Network Characteristics of LEO Satellite Constellations:

A Starlink-Based Measurement from End Users

Sami Ma∗ , Yi Ching Chou∗ , Haoyuan Zhao∗ , Long Chen∗ , Xiaoqiang Ma† , Jiangchuan Liu∗
∗ School
of Computing Science, Simon Fraser University, Canada
† CSIS Department, Douglas College, Canada

Emails: {masamim, ycchou, hza127}@sfu.ca; {longchen.cs, mxqcs}@ieee.org; [email protected]

Abstract—Low Earth orbit Satellite Networks (LSNs) have

been advocated as a key infrastructure for truly global cov- Starlink
Ping
arXiv:2212.13697v1 [cs.NI] 28 Dec 2022

erage in the forthcoming 6G. This paper presents our initial

Ground

AWS servers
measurement results and observations on the end-to-end network

(9 regions)
Station
characteristics of Starlink, arguably the largest LSN constellation
to date. Our findings confirm that LSNs are a promising solution
towards ubiquitous Internet coverage over the Earth; yet, we also Raspi 4B Starlink Starlink
find that the users of Starlink experience much more dynamics in router dish
throughput and latency than terrestrial network users, and even
Lab computer ISP router Internet
frequent outages. Its user experiences are heavily affected by
environmental factors such as terrain, solar storms, rain, clouds,
and temperature, so is the power consumption. We further Fig. 1. Starlink measurement with diverse scenarios and tools.
analyze Starlink’s current bent-pipe relay strategy and its limits,
particularly for cross-ocean routes. We have also explored its
mobility and portability potentials, and extended our experiments
from urban cities to wild remote areas that are facing distinct Yet, LSNs remain in their early stages with limited user
practical and cultural challenges. bases and user experiences. There are novel theoretical and
simulation-based studies on satellite communications for both
I. I NTRODUCTION GEO and LEO. There exist, however, few comprehensive
Low Earth orbit (LEO) satellites operate at around 180 studies that have a deep insight into the practical performance
km to 2,000 km above the Earth surface, which, compared of large-scale LSNs, except for scattered discussions in fo-
to the traditional Geosynchronous orbit (GEO) satellites at rums or with simulations [2], [3]. The complex and dynamic
around 35,780 km, enable much shorter delays and much topologies, as well as the heterogeneous and black-box ar-
higher throughput for space-ground communications, albeit chitectures incurred during incremental deployment, further
with smaller coverage.1 A large number of such LEO satellites, complicate how outstanding LSNs perform. To accelerate the
however, can form an LEO Satellite Network (LSN) constel- design, deployment, and optimization of LSNs, a systematic
lation that collectively offers high-quality network services to measurement is indispensable, though challenging.
ground users with truly global coverage. It has been suggested In this paper, we present our initial results and observations
as a key infrastructure towards the upcoming 6G networking on a systematic LSN measurement study. We focus on the Star-
and beyond [1]. We have seen commercial deployment of LSN link network, which is arguably the largest LSN constellation
constellations in the past decade with rapidly growing attention to date, in terms of both the number of satellites and the user
from the general public. One of the industrial leaders OneWeb, base. We are particularly interested in the end-to-end network
is building a constellation of 648 broadband satellites, which characteristics and performance with diverse configurations
would eventually expand to 7,000.2 Its rival, SpaceX’s Starlink, and applications. These are also the focus of most general
has launched more than 2,000 satellites to LEO and has users of the Internet, the mass consumer market that is targeted
received approval from the Federal Communications Commis- by Starlink. As a matter of fact, the Starlink network operator,
sion (FCC) to bring that number up to 12,000. FCC has further SpaceX, offers a plug-and-play black-box service to common
authorized Starlink to launch 7,500 generation 2 satellites in end users, who can simply connect to a Starlink router through
the near future.3 The next-generation Starlink constellation WiFi or Ethernet, without knowing the satellite communication
may eventually harbor up to 30,000.4 details and the constellation operations. We are therefore
interested in the following key questions:
A short version of this paper is to appear in IEEE INFOCOM 2023
(submitted on Aug. 1st, 2022; accepted on Dec. 1st, 2022) • Can today’s Starlink achieve comparable performance as
1 https://earthobservatory.nasa.gov/features/OrbitsCatalog
2 https://oneweb.net/resources/oneweb-streamlines-constellation
a typical modern terrestrial network does for end users?
3 https://www.cnbc.com/2022/12/01/fcc-authorizes-spacex-gen2-starlink- • What are the key influential factors of the Starlink net-
up-to-7500-satellites.html work services as perceived by end users?
4 https://www.space.com/spacex-starlink-satellites.html • Has Starlink achieved global coverage through its con-

1
 stellation? If not, what are the challenges? ISL

Our Starlink measurement, involving four dishes, has lasted

over half a year starting from the beginning of 2022. We use UL GSL
Internet
scripts to automatically launch various tools and applications
to communicate to servers distributed in different regions
worldwide. We have performed field tests in urban cities and Starlink Dishes Ground Station

remote wild areas with a per-region collection of over 3 million

(a) Starlink’s first shell (b) Starlink’s network structure.
total records. Our findings confirm that LSNs are a promising generated by Hypatia [2].
solution towards ubiquitous Internet coverage over the Earth;
yet, we also identify a series of significant issues on the Fig. 2. (a) The constellation of the Starlink’s first shell; (b) The Starlink’s
network structure with ISLs, GSLs, and ULs.
current Starlink’s network services; in particular, the following
observations:
1) The throughput and latency experienced by a Starlink the interaction between LEO satellites to ground stations [15],
user are much more dynamic compared to terrestrial [16], [19]–[21], constellation topology management, address-
networks; there are frequent outages, too. ing and routing [9] [22], as well as the potential impact of
2) Starlink’s user experience can be heavily affected by environmental factors, such as rain [15] and solar storms [23].
environments such as terrain, solar storm, rain, cloud, Recent years have seen real-world deployments like
and temperature, so is its power consumption. SpaceX’s Starlink, OneWeb, and Telesat. There have been
We are also interested in the impact of the constellation’s numerous articles and posts on these LSNs, in particular,
topology. Instead of establishing Inter-Satellite Links (ISLs) Starlink, from newsmedia and user forums,5 most of which
between satellites, Starlink is currently using a bent-pipe were simple reviews. Studies on LEO constellation networks
strategy [4], [5], where a LEO satellite must relay the traffic have largely relied on simulators, such as Hypatia [2]
to a Ground Station (GS) for further routing [6], as shown in and StarPerf [3]. They are good for model and design
Fig. 1. We have validated that the bent-pipe is generally of validation, but real-world measurement for such large-scale
one hop and involves the closest GS only before switching to complex systems as Starlink is irreplaceable. We have seen
the terrestrial network, instead of multiple. We analyze this earlier measurements [24]–[28], which were conducted with
strategy and its potential limits, particularly for cross-ocean relatively simple setups and limited dynamics of influential
routes. factors, and often focused on physical layer behaviors. There
We also examine Starlink’s mobility and portability support. have been recent measurement works with Ookla’s speedtest6
Even though throughput does not change much during move- and a browser extension [29], [30]. They have examined QUIC
ment, our experiments suggest that mobility with Starlink still performance and browser latency with Starlink. Our work,
requires significantly more work as outages happen frequently in parallel time as theirs, attempts to provide a systematic
with high latency. We further extend our experiments from measurement on Starlink’s performance from the common end
urban cities to the wild remote areas which reveal a series of users’ perspective. Our measurement has a more global cov-
issues regarding global coverage, including unique practical erage with distant endpoints, and considers video streaming,
and cultural challenges. remote wild northern areas, and power consumption.
The remainder of this paper is organized as follows. The
III. S YSTEM S ETUP AND M ETHODOLOGY
related works are reviewed in Section II, followed by the
measurement methodology in Section III. We then present and A. Starlink Network Overview
analyze the measurement results in an urban city in Section IV SpaceX’s Starlink is a prominent representative of the
and those in the wild in Section V, respectively. We discuss state-of-the-art LEO satellite networks, serving over 400,000
a series of advanced issues in Section VI and conclude the subscribers worldwide as of today. Its system consists of three
paper with future directions in Section VII. main components: a constellation of LEO satellites, a network
of ground stations, and user terminals. The constellation
II. R ELATED W ORK currently has over 2,000 satellites at different LEO groups,
There have been many pioneering works on satellite com- known as shells [2], [5]. Most of the deployed satellites are in
munications, from theoretical analysis [7]–[10], to system the first shell, which contains 72 orbits at an altitude of 550
design [4], [5], [11]–[13], and to practical deployment [14]– km in planes inclined 53°, each accommodating a maximum
[16]. Early measurement on real-world practical systems or of 22 satellites (see Fig. 2(a)).
simulations have focused on GEO satellite networking [17], The satellites employ phased array beamforming and digital
[18]. Ground-to-satellite distance and hence the signal latency processing in the Ku- and Ka-bands for data communications
is a physical barrier that can only be alleviated by moving [31]. At the ground, instead of directly connecting from
to the LEO. In the past decade, LEO constellations have re- 5 www.reddit.com/r/Starlink/comments/jllpet/starlink beta report
ceived much attention as a means towards ubiquitous Internet specifications dimensions/
connectivity beyond 5G. There have been extensive studies on 6 https://www.speedtest.net/

