Web Connectivity in Ghana: A survey
Kai Lüke
Institute of Computer Science
Department of Mathematics and Computer Science
Freie Universität Berlin
Berlin, Germany
kai.lueke@fu-berlin.de
ABSTRACT
Starting from the research“Dissecting web latency in Ghana”
of Zaki et al. the field of Internet connectivity in Ghana is
explored and the different aspects like latency, DNS resolu-
tion times, complexity of web sites, caching and prefetch-
ing, TLS/SSL, TCP and cellular network performance, role
of ISPs, Internet Exchange Points, are brought in relation.
Measurement approaches and tailored solutions for areas
with bad connectivity are listed. Possible ways of improve-
ment are discussed.
Categories and Subject Descriptors
C.2.6 [Internetworking]: Standards; C.4 [Performance
of Systems]: Reliability, availability, and serviceability
Keywords
Developing Countries, Internet, Latency, Performance, TCP,
Mobile, Cellular
1. INTRODUCTION
The Millennium Development Goals, which were supposed
to be reached at the end of this year, also include measure-
ments in target 8F for “mak[ing] available the benefits of new
technologies, especially information and communications”.
By these numbers almost 40% of the world population have
Internet access [26]. Despite the big gap between global south
and global north in a variety of aspects, the huge number of
6.9 billion mobile-cellular subscriptions could be utilised to
raise Internet connectivity quickly. As in Ghana the popu-
lation number was outnumbered [17], because people tend
to have more than one subscription, the distribution might
not be equal, but has led to 40.7% mobile Internet penetra-
tion in 2013 [11]. Concerning the ICT Development Index
Ghana ranks 6th among African countries [13] and has di-
rect connection to major undersea cables [30]. Recently LTE
is rolled out as 4G cellular standard and the licence requires
coverage of half the country within five years [25].
But the complexity of web site construction and data amount
per page load has grown faster than the Internet connectiv-
ity in developing countries, thus leading to slow page load-
ing times while bandwidth has increased in Ghana [30, 15].
These are not only frustration in usability but can also elim-
inate certain use cases. The following overview will show
that growing bandwidth alone can not solve the problem
due to the impact of latency and package loss on common
transport protocols and the nested dependencies of requests
during page load.
2. SPEED OF DATA TRANSMISSION
Round trip times to servers on the other half of the world can
not be faster than the theoretical limit which is the speed of
light through glass fiber. By example taking s = 12767 km
as distance from Accra to Seoul and maximum speed because
of the refractive index for glass is v =
c
1.5
200000
km
s
the
round trip takes always more than t =
2·s
v
128 ms.
The importance of low latency when interacting with web
sites is shown at the example of usage reduction for search
engines after introducing an artificial delay as shortly de-
scribed in the detailed survey of Briscoe et al. on reducing
latency [3].
The links between Europe and North America are quite close
to the theoretical limit and even almost half a billion dollar is
spend for few millisecond gains by establishing new undersea
cables [18]. Developing countries are not close to optimum
and so improvements are possible in many ways. But still
it must be questioned if an access outside the continent is
justified or whether the content and services can be located
at infrastructure nearby.
While throughput can be increased by adding more links,
the performance of TCP is affected by latency. Also the
recursive dependencies in requests while loading web pages
result, as described later, in connectedness of latency and
throughput.
3. METHODS OF DATA COLLECTION
Statistical data and specially detailed recordings can be taken
at various places. In 2006 the operator of a main Internet
cafe in Accra allowed analysis of logs containing 13.7 mil-
lion requests [9]. Also browser access to hundreds of web-
sites ranked popular by Alexa has been profiled at Accra,
Kumawu and Hohoe (all in the south) in 2012 and 2014 by
Zaki et al. [30] using a similar approach to [27]. These ap-
proaches help in tracing down how the slow loading times ex-
perienced by users are composed by the single parts. But as
they happen on application layer, the reasons for decreased
throughput and higher round trip times for packages can
not be examined. Regarding to P
¨
otsch the charts in [30] do
also only include web pages which were able to be received
fully which was not the case for all websites. However the
paper focuses rather on DNS latency and solution space in
the browser. Other works referenced in the chapter here on
TCP contain research for these incidents.
