovering several drawbacks (e.g., high packet loss

and unstable performance). In this work, we thoroughly investigate

the characteristics of the Starlink network. We scrutinize the TLS

handshake latency, packet loss, and the diurnal latency variation to

establish a correlation between these factors. To achieve this, we utilize

historical data measured by RIPE Atlas and Cloudﬂare Radar from

2022-01-01 to 2024-06-30.

We ﬁnd no statistically signiﬁcant correlation between latency and

packet loss in the Starlink satellite network. However, we discover an

intriguing pattern suggesting that Starlink exhibits speciﬁc latencies

more consistently than others. This ﬁnding contradicts recent research

that claims a signiﬁcantly better performance of Starlink with median

latencies substantially lower than 80 ms. Furthermore, our ﬁndings

reveal signiﬁcant geographical variations, where even highly developed

countries such as Germany experience packet loss ratios exceeding 10%.

Additionally, we examined Starlink’s routing behavior, which reveals

two sudden spikes in latency. The ﬁrst spike is attributable to the

transition between satellite and terrestrial networks, while the second

is seemingly unrelated to Starlink.

Fig. 1: Growth of number of satellites in satellite constellations from

2000 to June 2024 (according to N2YO [1]).

I. INTRODUCTION

Internet access is challenging in remote regions and areas with

unreliable terrestrial infrastructure (e.g., during wars or natural

disasters). Consequently, businesses have explored the concept of

Internet access via Low-Earth Orbit (LEO) satellites. Since space

is largely inaccessible, satellites are resilient to many catastrophes

and other threats. Additionally, communication occurs at the speed

of light, resulting in low latencies at low Earth orbits. Satellites

were expected to be predictable due to well-deﬁned properties

of a constellation, including the positioning of ground stations,

ﬂight paths of satellites, and the positions of neighboring satellites.

However, satellite Internet access has been observed to exhibit

signiﬁcant variation in latency and packet loss. In this study, we

examine the largest Satellite Network Operator (SNO): Starlink,

and aim to address the following Research Questions:

2018

2019

2020

2021

2022

2023

2024

Classiﬁcation

Starlink

Orbcomm

OneWeb

Beidou

O3B

120

943

104

1871

388

3481

498

5326 6396

628

Intelsat

TABLE I: Growth of number of satellite in satellite constellations

from 2017 to June 2024 (according to N2YO [1]).

Starlink Latency Improvements from 2022 to 2024: The

median latency for Starlink has shown signiﬁcant improvement

over the period from 2022 to 2024. The data demonstrates that

the median latency has decreased, with the lowest observed median

latency on a single weekday being approximately 80 ms in 2024.

This improvement suggests that Starlink has been optimizing its

network infrastructure, potentially through the addition of more

satellites and ground stations, resulting in better performance.

Bimodal Distribution of Latencies: The latency data exhibits a

bimodal distribution, with two distinct peaks. The ﬁrst peak repre-

sents lower latencies (approximately 80–100 ms), while the second

peak represents higher latencies (approximately 150–250 ms). This

bimodal distribution indicates that Starlink produces two distinct

latency ranges with particularly high frequency. The reasons for

this pattern could be related to factors such as the positioning of

ground stations, satellite orbits, weather conditions, and potentially

1) How does Starlink perform in terms of latency and packet

loss?

2) To what extent do latency and packet loss characteristics in

Starlink’s network correlate?

3) How do packets route through Starlink’s network and how do

network paths impact latency?

Data from RIPE Atlas and Cloudﬂare Radar is used to address

the research questions. The data was collected programmatically

from 2022 to 2024. We summarize our ﬁndings from the analysis

of this longitudinal data as follows.

different subscription models.

or not ideal. With n hops, the bent-pipe is called an n-hop-bent-pipe

Packet Loss Rates Above 10% for Countries with High [3].

Ground Station Density: Despite the high density of ground

B. Satellite Numbers in Different Satellite Constellations

stations in some countries, packet loss rates may exceed 10 %.

For example, Germany, a country with numerous ground stations,

experiences packet loss ratios exceeding 10 %. This ﬁnding suggests

that factors other than ground station density, such as network con-

gestion, satellite hardware limitations, or environmental conditions,

signiﬁcantly inﬂuence packet loss rates.

In recent years, satellite technology has advanced rapidly, driven

primarily by the growing demand for global connectivity and

communications. Consequently, companies have constructed their

own satellite constellations, resulting in a total of more than 29 000

objects in space at the time of this writing. As the satellite count

is highly relevant for the performance of an SNO, we collected

the numbers of satellites per constellation until June 2024 from

N2YO. Figure 1 shows various satellite constellations with the

number of satellites they comprised per year. Table I shows the

corresponding numbers, starting in 2017. We can observe that

Starlink is by far the largest constellation in terms of number of

satellites. At the time of this writing, it comprises 6396 satellites.

The Starlink constellation grew from 2022 to 2023 by nearly 2000

satellites. OneWeb, Starlink’s closest competitor, has a total of 628

satellites with no growth between 2023 and 2024. Other satellite

communication constellations such as Orbcomm or Intelsat did not

grow at all, or even experienced a reduction in satellites. Only

Starlink and OneWeb have seen signiﬁcant growth in recent years.

We note that this is due to the requirement of LEO constellations

(i.e., Starlink and OneWeb) to have signiﬁcantly more satellites,

compared to GEO constellations (e.g., Intelsat and O3B).

The remainder of this paper is organized as follows: We present

a brief background and related work in Section II. Section III de-

scribes the methodology, speciﬁcally the process of data collection.