2
 Clear View Obstructions
B. Measurement Topology and Environment
Our measurement study started in January 2022 and lasted
7 months. We have deployed 4 dish kits in different locations
for data generation, communication, and collection. Among
them, three are Gen 1 (referred to as A, B, and C, respectively)
Dish A: 2.7% Dish B: 4.7% Dish C: 24.9% Dish F: 0.0% and one is Gen 2 (referred to as F) which is easier to carry
Fig. 3. Samples of Starlink visibility maps, together with obstruction ratios to remote areas. Our experiments span over a wide range of
taken from the Starlink app logs. Dishes A and B are in Burnaby; dish F is terrains, including major cities, namely Vancouver (Burnaby
in Koeye Point.
and Coquitlam in the metro-Vancouver, in particular) in British
Columbia (BC), Canada, with open sky spaces, remote areas
with steep valleys or at the far north (Koeye Point, BC, which
handsets to satellites, a dish (nicknamed as Dishy McFlatface)
is near the current service boundary of Starlink), temperate
is needed to communicate through a User Link (UL) with
rain forests with heavy obstacles, etc. Dish A and B are sta-
a LEO satellite that is currently visible and accessible, and
tioned in Burnaby whereas dish C is setup in Coqutilam, and
the dish then serves local user devices. Two generations of
Dish F is stationed at Koeye Point. To collect the measurement
dishes are available for consumers: the round-shaped Gen 1
data with minimum signal, workload interference, and power
and the rectangular Gen 2. Gen 1 spans 23.2” in diameter and
demands, a Raspberry Pi 4B (Raspi) was directly connected
weighs in at 7.3 kg, whereas Gen 2 is slightly smaller (19” x
to each Starlink router through an Ethernet cable, except for
12”) and lighter (4.2 kg) with more bandwidth streams in the
dish F which has a WiFi interface only. We have performed
router (3x3 MU-MIMO versus 2x2 in Gen 1). The latter has
measurements to 9 destination servers deployed using Amazon
better portability but remains bulky. Not only the sky needs
Web Services (AWS), which are geo-distributed worldwide:
to be clear, but the setup location needs to be wide enough to
Sao Paulo, Singapore, Sydney, North California, Bahrain,
accommodate the dish size and any movement required for its
Tokyo, London, Cape Town, and Mumbai. This overall setup
self-adjusted tilt. A visibility map provided by the Starlink app
can be seen in Fig. 1.
can help with the placement (see Fig. 3 for an example), which
For the experiments where domestic terrestrial Internet
also gives a numerical obstruction ratio in its debug mode. We
access is available, we have used a PC with an i7-10700KF
have used both dishes in our measurement and found that their
CPU and the same AWS servers as the baseline for com-
network services experienced by end users are largely similar,
parison. The baseline presented in this paper is running over
though Gen 2 seems more friendly to TCP.
a terrestrial Cable service with a maximum download speed
Different from early satellite services that often serve a of 800 Mb/s. We have experimented with different terrestrial
niche market, Starlink targets the next generation broadband broadband Internet services of similar maximum speeds, and
Internet service for the global consumer market. To this have observed similar baseline performance. We also used
end, it provides end users a plug-and-play Internet access an Emporia smart plug8 to monitor the instant and historical
interface that hides the aforementioned technical details of power consumption of the dish kit. To understand the impact
satellite communications. The user kit consists of three main of weather conditions, we extract hourly data from a climate
components: the dish, a power-over-Ethernet power supply station at Vancouver Harbour, which is approximately 10 km
adapter, and a router with WiFi and Ethernet port (Gen 1) away from dish A.
or WiFi only (Gen 2) for local devices. The particular satellite
connecting to the dish at a time will act as a repeater, relaying C. Target Applications and Measurement Tools
the signal through the Ka-band to a Ground Station (GS) Our measurement cover different protocols and applications,
through a GS-to-Satellite Link (GSL), which further connects from standard Web browsing and file transfer to high-demand
to the global Internet, or vice versa (see Fig. 2(b)). Currently, video streaming. We use iperf3 to measure the TCP and
there are at least 65 GSes located in North America and 23 UDP throughput in the transport layer. We used cubic TCP
GSes in Europe, as well as others scattered worldwide.7 To as the default, though further studies would be conducted to
successfully establish this bent-pipe for traffic relay, the GS compare with other variants. For each destination server, two
and the dish must be both in the field of view of the satellite iperf3 are launched at different ports for downloading and
at the same time. Therefore, their distance should not exceed uploading, respectively, allowing simultaneous measurements
about 1,000 km. Also note that a LEO satellite has a fly- of both. We also use the Secure Copy Protocol (SCP) to
over dwell time of only around 10 minutes over one spot copy files with different data sizes bi-directionally to the AWS
on the ground [15]; when it moves out, the service needs to regions. fallocate has been used to create files with data
be handover to a new satellite moving into the area (if there sizes ranging from 1 to 300 MB. ping and traceroute
is). Starlink has started laser-based ISL (Inter-Satellite Link) have been used to track and analyze the network path charac-
experiments; yet large scale implementation will not happen teristics and routing strategies. We have used the Starlink Mo-
till the end of 2022 [32]. bile App and YouTube’s built-in tool Stats for nerds to
7 https://starlink.sx/ 8 https://www.emporiaenergy.com/emporia-smart-plug

3
 Terrestrial Regions
1.0 N. California
Regions 300
N. California Tokyo
CDF 0.5 London

Latency (ms)
Tokyo
London 200 Sydney
0.0 Singapore
Sydney
Starlink Singapore Sao Paulo
1.0 100 Bahrain
Sao Paulo
Bahrain Mumbai
CDF

0.5 Mumbai Cape Town

0 1 2 3 4 5 6 7 Type
Cape Town 6-1 6-1 6-1 6-1 6-1 6-1 6-1 6-1
0.0 2-0 2-0 2-0 2-0 2-0 2-0 2-0 2-0 Terrestrial
100 200 300 400 202 202 202 202 202 202 202 202 Starlink

Fig. 4. Cumulative distribution of latency (ms) globally. Fig. 6. Hourly latency over time during a one week period.