Others showed how measurement devices for broadband and
cellular connections can be deployed at diverse places in In-
dia and South Africa [14, 7]. The special benchmark routers
contribute to appropriate long term data and also eliminate
possible disturbance to accuracy when the line is shared
with other users or due to the local network setup (ibid.
and [22]). Also applications on smartphones gain interest
as they provide an easy way to reach data sources [7, 19].
All these works also note the faced challenges like unstable
power supply and absence of connectivity or suspiciousness
towards the smartphone application.
Measurement can help in empowering users which constantly
see a much slower connection speed than advertised by the
provider [7].
4. INTERNET SERVICE PROVIDERS AND
INTERNET EXCHANGE POINTS
The contrast between urban and rural areas regarding (fast)
internet access is not only present in developing countries,
but can therefore significantly determine whether a certain
limit is exceeded which makes usage of demanding services
not worthwhile any more. Particular for Ghana is also the
gradient from the south to the overall less prosperous north.
As urban inhabitants tend to be a better income source,
providers can be inveigled to privilege them in resource allo-
cation e.g. as it was observed by which gateways were given
to urban and rural cellular clients in India [14].
Studies where measurements were realised for more than one
provider indicate differences not only in bandwidth and high
load during popular usage hours but also latency, specially
last mile latency, which is the time to get to the first hop in
the provider’s IP infrastructure [22]. It might even be that
just network configurations bring betterment for cellular ac-
cess [14]. In addition to correlation of the previous findings
[7] also notes that the interconnection of ISPs in the region
including neighbour countries plays a huge role when routes
outside of the continent can be avoided.
Internet exchange points bring not only ISPs together but
also provide a place where content can be made available.
The GIX in Accra operates for a decade now and contributed
in reducing access times for those organisations who are in-
terconnected there. On average the traffic is around 500
MB/s. Specially the common content distribution networks
and popular web services need to take part by locating their
servers there. With the now abandoned AIX there used to
be an other IX in Accra run by different members. While
basis is laid, further steps are to be taken [4, 2, 8].
5. WEB SITE ARCHITECTURE, LATENCY
AND THE LOADING PROCESS
The work of Zaki et al. is the centre of this section. It
gives insights on how assumptions within design decisions
of web services on conditions much different from the re-
alities of Internet access in developing regions cause avoid-
able round trips leading to huge delays. Web pages become
more complex and featureful and the dynamically loaded
content through AJAX requests implies that even more re-
quests need to be done until the page is in a complete state.
Utilising browser automation through the Selenium frame-
work access to hundreds of Alexa’s TOP sites (global and
local) has been recorded to HAR reports (HTTP Archive),
allowing a deeper look on how loading times are constituted
and which dependencies are in the critical path through the
Wprof tool. In 80% of the cases an initial delay is incor-
porated because of HTTP redirects (like from name.tld to
www.name.tld) [30].
Similar work has been done by Sundaresan et al. to identify
bottlenecks [23, 24]. Like in [27] we can see in Zaki et al.
how blocking - as all TCP connections are in use - and com-
putation time have their share. But specially the time until
the connection is set up - which is name resolution and TCP
handshake - stands out with DNS having around 40% of the
time in 2012 and up to 14-24% in 2014 (although mind that
the amount to be downloaded increased in total). Time to
connect contributed 5% of page loading time in 2012 and
15-20% in 2014. All data according to the two clients in
Accra via campus WiFi and 3G as well as one 3G client in
a more remote area.
While it might be interesting trying to reduce the blocking
and speeding up receiving time by allowing more concurrent
TCP connections it is also observed that competing TCP
connections can have negative impact on each other [5].
Often it might be worth to reflect on which part of the web
site has the highest value to the user. Blocking of images or
JavaScript as well as using text based browsers will reduce
load time of media rich sites. It was found that filtering of
advertisement for products which are anyway not available
in developing countries should be considered [9]. There is
also a difference in browsers on the ability to render half
received HTML sites to get quick feedback. But as they
- like xlinks2 - are rather simple compared to established
browsers, they might just be used in addition. In contrast
the usage of lightweight REST APIs through AJAX could
also end up in faster interaction because it is not any more
the whole HTML site which has to be send by the server.