The research questions are addressed in Sections IV to VI, including

the data analysis of network latency, packet loss, and traceroute

measurements. Finally, Section VII concludes with a comprehensive

overview of the ﬁndings.

II. BACKGROUND AND RELATED WORK

A. Background

From a user’s perspective, satellite communication exchanges

packets with a target the same way as terrestrial internet connections

(except for the physical layer [2]). The difference lies in the

communication between the user and the ﬁrst terrestrial hop. The

user is provided with an antenna that allows communication with

the provider’s satellites. The satellites are within a speciﬁc satellite

constellation at Low-Earth Orbit (LEO) or Geostationary-Earth

Orbit (GEO). The altitude of the constellation has a major impact

on latency. Equation (1) shows the minimal latencies of the GEO

and LEO constellations. The LEO constellation provides a far better

latency in an ideal case.

C. Related Work

SNOss have been in existence for several decades, beginning

in the late 1990s, when Iridium announced its plans to build a

mega-satellite constellation [4], [5]. Even though their initial com-

mercial model proved unsustainable, modern SNOs have access to

newer, cost-effective technologies. With the emergence of Starlink,

OneWeb, Orbcomm, and others, research has taken the ﬁrst steps

in elucidating how ”satellite Internet” might function optimally in

the future. In 2020, the ﬁrst mega constellation simulators Hypatia

[6] and Starperf [7] were developed. However, this direction was

not further explored as it became apparent that the simulators did

not accurately model real-world conditions. Therefore, research

shifted toward performing measurements on existing constellations.

Performance of satellite network operators was presented in various

publications [8], [9], [10], [11], [12], [13], [3], [14], [15]. Most

measurements were performed on the Starlink constellation, due

to its relative affordability and widespread accessibility, even for

non-business customers. Resilience to disasters has been analyzed

by Stevensetal. [16]. It is also worth mentioning the inﬂuence of

weather [17], mobility [18], [19], and solar magnetic storms [9],

[20], [21]. All of these factors must be taken into account when

implementing routing [22], [23], [24] in a satellite constellation.

This includes ISLs [25], [26], the orbital dynamics of satellites [27],

[28], and the strategic placement of GSs [29].

2 · 35 786 km

2 · 550 km

≈ 0.240 s

≈ 0.004 s (1)

300 000

The broad process of satellite communication works as follows. The

sender utilizes an antenna for communicating with the satellites. It

sends packets to a satellite, which routes packets to their receiver.

The satellites may use an Inter-Satellite Link (ISL) to send packets

to other satellites until a suitable Ground Station (GS) is found.

The GS is usually connected to a terrestrial ISP, which is capable

of communicating with the target. The complicated part is routing

the packets through the satellite constellation itself. Such a route

is called a bent-pipe. The simplest case is sending packets to the

ground station right after they have been received by a satellite. This

is called a 1-hop-bent-pipe. Often, 1-hop-bent-pipes are not possible

Number

of Probes Country

Number

of Probes

Country

Philippines

Switzerland

United Kingdom

France

Greece

Poland

Italy

Benin

Czechia

In optimal conditions, Starlink could complement 6G deployment

[30]. However, the most signiﬁcant problem is the lack of knowl-

edge about the operational characteristics of Starlink. Therefore,

research has attempted to extract the ﬁrmware [31], [32].

Also, with the onset of the war in Ukraine, a new issue is the use

of satellite communication in conﬂicts: it operates independently

of a terrestrial infrastructure in the conﬂict zone and could even

exacerbate the war [33], [34], [35].

Kiribati

Spain

Canada

Re´union

Belgium

Austria

Haiti

Sweden

Honduras

Falkland Islands

Virgin Islands, U.S.

United States

Netherlands

Australia

Germany

TABLE II: The number of probes per country in the AS14593 on

the RIPE Atlas measurement platform.

However, existing literature predominantly concentrates on Star-

link deployment in a lab scenario (i.e., in nearly ideal conditions).

(a) USA

(b) Canada

(d) Philippines

Fig. 2: The history of median latencies from 01/2022 to 06/2024 for the USA, Canada, Germany, and the Philippines according to

RIPE Atlas data.

It remains unclear how Starlink performs from the user’s per- was obtained from the RIPE Atlas TLS tests, collected by built-

spective, where the network might suffer from external inﬂuences. in measurements (i.e., measurements that run continuously in each

This study presents a novel perspective by using RIPE Atlas and individual probe at a regular time interval).

Cloudﬂare Radar to facilitate comparison with the performance

Figure 2 shows the history of median latencies from January 2022

distribution of Starlink users. Additionally, this research provides

to June 2024 for the USA, Canada, Germany, and the Philippines.

a much longer time series interval that depicts the performance

The median latencies typically range from 100 to 150 ms for most

development from 01/2022 to 06/2024.

countries. Comparing the results of 2022 with those from 2023

reveals that most of the observed countries show an increase in

latency in the last months of 2022 (mostly in December). At the

III. METHODOLOGY

For analysis, we use RIPE Atlas [36] and Cloudﬂare Radar [37].

From RIPE Atlas, we analyze built-in measurements of all registered

Starlink probes. Built-in measurements are continuously running

measurements that perform measurements from probes toward root

servers (*.root-servers.org). The frequency of each measurement

type varies (e.g., Ping runs every 240s and Traceroute every 1800s).

We do not include any additional custom measurements. From

Cloudﬂare Radar, we include aggregated latency windows.

