Extended summary of "Cloudy with a chance of short RTTs Analyzing Cloud Connectivity in the Internet"

UNIVERSITA’ DEGLI STUDI DI TRIESTE
Dipartimento di Ingegneria e Architettura
Corso di Laurea in
Ingegneria Elettronica e Informatica
Extended summary of
“Cloudy with a Chance of Short RTTs
Analyzing Cloud Connectivity in the Internet ”
Candidata Relatore
Isabella Filippo Prof. Alberto Bartoli
Anno accademico 2022/2023
1

Contents
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
End-points selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
Vantage point selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
Cloud access Latency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
Influence of wireless last-mile . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
Cloud & ISP interconnections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
Riassunto in italiano . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2

Introduction
In today’s digital era, cloud computing revolutionizes data, application access and storage, by
utilizing internet-based services instead of local servers or personal computers. This technol-
ogy has become crucial for the ever-increasing growth of next-generation applications such
as AR/VR, autonomous vehicles, remote surgery and gaming. It also became essential for
user traffic, as most of the population shifted to mixed working methods (remote/presence)
since 2020. This paper [1] aims to investigate the state of end-user cloud connectivity through
widespread measurements and to provide insight into last-mile latency and edge computing.
The main aspects of this article include:
• A large-scale study of cloud computing involving nine major cloud providers;
• A comparison between Speedchecker measurements and previous experiments con-
ducted using over 8,000 RIPE Atlas probes, all targeting the same cloud region;
• An analysis of the impact of wireless last-mile connections;
• An identification of the various types of interconnections between ISPs (Internet Service
Providers) and cloud providers from the end-user’s perspective.
Background
In recent years, several studies focused on internet topology, which laid the foundation for in-
depth research on the spread of cloud infrastructures. Only a few studies assessed the latency
of global cloud access. Previously, the study conducted by Corneo et al.[2] had already used
probes of the RIPE Atlas platform but, due to the placement of the observation points within
managed infrastructures, such as network service providers and educational institutions, it
does not accurately represent the connectivity of real internet users on a global scale. In this
investigation, Speedchecker probes have been used to analyze the reachability and impact
of cloud expansion for internet users worldwide. They will also be employed to understand
whether the current infrastructure can meet the latency requirements of mission-critical ap-
plications, using the following Quality of Experience directives (Figure 1):
• Motion-to-Photon (MTP): the delay between user input and its display reflection, ap-
proximately 20 ms, for AR and VR applications to avoid motion dizziness;
• Human Perceivable Latency (HPL): the threshold when a user starts to experience lags,
around 100 ms, considered for cloud gaming;
• Human Reaction Time (HRT): the delay difference between a visual stimulus and the
motion response, estimated to be 250 ms, for remote surgery applications.
3

Figure 1. QoE requirements of next-generation applications.
Methods
End-points selection
The researchers selected 195 cloud regions, offered by nine cloud providers (Table 1) with
worldwide coverage and private Wide Area Networks (WANs) as end-points. Some providers
like Amazon, Google, and Microsoft have established extensive private WANs to protect users’
traffic from the public internet. Additionally, the authors incorporated data centers operated
by Alibaba Cloud, due to their extensive presence in Asia. The distribution of end-points is
represented in Figure 2.
Table 1. Density of cloud provider endpoints, and their type of backbone infrastructure.
4

Figure 2. Distribution of datacenters.
Vantage point selection
The primary data source for this study is the Speedchecker platform, which has many probes
deployed in over 170 countries. The researchers observed that Speedchecker probes (Figure 3)
showed wider deployment and greater geographical coverage than RIPE Atlas probes (Figure
4). Furthermore, Speedchecker probes were more frequently based on end-user devices. The
authors concluded that Speedchecker is a better platform for assessing global user accessibil-
ity to cloud services.
Figure 3. Distribution of 115,000+ Speedchecker Android probes.
Figure 4. Distribution of 8500+ RIPE Atlas probes.
5