Terrestrial
300 Starlink Ground Station
Latency (ms)

User Dish
Starlink Service Coverage
200 Starlink Satellite
Terrestrial Router
Starlink Path
100 Terrestrial Path

0
nia Tokyo ndon ydney apore Paulo hrain umbai Town
a lifor Lo S Sing Sao Ba M ape
N. C C
Regions

Fig. 5. Average latency in different regions. Vertical lines correspond to

standard deviations of data samples.

collect related network, user, and application information. We Fig. 7. Current Starlink network routing framework.
have designed Python-based scripts to automatically launch the
tests and collect the output data, which will be made publicly
available. Bahrain. Starlink’s latency being almost always just slightly
above that of the terrestrial network suggests that only a
IV. U RBAN C ITY M EASUREMENT: S TARLINK VERSUS small number of hops increase this latency, coinciding with
T ERRESTRIAL N ETWORKING the idea of the single bent-pipe transmission framework, i.e.,
We start our measurement in Vancouver, a major city at the only the first hop uses a bent-pipe [14]. Otherwise, if multiple
Pacific West Coast with all the state-of-the-art terrestrial Inter- bent-pipe transmissions are involved, there would be a larger
net infrastructures and Starlink coverage. We have deployed variation of the latency discrepancy between Starlink and the
the dishes in different locations in the city’s residential areas terrestrial network, and Starlink would have lower latency
with sufficient open surrounding spaces, and hence should to some destinations. Sydney being an outlier also reaffirms
reveal the best performance of Starlink. this. When Starlink started accepting requests for a Maritime
version of their dish9 in July 2022, its coverage map only spans
A. End-to-End Latency
across the coastal waters around in-land regions. A switch to
We have measured the latency across 9 AWS regions a terrestrial network would be required to cross the Atlantic
worldwide and looked at the difference between Starlink and or Pacific oceans through submarine fibre. For example, to
the terrestrial network’s latency. In summary, we have found communicate with Sydney, Starlink currently chooses to route
that Starlink’s latency is slightly higher (10%) than that of the packets through a Seattle AWS server after the first hop bent-
terrestrial network on average and is much more unstable (see pipe.
Fig. 4, 5). Obstruction, satellite movements, and ISP routing 2) Latency Variation: Starlink’s latency has a very high
decisions are potential factors that affect Starlink’s latency. variation, around 3.8 times that of the terrestrial network
1) Latency across Regions: We can see noticeable gaps (Shown in Fig. 5). Given that the terrestrial network keeps
between the latencies of Starlink and that of the terrestrial a relatively stable latency as seen in Fig. 6, the common rise
network in Fig. 4 with the average gaps shown in TABLE I. and falls for Starlink not observed in the terrestrial network are
The gaps typically range from 1.8 to 22.8 ms away from likely due to the constant network path changes as satellites
the terrestrial network, with Sydney being an outlier (3.4x to move closer or further away from the dish, affecting the
43.6x larger gap). There is a noticeable increase in latency latency of the single bent-pipe transmission. As shown in
for Bahrain, but since the terrestrial network also observes Fig. 7, when a connected satellite S is moving away from the
such an increase, the dish is likely not the culprit. This
could be due to a connection issue with the AWS server at 9 https://www.starlink.com/maritime

4
 TABLE I. Average Latency Gaps between Starlink and Terrestrial Networks
Bahrain Cape Town London Mumbai
N. California Sao Paulo Singapore Sydney Tokyo
(µs − µt )∗
Gap 8.4 14.3 16.0 12.9 20.9 22.8 1.8 78.4 9.2
∗ where µ and µ are, respectively, the Starlink’s and terrestrial network’s average latency (ms).
s t

TCP-T TCP-S UDP-T UDP-S

Transfer Time (s)

3 Terrestrial
10 75
Throughput (Mb/s)

Starlink
2 50
10
1
25
10
0
0
10 0 50 100 150 200 250 300
Data Sizes (MB)

(a) Download throughput

Fig. 9. Transfer time as a function of SCP data size (translucent area outlines
3
regression estimates).
10
Throughput (Mb/s)

2
10
1
10 packet losses that are much higher than the terrestrial network.
10
0 Such losses are not necessarily caused by congestion, as we
i
rnia Tokyo ondon ydney apore Paulo ahrain umba Town
have observed them with a low UDP throughput experiment
alifo L S Sing Sao B M ape
N. C C as well. For 20% of the presented throughput, which is low
enough to mitigate congestion factors of packet loss [33],
(b) Upload throughput we still see bursty losses of 0.24% and 1.24% on average
for downloads and uploads, respectively. This is consistent
Fig. 8. Throughput measurements across different regions for the terrestrial
(TCP-T/UDP-T) and Starlink (TCP-S/UDP-S) network, respectively. Vertical with early simulation studies for satellite communications
lines correspond to standard deviations of data samples. [34]. Interestingly, there is also a cycle around 12 hours. The
average and peak values differ from cycle to cycle but are
similar across regions during each cycle, suggesting that the
user dish (from S t0 to position S t1 , where the superscripts t0 losses happen most likely on the bent-pipe, in particular, the
and t1 represent the satellite S at different times), the latency UL, which will be further examined in our analysis of the
would increase continuously until a handover to a new satellite routing strategy in Section IV-C.
happens. This is also validated by a closer look at the arrival
time at the packet level — the jitters of Starlink are 0.255 ms For TCP traffic, Starlink has experienced more issues with
(download) and 2.715 ms (upload), both of which are over congestion control than the terrestrial network. Our results
5x higher than that of the terrestrial network (0.041 ms for show that Starlink has a bandwidth utilization [33] of 39.0% ,
download and 0.546 ms for upload, respectively). which is noticeably lower than that of the terrestrial network
(46.8%), suggesting TCP’s congestion control is quite sensitive
B. End-to-End Throughput to the dynamics in the satellite hop. The newer Gen 2 dish
We have measured the throughput of the Starlink dish using may optimize TCP uploads with a 1.76x more throughput as
iperf3 with both TCP and UDP. As shown in Fig. 8, compared to the Gen 1 dishes in the urban city. Unfortunately,
Starlink achieves good throughput (∼80 Mb/s on average) but UDP uploads do not see a similar increase, which is 0.73x
again lacks stability. The standard deviations of throughput of the Gen 1 dishes in terms of throughput. This suggests
for Starlink and terrestrial networks are 50.71% and 34.44%, that the improvement of TCP uploads could be due to a more
respectively. Again, this variation is likely due to constantly optimized access method in the updated dish or the revised
changing network paths from satellite movements and han- router design.
dovers, in addition to potentially low-cost terrestrial network We have also examined bulk data transfer through SCP. As
routing after being passed over from the Starlink network. displayed in Fig. 9, compared to the terrestrial network, there
For UDP measurements, we gradually increase iperf3’s is a much stronger upward trend for Starlink in transfer time
traffic flow until the path is saturated. Starlink’s achieved with higher variance when data sizes grow. While there is
throughput is similar across all regions (see TABLE II). This almost no difference for the 1 MB data size likely due to
suggests either the dish’s hop is physically the bottleneck SCP overhead, SCP over Starlink almost doubles in transfer
or traffic engineering has been applied to this hop at UL time compared to the terrestrial network for the 200 MB data
or GSL. The latter would be the root cause as the results size. This suggests that network dynamics of the satellite hop
in IV-D3 suggest that the throughput may be throttled for accumulate overtime and hence have amplified the impact with
power consumption control. We have also observed bursty larger data size in SCP.