The resulting data from RIPE Atlas and Cloudﬂare Radar is

stored in a PostgreSQL database. The data is exported into Parquet

ﬁles, which we publish to allow for reproduction of our results¹.

The database stores several measurement types, such as Ping,

Traceroute, TLS, HTTP, Disconnect Events, DNS measurements,

and details of the RIPE Atlas probes connected via AS14593

origin-Autonomous System (AS), and information on all Starlink

satellites launched until 06/2024. The whole database comprises ≈

150 GB in PostgreSQL (≈ 37 GB in Parquet ﬁles). At the time of

measurement, 146 relevant probes were connected to RIPE Atlas.

Table II shows the number of probes per country found in the

RIPE Atlas measurement platform.

end of 2023, latency started to decrease again. In the last few

months until June 2024, we observe an increase in latency again. We

perform a ﬁne-grained analysis of latency variation in Section V,

examining latency across weekly temporal patterns.

Figure 3 illustrates the CDF plots of 2022 to 2024. We observe

similar performance in 2022 and 2024, but an increase in latency

in 2023, corroborating our earlier observations from Figure 2. The

CDF is also continuous up to a certain point, where it ﬂattens out,

followed by a stronger increase again. This is also observed for

curves from other countries (e.g., France). This observation suggests

that speciﬁc latency ranges are encountered more frequently. The

range varies from country to country, but is usually located between

150 and 250 ms.

It becomes apparent that there is a bimodal distribution present

(i.e., there is a large gap within the latencies for most countries). By

2024, this pattern is became more pronounced. At this point, it is

not clear why the bimodal distribution occurs, but we hypothesize

that either the measurements were not consistent enough for such

a pattern to occur or the gap is a characteristics of the Starlink

system. The latter would suggest that Starlink serves certain laten-

cies better than others, meaning that some users experience lower

latencies under certain conditions, while others encounter higher

latencies. These conditions could include subscription models, ge-

ographic differences between countries, weather, user altitude, or

GS availability. In addition, we ﬁnd that approximately half of the

measurements are below 100 ms, while the other half are above

100 ms. This contradicts previous studies suggesting that Starlink

performance is mostly below 100,ms [3], [6], [11], [13].

A. Reproducibility

To enhance the reproducibility of our ﬁndings, we provide access

to the raw data, source code, and supplementary materials associated

with this study. These resources are publicly available under an

open-access license. Although these materials are currently withheld

to maintain the anonymity of the manuscript, we are committed to

transparency and will ensure that all relevant information is acces-

sible upon publication. This approach aligns with best practices in

research integrity and supports the scientiﬁc community’s efforts to

validate and build upon our work.

Figure 4 shows the median latencies in European countries.

Northwestern Europe exhibits the best latencies, probably due to

the presence of more GSs, as illustrated in Figure 5. The southern

and eastern European countries have higher latencies. Greece in

particular has a high average latency. One reason could be the

absence of GSs in the Eastern European region. Italy, on the other

hand, experiences a high average latency, despite GSs being present

in the country. The issue may be related to a more mountainous

topographu, such as northern Italy and Greece.

IV. LONGITUDAL VIEW: 2022–2024

A. Latency

To determine the performance of Starlink, we analyzed TLS

handshake latency over the period 01/2022 to 06/2024. The data

¹https://github.com/diic-starlink/performance-insights-into-starlink

(a) USA

(b) Canada

(d) Philippines

Fig. 3: CDF of RIPE Atlas TLS latencies in the USA, Canada, Germany, and the Philippines from 2022 to 2024.

Packet Loss

Ratio in %

B. Packet Loss

Sent

Received Country

Packet loss was quantiﬁed by analyzing the RIPE Atlas built-

in ping measurement data spanning January 2022 to June 2024.

Table III and Figure 6 present our results. They reveal different

values when comparing countries. Overall, most countries exhibit

packet loss ratios between one and four percent, with some having

even lower values. The Czech Republic demonstrates the lowest

packet loss rate at 0.23 %, followed by Chile at 0.24 %. The

Philippines exhibit the highest packet loss rate at 18.27 %. However,

it is unclear what the underlying pattern is, as some countries exhibit

anomalously high packet loss results, while their adjacent countries

do not (e.g., Germany at 10.52 % and Austria at 0.73 %).

Germany and the USA exhibit peaks in late 2022, followed by

reduced packet loss in 2023. However, the winter of 2023 also

demonstrates an increase in packet loss for all countries in Figure 3,

which might be related to the Starlink user base expansion (see

Section VI-B). The packet loss persisted until June 2024, when the

packet loss for all four visualized countries declined signiﬁcantly. It

remains uncertain whether this pattern will continue in the following

months.

2 150 628

65 021 654

22 727 113

1 176 124

124 263 104

2843

3 626 230

8 092 876

96 089 885

21 714 610

432 934

321 919 833

83 840 509

12 522 224

271 185

18 472 787

37 188 354

4 160 290

127 756

2 134 905

62 438 854

22 211 676

1 168 114

Austria

Australia

Belgium

Benin

0.73

3.97

2.27

0.68

2.50

0.39

0.24

0.23

10.52

3.28

3.24

6.86

3.72

1.62

2.24

1.38

4.32

7.95

4.55

14.19

18.27

0.37

1.14

0.53

3.69

0.58

121 160 149 Canada

2832

3 617 474

8 074 360

85 983 781

21 001 677

Switzerland

Chile

Czechia

Germany

Spain

418 925 Falkland Islands

299 847 062 France

80 720 387

12 318 835

United Kingdom

Greece

265 100 Guam

18 217 243

35 583 136

3 829 356

Haiti

Italy

Kiribati

121 937 Madagascar

20 450 037

26 654 399

18 739 794

7 047 181

9 975 257

578 852 548

18 571 694

17 548 642

21 785 387

18 670 207

6 967 111

9 922 734

Netherlands

Philippines

Poland

Re´union

Sweden

C. Correlation of Latency and Packet Loss

TLS handshake latency and packet loss are closely connected.