Experiments
In this paper [1] the authors analyzed two main aspects of cloud connectivity: the state of
user latency to current cloud deployment and the impact of cloud provider investments in
shortening user paths to their infrastructure on end-user connectivity. They ran TCP pings
and ICMP traceroutes from Speedchecker VPs (Vantage Points) to cloud region endpoints.
Both experiments were conducted in parallel for six months, from October 2020 to April 2021.
The research team faced some challenges while using the Speedchecker platform, including:
• The majority of Android probes on the platform were temporarily active across days.
• The research team had access to the platform with a limited measurement budget that
refreshed at the end of each day.
To address these challenges, the research team employed the following strategies:
• They used a few API calls to collect information about connected vantage points, which
they triggered every four hours. This strategy allowed the team to track related probes
on the platform worldwide.
• They set up their active network experiments to cover every country in each continent,
deploying a minimum of 100 probes and focusing on all cloud regions within the respec-
tive continent.
• They imposed a rate limit of one measurement request per minute to avoid overloading
the platform.
The researchers collected over 3.8 million ping data points and more than 7 million unique
traceroutes during their study.
Results
Cloud access Latency
The study found that:
• Geographical location has a significant impact on cloud connectivity latency. Countries
with in-land data centers had the best median latency (Figure 5). For example, Africa
exhibited the poorest performance, with less than ten percent of latency samples below
the HPL (Human Perceivable Latency) threshold.
• Speedchecker and RIPE Atlas probes have different deployment locations and connec-
tivity types, which can lead to differences in measured latency. RIPE Atlas probes are
generally located closer to data centers and have wired access, while Speedchecker probes
are deployed on end-user mobile devices with wireless last-mile access.
• Cloud connectivity latency can be improved in continents with limited data center de-
ployment by connecting to data centers in neighboring better-provisioned continents.
6

Figure 5. Average latency closest to datacenter.
Influence of wireless last-mile
The authors employed Speedchecker measurements to analyze the influence of wireless last-
mile between home (wireless connectivity) and cellular connection. In countries with well-
provisioned cloud infrastructures such as Europe and North America, the overall latency to
reach data centers is low so, the latency due to the last-mile is evident. On the other hand, in
Africa and Asia, the percentage of latency due to the last-mile is smaller because the deploy-
ment of data centers in these regions is sparse (Figure 6).
Figure 6. Percent of wireless last-mile latency to total cloud access for Speedchecker and RIPE
Atlas. SC home (USR-ISP) is wireless connection, SC home (RTR-ISP) is wired connection.
Figure (7) represents the last-mile absolute latency for home and cellular connection. It shows
that the nature of the last-mile does not affect the absolute latency over the globe. The median
value of the last-mile latency is 20-25 ms for both types of connections. The authors observed
that latency linked to the last-mile is lower for RIPE Atlas probes than Speedchecker. This ob-
servation highlights the wired nature of Atlas VP (Vantage Points) connections like the Speed-
checker measurements on the wired connection from home to ISP (Internet Service Provider).
7

Figure 7. Total of wireless last-mile latency for Speedchecker and RIPE Atlas probes.
Cloud & ISP interconnections
The researchers focused on identifying interconnections between VP ISPs (Vantage Point In-
ternet Service Providers) and cloud providers (Figure 8). Unresponsive IP addresses are re-
moved, IXPs (Internet Exchange Points) are identified and tagged in the path, and removed
from the AS-level topology as they only act as points of traffic exchange. The study classifies
paths into the following categories (Table 1):
• Direct Peering: The cloud provider and ISP AS are directly connected neighbors.
• Private Peering: An intermediate AS acts as a step between the cloud and VP ISP.
• Public Internet: Paths involve more than one step AS.
Figure 8. Different ISP-cloud interconnections.
The team also analyzed the pervasiveness of cloud providers’ ownership of routers along the
path (Figure 9): Cloud providers like Amazon, Google, and Microsoft own a significant portion
of the path in almost every continent, while providers with multiple ASes own fewer routers
on a path.
8