As web interfaces for mail services are commonly used Du et
al. suggest usage of SMTP because of advantages in saving
data transfer [9]. In addition it would also avoid slower inter-
action through the browser and prevent loss of the written
mail in case of transmission failure.
HTTP 1.1 pipelining allows to fetch more than one resource
through a TCP connection. Similarly, SPDY brings even
more flexible multiplexing of requests and the possibility
to mark priorities, but also adds in-built compression and
works exclusively over encrypted channels (i.e. to outper-
form HTTP without TLS the gain of SPDY must amortise
the increased start up time). The interesting parts for con-
nections with high round trip times are server push and
server hint, so that the server can continue transferring the
probably next wanted content without the need of a client
request or just indicate that the client should probably start
requesting some content which is often requested soon by
other clients. This technique is handy as the manual flatten-
ing of the dependencies to avoid too many recursive requests
(and shorten a long critical path) is not necessary any more.
On the downside usage of only one TCP connection results
in an entire delay under conditions of high package loss [29].
Yet in the tests in Ghana an overall speed up of 40% was
observed and further refinements might be possible [30].
6. INTERNET ACCESS WITH LOW OR IN-
TERMITTENT CONNECTIVITY
The global digital divide has also lead to solutions paying
attention to the conditions in the global south. Aggressive
caching and constructing an index for offline browsing take
care that benefits of Internet access are also available during
times of an unstable or non-existing connection and even
setups are used with the absence of a direct line and data
is physically transfered. As the approach on information
exchange is quite different, challenges of these mechanical
backhaul systems include how to break down the process of
web searches to be one round only as the “latency” is now
in the order of hours while much data can be transferred at
once on storage devices [6, 12, 20].
A more conventional system which serves a customised in-
terface utilising a proxy was tested in Hohoe [15]. To make
best usage of the connection, requests are queued and can
even be scheduled for less busy times. It also uses filters to
reduce page size and a shared cache which can be accessed
through an offline index.
1
The other popular Internet access method which is raising
analogical as in the global north is through second and third
generation mobile Internet subscriptions. Due to speed and
data volume costs being the regulative factors Sambasivan
et al. developed an application enabling mobile users to
be informed about the site prize (as by download size) and
gives short links/bookmarks instead of going through search
engines to reduce delay. They noted on the conditions of
mobile Internet usage and show observations with their cus-
tom infrastructure which was used to monitor page size by
a non-caching proxy [19].
7. CACHING AND PREFETCHING
Caching can play a major role in speeding up page loading
time under the here discussed conditions [9]. The emersion
of CDNs and common JavaScript libraries allow already re-
ceived content to be used between different web sites. Serv-
ing the common HTTP redirections from a cache resulted in
20% faster page load in the test set of Zaki et al. [30]. But a
caching mechanism which is more intrusive than the internal
browser cache may result to unwanted side effects. Also in
contrast to static pages dynamic web pages are altered and
personalised according to the situation of the request and
1
It should be considered to include access to a Wikipedia
dump as well like provided by the Evopedia application.
caching them is be not of much use because of their limited
life time. New approaches like micro-caching appear to be
promising as style and programme code of a web site are
highly cacheable [28].
Not much applicable to cellular access, caching and prefetch-
ing in home routers has various aspects. From obvious DNS
caching it reaches to caching of TCP connections and even
limited content caching. By also introducing popularity
based prefetching for DNS, TCP and HTTP requests the
extensive studies of Sundaresan et al. end in cutting down
page loading times to the half [23, 24].