As latency increases, we anticipate packet loss to increase and vice

versa. We examined the period from January 2022 to June 2024

and used the overall packet loss and median latency per month to

assess a possible correlation.

We analyzed individual years to gain a better understanding of

the trajectory of the correlation in 2022, 2023, and 2024 (up to

the end of June). Figure 7 shows the individual correlation values

from 2022 to 2024. Overall, the data yields no conclusive results.

The correlation coefﬁcients indicate values that are not close to 0,

1, or -1. However, this varies by country. Some countries show

557 475 491 United States

18 463 935

Virgin Islands, U.S.

TABLE III: RIPE Atlas packet loss from January 2022 to June 2024.

stronger correlation values (e.g., Greece in 2022), while others do

not exhibit signiﬁcant correlations. Overall, we cannot establish a

correlation between latency and packet loss for Starlink. However,

we also cannot assert that the two variables are uncorrelated.

It is probable that other variables need to be taken into account to

draw deﬁnitive conclusions (e.g., the number of users, the capacity

Fig. 4: Heatmap of median latencies in 2024 in Europe from Fig. 5: Map of the Ground Stations in Europe according to the

Cloudﬂare Radar. Unofﬁcial Starlink Global Gateways & PoPs Map [38].

(a) USA

(b) Canada

(d) Philippines

Fig. 6: Packet loss from 2022-01-01 to 2024-06-30 according to RIPE Atlas ping measurements in the USA, Canada, Germany and the

Philippines.

We used the correlation values with the number of probes per

country. This resulted in the following correlation coefﬁcients:

Pearson correlation: ≈ −0.44, Kendall correlation: ≈ −0.44,

Spearman correlation: ≈ −0.53. Since the correlation coefﬁcients

do not approach values close to 0, 1, or -1, we cannot establish

a correlation with the number of probes. It appears that, the lack

of probes does not cause the inconclusive correlation values for

latency and packet loss, but alternative factors that have not yet

been determined.

V. WEEKDAYS AND DIURNAL VARIATIONS

The question remains whether latencies vary across weekdays.

Using the TLS handshake latency measurements built into RIPE At-

las, we analyzed different weekdays from 2022 to 2024. We exam-

ined the measurements using standard statistical metrics (median,

average, maximum, and minimum latency). The results for Germany

are presented in Table V. We can conclude that there is no

discernible pattern that differentiates one weekday from another.

This consistency persists over the years.

Fig. 7: Individual correlation of latency and packet loss in 2024.

of the constellation [39], the intensity of solar magnetic storms

(see Section VI-A), geographical differences between countries, the

presence of GSs).

However, Table V also provides an interesting comparison be-

1) Correlation with the Number of Probes: We sought to explain tween 2022, 2023, and 2024. In 2022, the median TLS handshake

the lack of correlation between latency and packet loss by correlat- latency ranged between 81 and 84 ms. The average latency was

ing the results with the number of probes per country. It is possible approximately 20 ms higher. Peak latencies were as low as 43 ms.

that the lack of data causes a correlation between latency and packet In 2023, the median latencies were signiﬁcantly higher at 94–98 ms.

loss to be undetectable. One possibility is to examine the number The average latency was 14 ms higher at maximum, indicating

of probes available for each country. Therefore, we utilized the data a more stable connection, even if the performance was worse

from Figure 7 and correlated it with the number of probes available than in 2022. However, 2023 achieved a notably improved peak

for RIPE Atlas in each country.

performance of 26 ms. In 2024, Starlink achieved similar median

Mon.

Tue.

Wed.

Thu.

Fri.

Sat.

Sun.

Mon.

Tue.

Wed. Thu.

Fri.

Sat.

Sun.

2022

Med. 20.06 20.08

Avg. 15.81 17.23

2022

20.07

16.97

50.07

0.12

20.08

16.36

50.02

0.03

20.08

15.94

40.96

0.03

20.07 20.05

16.14 16.05

40.51 50.02

Median

Average

100

3090

106

3056

Max. 50.14 50.03

Maximum 1211

Minimum

703

1229

1106

672

Min.

0.05

0.11

0.02

2023

Med.

Avg.

2023

1.07

1.53

1.09

1.97

1.26

2.39

15.60

0.11

1.22

2.58

25.82

0.14

1.22

2.21

25.93

0.15

1.05

2.13

29.11 13.46

0.12 0.14

1.04

1.73

Median

Average

111

1227

108

1233

109

707

107

1088

108

1220

108

1052

Max. 13.43 13.59

Maximum 1245

Minimum

Min.

0.14

0.13

2024

Med. 13.39 12.95

14.05

13.19

25.33

0.92

14.09

13.68

28.49

0.94

14.01

13.44

29.69

0.94

13.80 13.14

13.70 13.14

29.75 29.95

Median

Average

Avg.

Max. 30.23 27.21

Min. 0.95 0.52

12.64 12.31

104

4374

105

3624

Maximum 3147

Minimum 26

1592

4368

1230

1320

0.82

0.81

TABLE IV: Packet loss in % per weekday in Germany according TABLE V: RIPE Atlas TLS latencies in ms per weekday in

to RIPE Atlas ping measurements. Germany.