5
 TABLE II. Average Throughput (Mb/s) across Regions and Protocols.
Terrestrial Starlink
TCP UDP TCP UDP
Download Upload Download Upload Download Upload Download Upload
N. California 798.89 64.21 804.97 104.42 107.77 6.26 201.42 10.41
Tokyo 647.55 42.69 807.79 99.96 87.28 6.24 205.72 11.18
London 644.71 43.98 806.91 103.49 78.51 6.08 203.56 10.16
Sydney 238.38 37.66 804.55 98.67 70.72 7.16 200.08 10.10
Singapore 475.39 43.25 805.45 101.02 78.11 6.33 202.50 10.56
Sao Paulo 481.63 47.68 807.84 103.21 73.83 6.13 205.39 10.49
Bahrain 398.05 54.07 810.59 103.20 74.49 6.14 194.27 9.90
Mumbai 346.69 51.35 803.79 101.65 66.43 6.29 196.20 10.33
Cape Town 318.90 55.62 816.03 105.58 68.51 6.34 193.67 10.07

Starlink Terrestrial
Bahrain 1.0
Cape Town 0.53 0.51
London 0.44 0.53 0.35 0.54 0.5
Mumbai 0.52 0.62 0.53 0.24 0.48 0.56
N. California 0.37 0.54 0.51 0.49 0.082 0.25 0.17 0.15 0.0
Sao Paulo 0.51 0.63 0.52 0.57 0.48 0.064 0.26 0.16 0.15 0.75
Singapore 0.47 0.62 0.56 0.57 0.46 0.49 0.041 0.3 0.18 0.15 0.65 0.66 0.5
Sydney 0.35 0.7 0.49 0.64 0.39 0.5 0.6 0.097 0.19 0.13 0.11 0.61 0.6 0.62
Tokyo 0.51 0.53 0.58 0.5 0.54 0.56 0.61 0.54 0.079 0.24 0.16 0.14 0.74 0.73 0.67 0.61 1.0
Bahrain

London

Mumbai

Singapore

Sydney

London

Mumbai

Singapore

Sydney
Sao Paulo

Bahrain
Tokyo

Sao Paulo

Tokyo
Cape Town

Cape Town
N. California

N. California
Fig. 10. Pearson latency correlation heatmap between regions.

C. Routing Strategy considering that London is closer to US East, the traffic from
our measurement location (at the Pacific West Coast) should
The results from our traceroute measurements suggest
be routed by Starlink to GSes in US East using multiple bent-
that the Starlink network only does one bent-pipe communica-
pipe transmissions before switching to the terrestrial network,
tion (e.g., the first UL and GSL) along the route and then enters
which should have lowered the correlations against other
the terrestrial network to arrive at the destination, or vice versa.
regions not in the same general direction as London, such
In our experiments, the Starlink network will always connect
as N. California, Sydney, and Tokyo. Unfortunately, this is
to the geographically closest GS from the dish no matter where
currently not the case as the correlations shown in Fig. 10 are
the destination is. We only see a single SpaceX Services ISP
moderate.
entry located at a GS nearest to the dish before the routing
So far our discussion has been limited to the case of only
switches to a terrestrial ISP. To the best of our efforts, we
one end being a Starlink user. When Starlink becomes a
have found no reference of Starlink using L2 routing so far.
ubiquitous Internet service, frequently, both network endpoints
As shown in Fig. 6, the regions, especially Sydney, Sao Paulo,
would be dish users. Ideally, one would expect that, when two
and London have similar latency patterns for the Starlink
nearby dishes are both within the service area of a satellite,
network. We use the Pearson correlation between all the
they can peer-to-peer exchange data using the satellite to
regions to further analyze this point.10 The correlation matrices
relay their respective ULs, without going through GSLs or
in Fig. 10 show that all the regions the dish communicates
other terrestrial Internet nodes. To verify this, we have placed
with have moderate correlations, further suggesting they have
dishes A and B together, attempting to connect them directly.
similar latency patterns. Despite communicating with different
Unfortunately, we have discovered that only ssh connections
regions, both observations suggest that similar satellites, GSes,
are allowed by Starlink in this setup. In fact, Starlink prevents
and routes have been used. In other words, this single-bent-
users from editing more advanced features in typical routers
pipe architecture does not participate in the routing strategy
such as port forwarding. In the end, an external proxy must be
other than serving as the first hop to bridge to the Internet (or
used, e.g. through remote.it,11 which, being customized
last hop should the Starlink user be the destination). Otherwise,
for Starlink, sets up an external AWS server as the public
10 Pearson correlation works by drawing a best fit line through the dat-
proxy. The throughput averaged 2-3 Mb/s with occasional
apoints on a plot and providing a coefficient that describe the distribu- bursts to 9-12 Mb/s; however, 27% of the throughput read-
tion from the line to denote the correlation between two variables. The ings are simply zero. Clearly, the tunnel-based solution does
closer the Pearson coefficient is to 1, the higher the correlation between
the two variables. See https://statistics.laerd.com/statistical-guides/pearson-
correlation-coefficient-statistical-guide.php 11 https://www.remote.it/resources/how-to-port-forward-on-starlink

6
not work well yet, needing improvements for pair-wise dish TCP UDP
communications. Download Upload