Yet cellular access already has a longer path until reach-
ing the point where such methods can be deployed and also
non-cacheable encrypted traffic in the “post-Snowden era”
is increasing and shipped by major web services. In 2014
15% of the Alexa sites accessed at the research of Zaki et al.
where using TLS/SSL [30]. Mid of this year the initiative
Let’s encrypt
2
plans to give free of charge TLS certificates
in an automatical manner and is backed by Mozilla, Cisco,
Akamai and IdenTrust SSL CA. This will influence not only
start up time for connections as already written in the study
but also the effectiveness of caches.
8. LATENCY IN DNS RESOLUTION AND
POSSIBLE IMPROVEMENTS
The bad performance of DNS request in Ghana was the out-
standing part of the recent study of Zaki et al.. Specially
when a page load incorporates 11 name resolutions, as ob-
served during their research, the round trip times slow down
the page load tremendously by making up 72% of the time.
Thus they do not only propose caching
3
but also find that a
better DNS infrastructure is needed in Ghana. Concerning
caching is has to be considered that DNS resolution is also
used for load balancing and localisation.
The distance to root servers which are mainly located in
North America and Europe is one experienced nuisance. But
also the often occurring recursive requests to the local DNS
server, root server, top level domain server to sub domain
servers in combination with the high round trip times sum
up to a big delay. It was shown that by assuming a fictive
zero latency for root servers in Ghana the access times could
meet with the ones from Europe and North America. But
not only placing root servers closer is giving a solution, also
routes have to be optimally chosen:
“We found that nearly all of the traceroutes’ first
hops were outside Ghana and somewhere in Eu-
rope (Switzerland, UK, etc.), from where, the
route sometimes diverts to the US or Asia. None
of the routes went to DNS root servers that are
geographically nearby.” [30]
Others say that “ISPs should be more proactive to deploy
local root anycast instances to improve their DNS query
2
https://letsencrypt.org/
3
In the Accra WiFi case caching is done on university cam-
pus which leads to improvement from 2012 to 2014.
performance” [16]. Coming back to the issue of prefetching
quite a while ago it was proposed to practise piggybacking
and putting answers to queries for related domain names in
the additional answer field of a DNS response [21].
9. PERFORMANCE OF TCP UNDER CON-
DITIONS OF HIGH LATENCY AND PACK-
AGE LOSS
The TCP slow start behaviour for congestion control is well
known and the three-way handshake for opening a connec-
tion and the way reliability is achieved by receiving ACKs
has constraints on throughput (e.g. through head-of-line
blocking) when round trip times are high and/or package
loss occurs. A study claims that under their conditions loss
makes web page latency 5 times slower, transfers take 5
times longer on average and 77% of losses were recovered
by timeout instead of fast recovery. It is up to the algo-
rithms to reduce these effects and contributions of the pro-
posed methods to the Linux kernel have been done in 2013
[10]. Also RFC 6298 wants to improve the situation, but
also spreading of newer software often needs longer time.
While the focus lays on speed and latency, stability of TCP
connections is a challenge as well. Traditionally package
drop is considered to come from congestion of a router. But
this assumption is not accurate for unstable links and wire-
less transfers. Instead of reducing sending rate because of
timeouts it might even help to retransmit quickly to pre-
vent further collapse of the connection. This is also the case
when competing connections share one link and some flows
take the whole bandwidth while others do not recover and
experience, thus leading to unfairness. A middlebox solu-
tion by Chen et al. has been tested to reduce these issues by
preventing timeouts through scheduling of TCP connections
and tactical package drops [5]. An other aspect impacting
TCP latency is the bufferbloat in cellular networks which is
discussed in the last section.
Adjustments which, like changing initial TCP window size
to 10 segments, might have helped in certain situations, but
research in general solutions by more robust algorithms per-
forming in any corner of the world appears to be more de-
sirable.
9.1 Alternatives to TCP
Other protocols like SCTP could in fact rule out the short-
comings of TCP which dates back to beginnings in 1974.
But changing away from such an established protocol is a
long way and promoting improvements to TCP instead of
new protocol are adopted.
But in the context of Internet connectivity in developing
countries SCTP
4
has some properties to note on. On one
hand the four-way handshake which should prevent DOS
attacks seems to go contrary to the efforts to reduce start
up time. On the other the flow control algorithms might
match better with the conditions and thus reducing loading
times - but this is to be investigated. Clearly the features of
taking multiple paths and having multiple addresses (multi
home) do fit to the unsteady nature of connectivity in remote
4
http://tools.ietf.org/html/rfc7053
areas.