(a) USA

(b) Canada

(d) Philippines

Fig. 8: Number of hops of traceroute measurements on RIPE Atlas from probes to k.root-servers.org.

and average latencies compared to 2022, while also maintaining days of the week (as demonstrated in Table V).

the peak performance observed in 2023.

VI. NETWORK PATHS

Hops per Route: We conducted a closer examination of the

routing behavior of Starlink. For this, we utilized the traceroute

measurements built into RIPE Atlas probes. Figure 8 shows the

histogram of the number of hops per route for the USA, Canada,

Germany, and the Philippines. The histograms indicate that most

routes require between six and thirteen hops. Note that this does

not include the individual satellites in the Starlink constellation.

These are not detectable by the traceroute measurement due to the

fact that they operate below the IP layer (which traceroute cannot

detect). For the exact number of hops, a traceroute operating below

the IP layer would be necessary.

Latency per Hop: We analyzed the change in latency from hop

to hop (measured in traceroute). Figure 10 shows how the latency

progresses in successive hops. We observed dramatic differences

between the target root servers. Therefore, we focused on traceroute

measurements to the target k.root-servers.org. It becomes apparent

that there is at least one hop that is associated with a signiﬁcant

increase in latency. We hypothesize that this is usually the hop

between the user’s antenna and the provider’s GS. This hop in-

creases latency signiﬁcantly as it is routed through the entire satellite

constellation including signal transmission from and to Earth. The

satellites serve as middleboxes that cannot be detected by using

a traceroute measurement. Therefore, such a behavior is consistent

with the expected network architecture. We also observe that another

substantial increase may occur in a later hop (e.g., in Figure 10b).

To determine the responsible network segment, we have mapped

the most common AS to the speciﬁc hop. The AS is acquired by

using IPinfo data for the IPs associated with each hop. Usually,

Starlink is no longer a dominant part of the trace after the ﬁfth hop.

Therefore, Starlink external network entities are responsible for the

second major increase.

(a) Monday, 2024-04-01

(e) 2024-04-15 to 2024-04-22

(b) Tuesday, 2024-04-02

(d) 2024-04-08 to 2024-04-15

(f) 2024-04-22 to 2024-04-29

Fig. 9: Cloudﬂare Radar latencies for the ﬁrst four weeks of April

2024.

Cloudﬂare Radar provides a different perspective on the data

as it is collected via the Cloudﬂare Radar speedtest, rather than

the TLS handshake measurement used in RIPE Atlas. We utilized

this data to analyze how Starlink performance ﬂuctuates over the

course of a day. Similar results have been observed with RIPE Atlas

A. Inﬂuence of Solar Magnetic Storms

Recent research [9] has claimed that solar magnetic storms have

measurements. For completeness and availability, we chose to a signiﬁcant impact on the performance of Starlink. We have

analyze data from April 2024. We examined single days as well investigated TLS handshake latency data and correlated it with the

as the development over the week. Figure 9a and Figure 9b show intensity of solar magnetic storms. The intensity of solar magnetic

the latency patterns for the ﬁrst two days of April 2024. On April storms is conventionally measured by the Kp index [40] (a value

1st and April 2nd, a diurnal variation is clearly evident. To further between zero and nine). Historic Kp indices are provided by the

investigate the diurnal variation, we analyzed the rest of the month. G. F. Z. [41]. We hypothesized that as the intensity of solar magnetic

Figure 9 shows the weeks between 2024-04-01 and 2024-04-29. The storms increases, the latency would also increase. Therefore, we

latency varies during this time, which we attribute to be caused by used the average Kp index over a single day and correlated it with

diurnal variation. Therefore, we conclude that Starlink exhibits a the median latency over a single day. The following correlations

diurnal variation over the hours of the day, but not across different were obtained: Pearson correlation: ≈ 0.03, Kendall correlation:

(a) USA

(b) Canada

(d) Philippines

Fig. 10: Average latency per hop in the USA, Canada, Germany, and the Philippines on RIPE Atlas.

≈ 0.01, and Spearman correlation: ≈ 0.01. These values are close

VII. CONCLUSION

to zero (i.e., the dimensions are almost orthogonal). Thus, we

concluded that latency and Kp index are not statistically correlated

in our dataset.

In summary, our analysis of historical network measurements data

from January 2022 to June 2024, utilizing metrics from RIPE Atlas

and Cloudﬂare Radar, reveals critical insights into the performance

of networked satellite systems, particularly regarding latency and

packet loss, thereby addressing our research questions on their

operational efﬁciency.

B. Starlink User Numbers

Utilization plays an important role in the variation of the pre-

sented measurements. Unfortunately, the exact user numbers are

not published. However, other sources provide estimates.

Research Question (RQ) 1: How do networked satellite systems

perform in terms of latency and packet loss?

One source of user numbers is Starlink’s X page [42], [43].

It mentions Starlink having more than four million users. AP-

NIC also provides a list of user numbers per AS [44]. In to-

tal, it lists 16 512 033 users distributed among 114 countries.

The most users are concentrated in the USA (2 634 629), Yemen

(1 511 944), the Philippines (1 213 642), Nigeria (1 171 687), and

Mexico (1 122 041). However, those numbers are far higher than the

ones ofﬁcially published by Starlink, which suggests the numbers

are signiﬁcantly overestimated. Additionally, it is not entirely clear

how the numbers were derived.