Throughput (Mb/s)
D. Environmental Influential Factors 200 20
1) Obstruction: The Starlink App gives the sky visibility
map of the dish at the current location (Fig. 3), together 100 10

with the obstruction status. The latter is only visible in the

debug mode of the App through ObstructionStats → 0 2 4 0 2 4
Precip. Amount (mm) Precip. Amount (mm)
CurrentlyObstructed and FractionObstructed.
We referred to the fraction as obstruction ratio in this paper, Fig. 11. Starlink throughput with precipitation (vertical lines correspond to
which is not always correlated with the visibility map. It takes 95% confidence interval; translucent area outlines regression estimates).
around 12-24 hours to build the initial map, but the obstruction
ratio is updated more frequently to capture live dynamics, e.g.,
fallen leaves that cover the dish. We have found that, even if downloads is on average 215 Mb/s with no precipitation and
a location has a good visibility map, a transient obstruction almost halves to around 120 Mb/s with 4.1 to 5.2 mm of
due to foreign objects can lead to spikes in latency or even precipitation. Rain attenuation is the likely culprit for the drop
network outage, implying that the dish is very sensitive to in throughput as Ka- and Ku-band radio waves can suffer
surface obstructions.12 In short, Starlink’s end user needs both heavily in rainfall [36], [37]. Additionally, clouds that form
a clear sky and a clean dish to achieve the best performance. in the sky could also interrupt data communications with
2) Solar and Geomagnetic Storms: Starlink’s throughput satellites. It is known that even a light cloud can affect the
can be heavily affected by solar and geomagnetic storms. A satellite signal strength by around 10% [15]. Thick clouds
solar storm, or Coronal Mass Ejection (CME), is when massive are usually coupled with heavy rainfall, further blocking the
magnetized particles from the Sun are ejected in a specific network paths in ULs and GSLs.
direction, e.g., towards the Earth, which can interact with We have also observed a 5-26% drop in throughput when
the earth’s magnetic field [23].13 We have found that, during the temperatures are above 12◦ C. The Starlink dish does not
February 3rd and 4th, 2022, the dish’s throughput dropped dra- have an active cooling mechanism, only passively cooling
matically from 100 Mb/s to 5 Mb/s. This coincides with two through its aluminum back-plate,16 which can be a significant
events: A SpaceX mission that launched 49 more satellites to challenge when used in the summer or tropics.
space on Feb. 3rd and a geo/electro-magnetic storm on Feb. 4th Using an Emporia smart plug, we find that a current dish’s
that forced 40 satellites to re-enter the Earth’s atmosphere.14 power consumption averages around 56.3 Watts but can go as
The storm may have strongly interfered with the UL/GSL high as 144.5 Watts. This is considerably lower than an earlier
similar to how GPS radio waves were affected by the change reported measurement of 105 Watts on average and 190 Watts
in atmosphere due to a previous geomagnetic storm [35]. maximum [26]. Such power saving is likely done through
Another reason could be that the satellite constellation may firmware optimization as the dishes’ form factors remain the
have been put in a less serviceable maintenance mode, leading same, and it is certainly helpful for heat control. As an extra
to poor communication performance. Interestingly, uploads do experiment during a snow storm in Burnaby at night from
not seem to be affected as the throughput stayed relatively Dec. 19 to 20, 2022, Starlink enters a snow melting mode for
stable. This implies that the two directions are not symmetric 10 hours, increasing its power consumption to an average of
when facing such storms, which can be an important factor to 146.3 Watts with a maximum of 188.6 Watts. If Starlink is
consider when integrating Starlink and the terrestrial network used in cold and snowy climates, the user must be careful as
infrastructures. the dish requires significantly more energy to melt the snow
3) Precipitation, Temperature, and Dish Power: We have that could accumulate. This feature however can be turned
observed that the Starlink’s throughput can be significantly off in the Starlink app if the user has alternative methods of
affected by the weather. The results imply inverse correlations keeping the Starlink clear from snow.
for throughput in comparison to precipitation and temperature. Since the dish kit is a blackbox, we cannot directly measure
Throughput drops on average 27% during any precipitation the power consumption of individual modules. However, we
(see details in Fig. 11). We have also found that the throughput compared the power consumption against throughput and find
is capped with heavy rain (> 4 mm per hour).15 Downloads that there is no obvious correlation between them, which
are affected more than uploads with UDP download having contrasts with the intuition that higher throughput requires
the maximum inverse correlation of 0.34. For instance, UDP more power. In fact, in some cases, we have seen that they
are inversely correlated for both upload and download. These
12 Though the Starlink App conceals obstruction status in its debug mode
suggest that the dish does not necessarily amplify the signal
with no further details, certain information can be extracted from the debug
log through a third-party tool Obstruction Viewer (http://starlink.dsmfaq.com/ further for higher throughput and that the dish may throttle
tools/obstructions.html). the maximum achievable throughput to control the power
13 https://www.swpc.noaa.gov/phenomena/geomagnetic-storms
14 https://www.spacex.com/updates/#sl-geostorm 16 https://www.vice.com/en/article/7kvkga/spacexs-satellite-internet-shuts-
15 https://water.usgs.gov/edu/activity-howmuchrain-metric.html down-because-its-too-hot-in-arizona

7
 150

Power (Watts)
100

50

0
0.0 0.5 1.0 1.5 2.0 2.5
Precip. Amount (mm)

Fig. 12. Power consumption versus hourly precipitation.

consumption. Consider the summary in TABLE II, the highest

(a) Dish F facing towards a clear (b) Google satellite map.
average TCP throughput in our experiments is around 108 sky.
Mb/s for download and 6 Mb/s for upload; yet a previous
experiment in Austria has observed much higher throughput Fig. 13. (a) Dish F (Gen 2) setup at (b) Koeye point (Estuary of Koeye River).
for both directions: 175 Mb/s for download and 17 Mb/s for
upload [26]. This could be due to obstructions and coverage
differences; however, the Starlink software update to lower
power consumption may have also affected the throughput.17
An updated future measurement in the same environment
would be required to test this conjecture.
On the other hand, as shown in Fig. 12, the dish’s power
consumption is correlated with precipitation. Without rain,
power consumption is lower in general, with occasional spikes
due to other factors such as tuning the dish direction or
establishing a UL with farther away satellites. When there is
rain or heavy clouds, power consumption becomes persistently
higher, likely in an attempt to deal with the interference from
the rain or clouds.
(a) The surrounding terrain map. (b) High elevation mountains.
V. S TARLINK IN THE W ILD
Fig. 14. The terrain of the valley in Manning Park. (a) Terrain map; (b)
Serving remote areas has always been a key target of surrounding mountains.
satellite communication systems, especially for Starlink. We
have performed a series of measurements in these challenging
environments with no terrestrial Internet coverage, not even about 68% lower than that in the urban city and the latency
cellular or power services. We now present the results for two is about 11% to 30% longer, with interruptions from time to
representatives: an estuary in the far north, and a deep valley time. This is likely due to there being fewer satellites and
surrounded by steep mountains. GSes being able to cover the northern areas (Shown in Fig. 7)
and the inclination constraint of the first shell. As such, more
A. Northern Shoreline efforts are required to find the next satellite to handover and
From Vancouver to the estuary (Koeye Point) as seen in set up the GSL and UL.
Fig. 13(b) involves a 2-hour flight to Bella Bella in the north
and then a 2-hour motor boat ride along the Pacific coastline. B. High-Elevation Deep Valley
Our trip is for a collaborative project with biologists and the Our second test is at the center of Manning Park, BC,
Heiltsuk native people for salmon and forest conservation. which has a rough terrain at around 1,000 meters above the
Building a communication service plays an important role in sea surrounded by mountains over 2,000 meters as shown in
the project. Our test is based on dish F shown in Fig. 13(a), Fig. 14. It is a 3-hour driving distance from Vancouver and has
a Gen 2 that is easier to carry over on airplanes. basically no cellular network coverage, not to mention Internet
The simple lodge we were living at the Koeye Point was access. We carried over dish C in our car to the site with the
built by the Heiltsuk people. Near the Pacific shoreline, it has portability option turned on, allowing the dish to be used out
a nice sky clear of obstructions. There is low light pollution, of its original registered location.
allowing clear sight of satellites (100%) and space stations Despite this valley not yet being officially serviced by Star-
orbiting above. Unfortunately, even though this location is link till the end of 2023 as shown in Fig. 15,18 we were able to
recently covered by Starlink with Dish F being usable for connect to satellites without denials, albeit with poor services.
simple day-to-day applications, its download throughput is Latency ranged from 90-350 ms on average across the different
17 https://www.tesmanian.com/blogs/tesmanian-blog/ukraine-starlink 18 https://www.starlink.com/map

8
 TABLE III. Outage counts for 2-dish synchronous streaming.

Buffer App Reported Overlapped Overlapped

Outage Outage Outage Outage Ratio
Dish A 8 8 3 37.50%
Dish B 19 13 3 15.79%

Fig. 15. Starlink coverage map around Manning Park.