Figure 9. Pervasiveness of different cloud providers globally.
Discussion
Different factors influence cloud computing performances. Firstly, there may be limitations in
identifying peering relationships, such as IXP (Internet eXchange point) hops may not always
appear in traceroutes, and routes that traverse IXPs may be misclassified. Secondly, the article
[1] doesn’t discuss the access type of last-mile, such as Wi-Fi or cellular. Since the nature of
the last-mile represents the primary bottleneck for users’ traffic.
Recently, interest has grown in edge computing, in which the data elaboration phase is closer
to the users. The belief is that this approach could help to reach the low latency requirements
for next-generation applications. In regions with a dense distribution of cloud data centers,
latency is typically stable on last-mile connectivity; in these regions, the deployment of edge
computing won’t change the medium latency unless it becomes widespread. However, in de-
veloping regions with poor connectivity to cloud data centers, edge computing could help to
reach significant improvements.
The cloud can meet the requirements of most applications. MTP (Motion-To-Photon) appli-
cations that require low latency or high transfer speeds may be impractical in regions with
poor connectivity also, in developed areas as the absolute latency measurement is about 20+
ms. Peering agreements between operators can help reduce latency variation, but this does
not significantly affect base latency.
Conclusions
In the last decade, cloud service providers have made huge investments in expanding their
global network including the construction of new data centers and the expansion of private
networks to get closer to users. These advancements have led to better cloud performance
across continents with developed economies, thanks to a well-distributed presence of data
centers. In developing regions, however, users experience suboptimal cloud latencies due to
geographical distance. Investments in private networks and direct links with ISPs (Internet
Service Providers) are more evident in these regions, enabling sometimes better performance.
9

In conclusion, the last-mile wireless connection remains a significant challenge for cloud ac-
cess over the globe.
References
[1] The Khang Dang, Nitinder Mohan, Lorenzo Corneo, Aleksandr Zavodovski, Jörg Ott, Jussi
Kangasharju. Cloudy with a Chance of Short RTTs Analyzing Cloud Connectivity in the Inter-
net. November 2–4, 2021, Virtual Event, USA
[2] Lorenzo Corneo, Maximilian Eder, Nitinder Mohan, Aleksandr Zavodovski, Suzan Bayhan,
Walter Wong, Per Gunningberg, Jussi Kangasharju, and Jörg Ott. 2021. Surrounded by the
Clouds. In Proceedings of The Web Conference 2021 (WWW ’21). Association for Computing
Machinery, NewYork, NY, USA. https://doi.org/10.1145/3442381.3449854
Riassunto in italiano
Nell’attuale era digitale, il cloud computing ha rivoluzionato l’accesso e l’archiviazione dei
dati, sostituendo i server locali e i computer personali con servizi basati su Internet. Questa
tecnologia è cruciale per la crescita delle applicazioni di prossima generazione, specialmente
dal 2020 con il diffondersi dello smart working. Il documento [1] esplora la connettività cloud
dell’utente finale, concentrandosi sulla latenza dell’ultimo miglio e in particolare sui seguenti
aspetti chiave: uno studio su larga scala che coinvolge i principali fornitori di cloud, un con-
fronto con un esperimento precedente, un’analisi delle connessioni dell’ultimo miglio e l’iden-
tificazione delle interconnessioni tra ISP e fornitori di cloud.
Nelle regioni con connettività limitata cresce l’interesse nell’edge computing, in quanto potreb-
be migliorare significativamente la latenza. Inoltre, sebbene il cloud soddisfi la maggior parte
dei requisiti delle applicazioni, quelle che richiedono bassa latenza o alte velocità di trasferi-
mento potrebbero incontrare delle difficoltà. I fornitori di servizi cloud hanno investito notevol-
mente nell’espansione della loro rete globale, migliorando le prestazioni del cloud nelle re-
gioni sviluppate. Tuttavia, nelle regioni in via di sviluppo persistono latenze subottimali, prin-
cipalmente a causa della distanza geografica. La connessione wireless dell’ultimo miglio ri-
mane una sfida significativa per l’accesso globale al cloud.
10

Extended summary of "Cloudy with a chance of short RTTs Analyzing Cloud Connectivity in the Internet"

More Related Content

Similar to Extended summary of "Cloudy with a chance of short RTTs Analyzing Cloud Connectivity in the Internet" (20)

Recently uploaded (20)

Extended summary of "Cloudy with a chance of short RTTs Analyzing Cloud Connectivity in the Internet"