We analyzed the TLS handshake latency and found that the

Starlink latency was approximately 80 ms median in 2024. The

latency has improved since 2022 to reach 26 ms minimum, 80 ms

median, and 100 ms average latency. We observed a behavior in

the latency characterized by a bimodal distribution. The bimodal

distribution reﬂects the behavior of serving two ranges of latency

particularly well. First lower latencies (≈ 80–100 ms) and second,

higher latencies (≈ 150–250 ms) with a large gap in between.

This pattern appears in most countries and is more pronounced in

2024 compared to 2022. The reason for this pattern is unclear. We

Please note that both sources are not sufﬁciently reliable for hypothesize that the reasons are related to the location of a probe.

rigorous analytical purposes [45]. Speciﬁcally, we postulate that GS positioning, satellite orbits, and

weather are the most relevant factors for varying performance. Dif-

ferent subscription models may also contribute to this phenomenon.

We also examined packet loss and found a wide variation from

country to country. Countries such as the Philippines have packet

loss ratios as high as 18 %, while the Czech Republic and Chile have

less than 0.25 %. On the other hand, Central European countries

such as Germany and the Netherlands have high packet loss ratios.

Overall, most countries have a packet loss ratio of 1 to 4 %.

RQ 2: Do latency and packet loss correlate?

[9] S. Ma, Y. C. Chou, H. Zhao, L. Chen, X. Ma, and J. Liu,

“Network characteristics of LEO satellite constellations: A starlink-

based measurement from end users,” in IEEE INFOCOM 2023 -

IEEE Conference on Computer Communications, New York City,

NY, USA, May 17-20, 2023. IEEE, 2023, pp. 1–10, movement,

Environment, LEO, Performance, Comparison LEO vs Terrestrial ,

TCP vs UDP, Power Consumption, Bent-Pipe. [Online]. Available:

https://doi.org/10.1109/INFOCOM53939.2023.10228912

[10] S. Pekhterev, “The bandwidth of the starlink constellation...and the

assessment of its potential subscriber base in the usa.” SatMagazine,

Nov. 2021.

We expected packet loss and latency to exhibit correlation,

therefore we analyzed their relationship. We separated the values

by year and by country. The results were inconclusive, meaning

we could not establish a correlation between latency and packet

loss. We surmise that there are other parameters that affect the

relationship between latency and packet loss that have not been

adequately investigated. These factors include the version of the

Starlink terminal, the weather, the satellite hardware components,

or the measurement targets.

[11] A. Raman, M. Varvello, H. Chang, N. Sastry, and Y. Zaki, “Dissecting

the performance of satellite network operators,” PACMNET, vol. 1,

no. CoNEXT3, pp. 15:1–15:25, 2023. [Online]. Available: https:

//doi.org/10.1145/3629137

[12] J. Garcia, S. Sundberg, G. Caso, and A. Brunstr o¨m, “Multi-timescale

evaluation of starlink throughput,” in Proceedings of the 1st ACM

Workshop on LEO Networking and Communication, LEO-NET 2023,

Madrid, Spain, 6 October 2023. ACM, 2023, pp. 31–36. [Online].

Available: https://doi.org/10.1145/3614204.3616108

[13] F. Michel, M. Trevisan, D. Giordano, and O. Bonaventure, “A

ﬁrst look at starlink performance,” in Proceedings of the 22nd

ACM Internet Measurement Conference, IMC 2022, Nice, France,

October 25-27, 2022, C. Barakat, C. Pelsser, T. A. Benson, and

D. R. Choffnes, Eds. ACM, 2022, pp. 130–136. [Online]. Available:

https://doi.org/10.1145/3517745.3561416

[14] L. Izhikevich, R. Enghardt, T.-Y. Huang, and R. Teixeira, “A global

perspective on the past, present, and future of video streaming over

starlink,” 2024.

RQ 3: What happens to latency when routing through the Starlink

constellation?

First, we determined the number of hops a route typically

traverses. We found that most routes encompass 6 to 13 hops,

excluding hops through the Starlink constellation, as we can infer

from the change in latency per hop. The histogram of hops also

exhibits the aforementioned bimodal distribution, which may be

related to the pattern we found for latencies.

Finally, we conclude that Starlink currently provides reliable and

consistent performance. It should be noted that Starlink has not yet

reached its full potential, as demonstrated by the trend over the last

few years. The infrastructure is likely to expand and offer enhanced

performance in countries that currently show suboptimal results.

[15] O. O. N. B. V. A. S. Akpala, “Performance test of spacex’s starlink:

An empirical review,” JES. Journal of Engineering Sciences, vol. 52,

no. 5, pp. 73–87, Sep. 2024.

[16] A. Stevens, B. Iradukunda, B. Bailey, and R. Durairajan, “Can LEO

satellites enhance the resilience of internet to multi-hazard risks?”

in Passive and Active Measurement - 25th International Conference,

PAM 2024, Virtual Event, March 11-13, 2024, Proceedings, Part II

, ser. Lecture Notes in Computer Science, P. Richter, V. Bajpai,

and E. Carisimo, Eds., vol. 14538. Springer, 2024, pp. 170–195.

[Online]. Available: https://doi.org/10.1007/978-3-031-56252-5 9

[17] D. Laniewski, E. Lanfer, B. Meijerink, R. van Rijswijk-Deij,

and N. Aschenbruck, “Wetlinks: A large-scale longitudinal starlink

dataset with contiguous weather data,” in 8th Network Trafﬁc

Measurement and Analysis Conference, TMA 2024, Dresden,

Germany, May 21-24, 2024. IEEE, 2024, pp. 1–9. [Online].