40 Fig. 17. A Gen-1 dish secured on the roof of a minivan.

Buffer Health (s)

30 Source
Dish A
Dish B place two dishes (A and B) in nearby locations within the cov-
20
erage of the same satellite. In other words, they share the same
01:00:00 02:00:00 03:00:00 04:00:00 05:00:00 06:00:00
Timestamp GSL but have their respective ULs. We then perform YouTube
streaming synchronously to them, each being attached to a PC
Fig. 16. Buffer health over time for 2-dish synchronous streaming. Each bar to playback the video. The synchronous streaming sessions last
outlines the standard deviation for 7-min samples. Outages are highlighted by
vertical dotted lines. over 24 hours by loop playing back the video, and the PCs’
browser buffers were cleaned up after each playback to ensure
continuous data downloading. We use a script to automatically
regions with huge fluctuations over 1,000 ms. Throughput was scrape YouTube’s built-in tool Stats for nerds for real-
on average 13 Mb/s and 4 Mb/s with rare bursts to 100 Mb/s time data statistics [26], including connection speed, network
and 20 Mb/s for download and upload, respectively. Frequent activity, and buffer health. The buffer health here represents
outages happened every 1-3 minutes, even though the App the length (in seconds) of pre-loaded video content, which,
reports only a 2% obstruction ratio. Given the terrain, we being sensitive to network outages, is directly correlated with
believe that any LEO satellite, if not being directly above, the viewer’s perceived quality.
would be obstructed by the mountains. The clear short-distance As can be seen from Fig. 16, both dishes have experienced
visibility to the sky within the valley would mislead the App’s a series of buffer outages in this synchronous test, some of
algorithm in calculating the obstruction. As a matter of fact, which are overlapped. For instance, two pairs of buffer outages
dish C used to be placed within a balcony in the urban city, occurred at 01:18 and 03:26 simultaneously for both dishes,
suffering from a 24.9% obstruction ratio, but its end user and their Starlink Apps both reported network outages at the
experience there was nearly as good as dishes A and B. same time. The overlapped outages indicate that there can be
This suggests that location (and terrain in particular) plays a common factor, the bent-pipe, causing them between the
an important role and not all obstructions are equal to the two dishes. Table. III summarizes the outage events, from
end user experience. With more satellites being deployed and which we can confirm that a non-negligible amount of outage
the valley becoming officially serviceable, the service quality events are caused by their bent-pipes (over 37.5% for A and
would be improved, if not to the level of flat open areas. 15.79% for B). The data also show that Dish A is more stable
It is also worth noting that both remote locations have not in streaming than B, which is likely because it has clearer
been connected to the power grid, nor would be in the near visibility. There are more trees around dish B, so the Starlink
future. Hence we have to rely on a solar-diesel hybrid power App has reported a 2.7% and a 4.7% obstruction ratio for Dish
system at the estuary and a battery power pack in the valley. A and B, respectively. Hence, even in the same service area,
Given the current Starlink kit’s power consumption, a typical different Starlink users may experience different UL quality
1,500 Ah power pack lasts around only 1 hour. and hence network service quality.
VI. F URTHER D ISCUSSIONS
B. Mobility Potentials
A. Stability of Bent-Pipe
Our early experiments have suggested that, despite being FCC has recently authorized Starlink Internet for use on
reasonably good in throughput, the bent-pipe is often unstable vehicles in motion,20 though it has yet to be fully supported
in communication. To closely examine its stability, we have by Starlink. Hence, we have run a short test at this stage to
done a stress test under continuous streaming of an ultra-high- gauge the work needed to get Starlink ready for full in-motion
resolution 8K (7,680 x 4,320 pixels) YouTube video.19 We
20 https://www.cnbc.com/2022/06/30/fcc-approves-spacex-starlink-service-
19 https://www.youtube.com/watch?v=KcMlPl9jArM to-vehicles-boats-planes.html

9
 2
4 × 102
3 × 10
2
Latency (ms) 2 × 10
2
10
1
6 × 10

20:44 20:50 20:56 21:02 21:08

Fig. 18. Latency to N. California while in motion (vertical lines correspond
to 95% confidence intervals).
(a) Aerial view with 30m tall trees

Download Upload
2
10
Throughput (Mb/s)

1
10

0
10

20:44 20:50 20:56 21:02

Fig. 19. Throughput to N. California while in motion (vertical lines correspond (b) Weir setup and solar panels
to 95% confidence intervals).
Fig. 20. An aerial view of the weir site upstream of the Koeye River in the
Great Bear rainforest.

connectivity.21 We secured dish C with two crossbars on the

roof-rack of a mini van (see Fig. 17), with the dish’s portability
that of the stationary setup. Both observe a noticeable drop at
option being turned on. In our 30-minute test, the car went
around 20:58, likely due to a disconnection from a satellite.
at 40 to 70 km/h around three different driving routes with
The last 6 minutes have been mostly a straight drive in a valley,
twists and turns and then a full stop near the end. Most of
suggesting that perhaps the dish was forced to switch satellites
the time, the dish was flat towards the sky with some minor
faster than expected.
self-adjustment. Note that the dish in the stationary setup is
Note that the Starlink kit typically takes 3-7 minutes to boot
mostly tilted towards one direction. Hence, likely the dish can
up and be Internet ready, but we have observed cases over 20
sense motion and accordingly adjust the flat position, which
minutes. Such a long time will be a barrier for mobile use.
works best for omnidirectional reception.
The current Starlink dish works while in motion but still C. Global Coverage? Technical and Cultural Implications
has a long way to go to match the current stationary setup.
To achieve truly global coverage, more Starlink satellites
We have observed outages at the median every 16.5 seconds
are needed, including other shells covering the North and
for 5 seconds, and some every 162 seconds lasting as high as
South Poles. Our remote area experiments were done 14°
36 seconds (95-th percentile). Latency as shown in Fig. 18
south of the Arctic Circle (66°34’N), which is already almost
is on average 100 ms with the highest being 2,800 ms
out of the current service area in North America. Inter-
at 20:42. The latency spikes greater than 200 ms at least
satellite links are necessary to be enabled, too, for efficient
twice every minute on average with as high as 8 times in
cross-continent communications and for robust operation when
a minute. This would frequently interrupt applications that
disasters happen near the Earth’s surface.
require sub-200 ms for satisfying experiences. Considering
Even if these technical issues would eventually be over-
that a stationary dish averages around 50 ms as we have
come, we have realized many challenges during our site visit
discussed earlier, the latency while mobile is high with larger
deep into the Great Bear temperate rainforest at the estuary’s
fluctuations. Some of these latency increases could be due to
upstream, where we planned to install a Starlink dish near the
sudden orientation changes as the car turns around corners or
Heiltsuk aboriginal people’s reconstructed weir for a fish trap
intermittent obstructions from trees passing by. For example, a
and its monitoring system (Shown in Fig. 20(b)).
right-hand turn occurred at around 20:40, which coincides with
Practical challenges include clearance, wildlife, and power.
the sharp spike in latency at around the same time. Download
The old-growth trees are 30 meters tall as can be seen in
throughput has reached similar averages to the stationary test
Fig. 20(a), and the only clear sky available is at the weir on
at around 80 to 100 Mb/s but immediately drops off near the
the riverside, which is subject to flooding. Wildlife such as
end, as shown in Fig. 19. Upload throughput, interestingly,
bears, wolves, and birds could easily cross the narrow weir
has been able to keep up around 5 to 10 Mb/s, on par with
and disrupt the dish. As a matter of fact, one hundred distinct
21 This test was for academic research only. Currently, Starlink does not
individual grizzly bears have been identified in the Koeye
suggest general use of its dishes in motion, although there have been trials, watershed. Also, solar is the only source of power available
notably in Ukraine, and Starlink also plans to release an RV version. in such a remote area in the deep forest (The diesel generator