Available: https://doi.org/10.23919/TMA62044.2024.10558998

[18] D. Laniewski, E. Lanfer, S. Beginn, J. Dunker, M. Du¨ckers, and

N. Aschenbruck, “Starlink on the road: A ﬁrst look at mobile starlink

performance in central europe,” CoRR, vol. abs/2403.13497, 2024.

[Online]. Available: https://doi.org/10.48550/arXiv.2403.13497

[19] D. Laniewski, E. Lanfer, and N. Aschenbruck, “Measuring mobile

starlink performance: A comprehensive look,” IEEE Open Journal of

the Communications Society, vol. 6, p. 1266–1283, Feb 2025.

[20] T. Fang, A. Kubaryk, D. Goldstein, Z. Li, T. Fuller-Rowell, G. Mill-

ward, H. J. Singer, R. Steenburgh, S. Westerman, and E. Babcock,

“Space weather environment during the spacex starlink satellite loss in

february 2022,” Space Weather, vol. 20, no. 11, Nov. 2022.

APPENDIX A

ETHICAL CONSIDERATIONS

This study does not raise any ethical concerns. It complies with

the ethical guidelines outlined in [46] and [47].

REFERENCES

[1] N2YO, “Live real time satellite tracking and predictions,” 2024.

[Online]. Available: https://www.n2yo.com/

[2] T. Pratt and J. E. Allnutt, Satellite communications. John Wiley &

Sons, 2019.

[3] N. Mohan, A. E. Ferguson, H. Cech, R. Bose, P. R. Renatin, M. K.

Marina, and J. Ott, “A multifaceted look at starlink performance,”

in Proceedings of the ACM on Web Conference 2024, WWW 2024,

Singapore, May 13-17, 2024, T. Chua, C. Ngo, R. Kumar, H. W.

Lauw, and R. K. Lee, Eds. ACM, 2024, pp. 2723–2734. [Online].

Available: https://doi.org/10.1145/3589334.3645328

[21] M. Hapgood, H. Liu, and N. Lugaz, “Spacex—sailing close to the

[4] S. P. Chan, “The birth and demise of an idea: Teledesic’s ’internet in

the sky’,” The Seattle Times, 10 2002.

space weather?” Space Weather, vol. 20, no. 3, Mar. 2022.

[22] M. Handley, “Delay is not an option: Low latency routing

in space,” in Proceedings of the 17th ACM Workshop on

Hot Topics in Networks, HotNets 2018, Redmond, WA, USA,

November 15-16, 2018. ACM, 2018, pp. 85–91. [Online]. Available:

https://doi.org/10.1145/3286062.3286075

[23] S. Zhang and K. L. Yeung, “Scalable routing in low-earth

orbit satellite constellations: Architecture and algorithms,” Comput.

Commun., vol. 188, pp. 26–38, 2022. [Online]. Available: https:

//doi.org/10.1016/j.comcom.2022.02.015

[24] V. Bhosale, A. Saeed, K. Bhardwaj, and A. Gavrilovska, “A

characterization of route variability in LEO satellite networks,” in

Passive and Active Measurement - 24th International Conference,

PAM 2023, Virtual Event, March 21-23, 2023, Proceedings, ser.

Lecture Notes in Computer Science, A. Brunstro¨m, M. Flores, and

[5] D. Barboza, “Iridium, bankrupt, is planning a ﬁery ending for its 88

satellites,” The New York Times, pp. 1–1, 4 2000.

[6] Z. Lai, H. Li, and J. Li, “Starperf: Characterizing network

performance for emerging mega-constellations,” in 28th IEEE

International Conference on Network Protocols, ICNP 2020, Madrid,

Spain, October 13-16, 2020. IEEE, 2020, pp. 1–11. [Online].

Available: https://doi.org/10.1109/ICNP49622.2020.9259357

[7] S. Kassing, D. Bhattacherjee, A. B. Aguas, J. E. Saethre, and

A. Singla, “Exploring the ”internet from space” with hypatia,” in IMC

’20: ACM Internet Measurement Conference, Virtual Event, USA,

October 27-29, 2020. ACM, 2020, pp. 214–229. [Online]. Available:

https://doi.org/10.1145/3419394.3423635

[8] S. Segan, “Tested: Spacex’s starlink satellite internet service is fast,

but it’ll cost you,” PC Mag, 10 2020.

M. Fiore, Eds., vol. 13882. Springer, 2023, pp. 313–342. [Online].

Available: https://doi.org/10.1007/978-3-031-28486-1 14

[44] APNIC, “Customers per as measurements.” [Online]. Available:

https://stats.labs.apnic.net/aspop/

[45] L. Salamatian, C. Ardi, V. Giotsas, M. Calder, E. Katz-Bassett, and

T. Arnold, “What’s in the dataset? unboxing the apnic per as user

population dataset,” in Proceedings of the 2024 ACM on Internet

Measurement Conference, ser. IMC ’24. New York, NY, USA:

Association for Computing Machinery, 2024, p. 165–182. [Online].

Available: https://doi.org/10.1145/3646547.3688411

[25] Y. Hauri, D. Bhattacherjee, M. Grossmann, and A. Singla, “”internet

from space” without inter-satellite links,” in HotNets ’20: The 19th

ACM Workshop on Hot Topics in Networks, Virtual Event, USA,

November 4-6, 2020, B. Y. Zhao, H. Zheng, H. V. Madhyastha,

and V. N. Padmanabhan, Eds. ACM, 2020, pp. 205–211. [Online].

Available: https://doi.org/10.1145/3422604.3425938

[46] B. Jones, R. Ensaﬁ, N. Feamster, V. Paxson, and N. Weaver, “Ethical

Concerns for Censorship Measurement,” in Proceedings of the 2015

ACM SIGCOMM Workshop on Ethics in Networked Systems Research.

London United Kingdom: ACM, Aug. 2015, pp. 17–19. [Online].

Available: https://dl.acm.org/doi/10.1145/2793013.2793015

[47] C. Partridge and M. Allman, “Ethical considerations in network

measurement papers,” Communications of the ACM, vol. 59, no. 10,

pp. 58–64, Sep. 2016. [Online]. Available: https://dl.acm.org/doi/10.

1145/2896816

[26] G. Stock, J. A. Fraire, and H. Hermanns, “Distributed on-demand

routing for leo mega-constellations: A starlink case study,” in 2022

11th Advanced Satellite Multimedia Systems Conference and the 17th

Signal Processing for Space Communications Workshop (ASMS/SPSC),

2022, pp. 1–8.

[27] Z. M. Kassas, S. E. Kozhaya, H. Kanj, J. Sarouﬁm, S. W. Hayek,

M. Neinavaie, N. Khairallah, and J. Khalife, “Navigation with

multi-constellation LEO satellite signals of opportunity: Starlink,

oneweb, orbcomm, and iridium,” in IEEE/ION Position, Location

and Navigation Symposium, PLANS 2023, Monterey, CA, USA,

April 24-27, 2023. IEEE, 2023, pp. 338–343. [Online]. Available:

https://doi.org/10.1109/PLANS53410.2023.10140066

[28] D. P. Shorten, W. Karunarathne, and M. Roughan, “How is starlink

manoeuvring? an analysis of patterns in the manoeuvres of starlink

satellites,” in Proceedings of the 9th International Conference

on Internet of Things, Big Data and Security, IoTBDS 2024,

Angers, France, April 28-30, 2024, A. Kobusinska, A. Jacobsson,

and V. Chang, Eds. SCITEPRESS, 2024, pp. 174–184. [Online].

Available: https://doi.org/10.5220/0012619900003705

[29] D. Vasisht, J. Shenoy, and R. Chandra, “L2D2: low latency distributed

downlink for LEO satellites,” in ACM SIGCOMM 2021 Conference,

Virtual Event, USA, August 23-27, 2021, F. A. Kuipers and

M. C. Caesar, Eds. ACM, 2021, pp. 151–164. [Online]. Available:

https://doi.org/10.1145/3452296.3472932

[30] X. Lin, S. Cioni, G. Charbit, N. Chuberre, S. Hellsten, and J. Boutillon,

“On the path to 6g: Embracing the next wave of low earth orbit satellite

access,” IEEE Commun. Mag., vol. 59, no. 12, pp. 36–42, 2021.

[Online]. Available: https://doi.org/10.1109/MCOM.001.2100298

[31] L.

spacex

[Online].

Wouters,

“Dumping

user terminal

and

extracting

the

starlink

ﬁrmware,”

2021.

Available:

https://www.esat.kuleuven.be/cosic/blog/

dumping-and-extracting-the-spacex-starlink-user-terminal-ﬁrmware/

[32] C. Ramponi, “Diving into starlink’s user terminal ﬁrmware,” 8 2023.

[Online]. Available: https://blog.quarkslab.com/starlink.html

[33] RightsLabDigital, “Starlink in sudan: A lifeline or war facilitator?”

GlobalVoices, Aug. 2024. [Online]. Available: https://globalvoices.

org/2024/08/20/starlink-in-sudan-a-lifeline-or-war-facilitator/#:

^∼:text=Starlink’sstrategicroleinconﬂictzones&text=Itisalsousedin,

hotspotsaffectedbyarmedconﬂict.

[34] N. P. W. A. M. F. Davey-Attlee, “Ukraine relies on starlink for its

drone war. russia appears to be bypassing sanctions to use the devices

too,” CNN World, Mar. 2024.

[35] J. Abels, “Private infrastructure in geopolitical conﬂicts: the case of

starlink and the war in ukraine,” European Journal of International

Relations, Jun. 2024.

[36] RIPE-NCC, “Ripe atlas,”

//atlas.ripe.net

2025. [Online]. Available: https:

[37] Cloudﬂare, “Cloudﬂare radar,” 3 2025. [Online]. Available: https:

//radar.cloudﬂare.com/

[38] panuvic, “Unofﬁcial starlink global gateways and pops.”

[Online]. Available: https://www.google.com/maps/d/viewer?mid=

1805q6rlePY4WZd8QMOaNe2BqAgFkYBY

[39] D. Rozenvasser and K. Shulakova, “Estimation of the starlink global

satellite system capacity,” 2023.

[40] J. Bartels, “Solar inﬂuences of geomagnetism,” Proceedings of the

National Academy of Sciences, vol. 43, no. 1, pp. 75–81, 1957.

[Online]. Available: https://www.pnas.org/doi/abs/10.1073/pnas.43.1.

[41] Helmholtz-Zentrum Potsdam

Deutsches GeoForschungsZentrum

GFZ, “Geomagnetic kp index,” Website, 2023. [Online]. Available:

https://kp.gfz-potsdam.de/en/

[42] Starlink, “X - starlink usernumbers september 2023,” 9 2023. [Online].

Available: https://x.com/Starlink/status/1705695980325323023

[43] ——, “X - starlink usernumbers september 2024,” 9 2024. [Online].

Available: https://x.com/Starlink/status/1839424733198344617