10
is not allowed nor practical). As such, intermittent operations First Nation, for their great support. We also thank the QQS
are inevitable as the dish needs a max of 145 watts of power (EYES) Projects Society for hosting us at the Koeye Lodge
if it does not enter snow melting mode. (in particular, the lodgekeepers Ian and Emily Files) and for
Cultural challenges are more complicated, which we have their effort in protecting the natural environment of the Great
not considered before the site visit. The Koeye river and its Bear Rainforest and the Heiltsuk heritage.
surrounded temperate rainforest is a holy land of the Heiltsuk
native people, hosting their traditional big house. The people R EFERENCES
there bring their set of conflicts with their traditional values. [1] X. Lin, S. Cioni, G. Charbit, N. Chuberre, S. Hellsten, and J.-F.
Signal pollution includes more philosophical thought of pris- Boutillon, “On the Path to 6G: Embracing the Next Wave of Low
tine wilderness and as more of the world is connected, fewer Earth Orbit Satellite Access,” IEEE Communications Magazine, vol. 59,
no. 12, pp. 36–42, 2021.
places will exist as a getaway from constant connectivity. [2] S. Kassing, D. Bhattacherjee, A. B. Águas, J. E. Saethre, and A. Singla,
Light pollution is commonly known nowadays, and a global “Exploring the ”Internet from space” with Hypatia,” in Proc. ACM IMC,
coverage brought by the LEO satellite constellation would 2020.
[3] Z. Lai, H. Li, and J. Li, “StarPerf: Characterizing Network Performance
bring another pollution of invisible electromagnetic waves, the for Emerging Mega-Constellations,” in Proc. IEEE ICNP, 2020.
WiFi pollution. This is not in terms of the technological aspect [4] G. Giuliari, T. Ciussani, A. Perrig, and A. Singla, “ICARUS: Attacking
but in the social aspect. This has been the major concern of low Earth orbit satellite networks,” in Proc. USENIX ATC, 2021.
[5] Y. Hauri, D. Bhattacherjee, M. Grossmann, and A. Singla, “”Internet
the biologists working with us. As a matter of fact, in the true from Space” without Inter-satellite Links,” in Proc. ACM HotNets, 2020.
wild, individuals will have to always stay alert and mostly take [6] D. Fischer, D. Basin, K. Eckstein, and T. Engel, “Predictable Mobile
care of themselves; hence, the distraction of using mobile apps Routing for Spacecraft Networks,” IEEE Transactions on Mobile Com-
puting, vol. 12, no. 6, pp. 1174–1187, 2013.
would overshadow the benefit of staying connected. Policies,
[7] M. Handley, “Delay is Not an Option: Low Latency Routing in Space,”
professional guidelines, and ethics across different disciplines in Proc. ACM HotNets, 2018.
and cultures are therefore urgently needed for the forthcoming [8] D. Bhattacherjee, W. Aqeel, I. N. Bozkurt, A. Aguirre, B. Chan-
global Internet coverage on Earth and in space. drasekaran, P. B. Godfrey, G. Laughlin, B. Maggs, and A. Singla,
“Gearing up for the 21st century space race,” in Proc. ACM HotNets,
2018.
VII. C ONCLUSION AND F UTURE W ORKS [9] G. Giuliari, T. Klenze, M. Legner, D. Basin, A. Perrig, and A. Singla,
This paper has presented our initial measurements on the “Internet backbones in space,” ACM SIGCOMM Computer Communi-
cation Review, vol. 50, no. 1, pp. 25–37, 2020.
network characteristics of today’s Starlink from an end-user’s [10] L. Chen, F. Tang, Z. Li, L. T. Yang, J. Yu, and B. Yao, “Time-Varying
perspective. Many of the issues identified from our study Resource Graph Based Resource Model for Space-Terrestrial Integrated
are not just confined to Starlink, but are common for LSNs. Networks,” in Proc. IEEE INFOCOM, 2021.
[11] B. Denby and B. Lucia, “Orbital Edge Computing: Nanosatellite Con-
Our observations therefore will help with optimizing the LSN stellations as a New Class of Computer System,” in Proc. ACM ASPLOS,
deployment and operation, and will also help developers and 2020.
users customize their networks and applications. [12] D. Bhattacherjee and A. Singla, “Network topology design at 27,000
km/hour,” in Proc. ACM CoNEXT, 2019.
LSN services are rapidly evolving, so we will be con- [13] Z. Lai, W. Liu, Q. Wu, H. Li, J. Xu, and J. Wu, “SpaceRTC: Unleashing
tinuing to monitor the performance evolution of Starlink’s the Low-latency Potential of Mega-constellations for Real-Time Com-
Internet service. For instance, the weather or cloud effects munications,” in Proc. IEEE INFOCOM, 2022.
[14] M. Handley, “Using ground relays for low-latency wide-area routing in
on Starlink can be further expanded with rain probability megaconstellations,” in Proc. ACM HotNets, 2019.
[38] and cloud attenuation [39] models in future experiments. [15] D. Vasisht, J. Shenoy, and R. Chandra, “L2D2: low latency distributed
They can include optimization methods that take into account downlink for LEO satellites,” in Proc. ACM SIGCOMM, 2021.
[16] V. Singh, A. Prabhakara, D. Zhang, O. Yağan, and S. Kumar, “A
the probability of weather attenuation and act accordingly to community-driven approach to democratize access to satellite ground
provide uninterruptible service. For the bent-pipe’s stability, stations,” in Proc. ACM MobiCom, 2021.
a future repeat experiment could be conducted by mounting [17] T. Henderson and R. Katz, “Transport protocols for Internet-compatible
satellite networks,” IEEE Journal on Selected Areas in Communications,
the dishes to the roof of each respective house to mitigate the vol. 17, no. 2, pp. 326–344, 1999.
obstruction ratio variable. [18] C. Caini, R. Firrincieli, D. Lacamera, T. de Cola, M. Marchese, C. Mar-
We are also interested in the effectiveness of the upcoming condes, M. Sanadidi, and M. Gerla, “Analysis of TCP live experiments
on a real GEO satellite testbed,” Performance Evaluation, vol. 66, no. 6,
new functions, e.g., satellite-to-satellite links, multiple bent- pp. 287–300, 2009.
pipe, and mobility support, to name but a few. We are also [19] Z. Lai, Q. Wu, H. Li, M. Lv, and J. Wu, “OrbitCast: Exploiting
interested in the seamless integration of LSN and terrestrial Mega-Constellations for Low-Latency Earth Observation,” in Proc. IEEE
ICNP, 2021.
Internet and 5G cellular networks towards truly global cover-
[20] P. Wang, H. Li, B. Chen, and S. Zhang, “Enhancing Earth Observation
age and the related inter-disciplinary issues. Throughput Using Inter-satellite Communication,” IEEE Transactions
on Wireless Communications, pp. 1–1, 2022.
ACKNOWLEDGEMENT [21] Y. Zhang, Q. Wu, Z. Lai, and H. Li, “Enabling Low-latency-capable
Satellite-Ground Topology for Emerging LEO Satellite Networks,” in
This project was supported by a Canada NSERC Discovery Proc. IEEE INFOCOM, 2022.
Grant and a British Columbia Salmon Recovery and Innova- [22] Y. Li, H. Li, L. Liu, W. Liu, J. Liu, J. Wu, Q. Wu, J. Liu, and
tion Fund (No. 2019-045). The authors thank Dr. Will Atlas Z. Lai, “”Internet in Space” for Terrestrial Users via Cyber-Physical
Convergence,” in Proc. ACM HotNets, 2021.
from the Pacific Salmon Foundation and Ian Clevenger from [23] S. A. Jyothi, “Solar superstorms: planning for an internet apocalypse,”
the Salmon Watersheds Lab at SFU, as well as the Heiltsuk in Proc. ACM SIGCOMM, 2021.

11
[24] J. Tregloan-Reed, A. Otarola, E. Ortiz, V. Molina, J. Anais, R. González, [32] A. Garreffa, “SpaceX to enable laser satellites for Starlink
J. Colque, and E. Unda-Sanzana, “First observations and magnitude internet connectivity | TweakTown,” Jun. 2022. [Online].
measurement of Starlink’s Darksat,” Astronomy & Astrophysics, vol. Available: https://www.tweaktown.com/news/86807/spacex-to-enable-
637, p. L1, 2020. laser-satellites-for-starlink-internet-connectivity/index.html
[25] J. Tregloan-Reed, A. Otarola, E. Unda-Sanzana, B. Haeussler, F. Gaete, [33] D. Xu, A. Zhou, X. Zhang, G. Wang, X. Liu, C. An, Y. Shi, L. Liu, and
J. Colque, C. González-Fernández, J. Anais, V. Molina, R. González H. Ma, “Understanding Operational 5G: A First Measurement Study on
et al., “Optical-to-NIR magnitude measurements of the Starlink LEO Its Coverage, Performance and Energy Consumption,” in Proc. ACM
Darksat satellite and effectiveness of the darkening treatment,” Astron- SIGCOMM, 2020.
omy & Astrophysics, vol. 647, p. A54, 2021. [34] X. Shao, Z. Sun, M. Yang, S. Gu, and Q. Guo, “NOMA-Based Irregular
[26] C. Uran, K. Horvath, and H. Wöllik, “Analysis of a Repetition Slotted ALOHA for Satellite Networks,” IEEE Communica-
Starlink-based Internet connection – ROADMAP-5G,” Jul. tions Letters, vol. 23, no. 4, pp. 624–627, Apr. 2019.
2021. [Online]. Available: https://forschung.fh-kaernten.at/roadmap- [35] N. Linty, A. Minetto, F. Dovis, and L. Spogli, “Effects
5g/en/2021/07/24/analysis-of-a-starlink-based-internet-connection/ of Phase Scintillation on the GNSS Positioning Error
[27] Khalife, Joe and Neinavaie, Mohammad and Kassas, Zaher M., “The first During the September 2017 Storm at Svalbard,” Space
carrier phase tracking and positioning results with starlink leo satellite Weather, vol. 16, no. 9, pp. 1317–1329, 2018, eprint:
signals,” IEEE Transactions on Aerospace and Electronic Systems, https://onlinelibrary.wiley.com/doi/pdf/10.1029/2018SW001940. [On-
vol. 58, no. 2, pp. 1487–1491, 2022. line]. Available: https://onlinelibrary.wiley.com/doi/abs/10.1029/
[28] M. Neinavaie, J. Khalife, and Z. M. Kassas, “Acquisition, Doppler 2018SW001940
Tracking, and Positioning With Starlink LEO Satellites: First Results,” [36] A. D. Panagopoulos, P.-D. M. Arapoglou, and P. G. Cottis, “Satellite
IEEE Transactions on Aerospace and Electronic Systems, vol. 58, no. 3, communications at KU, KA, and V bands: Propagation impairments
pp. 2606–2610, 2022. and mitigation techniques,” IEEE Communications Surveys & Tutorials,
[29] F. Michel, M. Trevisan, D. Giordano, and O. Bonaventure, “A first vol. 6, no. 3, pp. 2–14, 2004, conference Name: IEEE Communications
look at starlink performance,” in Proceedings of the 22nd ACM Surveys & Tutorials.
Internet Measurement Conference, ser. IMC ’22. New York, NY, [37] C. I. Kourogiorgas and A. D. Panagopoulos, “Space-Time Stochastic
USA: Association for Computing Machinery, Oct. 2022, pp. 130–136. Rain Fading Channel for Multiple LEO or MEO Satellite Slant Paths,”
[Online]. Available: http://doi.org/10.1145/3517745.3561416 IEEE Wireless Communications Letters, vol. 7, no. 3, pp. 284–287, Jun.
[30] M. M. Kassem, A. Raman, D. Perino, and N. Sastry, “A browser-side 2018, conference Name: IEEE Wireless Communications Letters.
view of starlink connectivity,” in Proceedings of the 22nd ACM [38] “RECOMMENDATION ITU-R P.837-6 - Characteristics of precipita-
Internet Measurement Conference, ser. IMC ’22. New York, NY, tion for propagation modelling.”
USA: Association for Computing Machinery, Oct. 2022, pp. 151–158. [39] F. Yuan, Y. H. Lee, Y. S. Meng, S. Manandhar, and J. T. Ong,
[Online]. Available: http://doi.org/10.1145/3517745.3561457 “High-Resolution ITU-R Cloud Attenuation Model for Satellite Com-
[31] Space Exploration Technologies, “SpaceX non-geostationary satellite munications in Tropical Region,” IEEE Transactions on Antennas and
system Attachment A: technical information to supplement Schedule Propagation, vol. 67, no. 9, pp. 6115–6122, Sep. 2019, conference
S.” [Online]. Available: https://licensing.fcc.gov/myibfs/download.do? Name: IEEE Transactions on Antennas and Propagation.
attachment key=1158350

12