9.2 SSL/TLS
Normal start up has almost 10 round trips and it is an on-
going field to search for optimisations. False start was one
attempt, now caching of server parameters - called fast-track
- is promising [3, 30].
10. THE MOBILE CASE, RELIABILITY AND
LAST MILE LATENCY
Much can be learned from the studies by Koradia et al.
on Indian cellular networks in urban and rural areas [14].
The gap between urban and rural network performance is
not necessarily having technical reasons but also reflects
provider decisions of resource allocation. Also influence of
network configuration at the gateway on latency of 3G links
and throughput has been observed. The occurrence of stalled
connections has been investigated and it is related to bursty
transmission and the number of packages in flight, but some
questions remain open. Some packets were delayed by 30
seconds. It was seen that in general connections without
stall were 1.25 to 3 times faster, 2G connections were af-
fected 40%, 3G 30%. The 5 minute transfers spend 15%
of the time in this state. The effects could be reduced by
changing the receive window size, but on cost of throughput.
Bufferbloat is a phenomenon in cellular network where pack-
ets are not directly forwarded but kept in a per client buffer
which can have impact on TCP performance [1, 22]. Under
these conditions the congestion control algorithms NewReno,
Westwood+ and CUBIC performed differently when more
then one connection is in use, specially with mixed short
and long staying connections.
When new cellular networks are installed it is thus not enough
to only look on the higher bandwidth but also efforts must
be taken towards optimal latency. As there are situations
where TCP recovery does not work well, providers have to
make sure that they rule out reasons contributing to unsta-
ble TCP connections.
11. CONCLUSIONS
The aspects of slow Internet access in Ghana are diverse.
The total latency in web requests consists of the various
parts taking their share. This also includes that there are
many ways to improve the situation, be it on the interna-
tional level down to even improving the behaviour of one’s
own device. Yet it needs people working on these issues,
be it in the non-profit GIX, in rural communities and school
LANs which are not connected to the Internet or in develop-
ing robust and efficient communication protocols. But after
all the digital divide is connected to the economical situa-
tion and relation of the global south and north towards each
other.
12. REFERENCES
[1] S. Alfredsson, G. D. Giudice, J. Garcia, A. Brunstrom,
L. D. Cicco, and S. Mascolo. Impact of TCP
congestion control on bufferbloat in cellular networks.
In IEEE 14th International Symposium on A World
of Wireless, Mobile and Multimedia Networks”,
WoWMoM 2013, Madrid, Spain, June 4-7, 2013,
pages 1–7, 2013.
[2] Ayitey Bulley. The Peering Scene in Ghana. Africa
Peering and Interconnect Forum Accra, Aug. 2011.
[3] Bob Briscoe and Anna Brunstrom and Andreas
Petlund and David Hayes and David Ros and Ing-Jyh
Tsang and Stein Gjessing and Gorry Fairhurst and
Carsten Griwodz and Michael Welzl. Reducing
Internet Latency: A Survey of Techniques and their
Merits. IEEE Commun. Surveys Tuts., 2014.
[4] D. Charles Amega-Selorm, Muriuki Mureithi. Impact
of IXPs - A review of the experiences of Ghana, Kenya
and South Africa. Technical report, Open Society
Institute, Aug. 2009.
[5] J. Chen, L. Subramanian, J. R. Iyengar, and B. Ford.
TAQ: enhancing fairness and performance
predictability in small packet regimes. In Ninth
Eurosys Conference 2014, EuroSys 2014, Amsterdam,
The Netherlands, April 13-16, 2014, page 7, 2014.
[6] J. Chen, L. Subramanian, and J. Li. RuralCafe: Web
Search in the Rural Developing World. In Proceedings
of the 18th International Conference on World Wide
Web, WWW 2009, Madrid, Spain, April 20-24, 2009,
pages 411–420, 2009.
[7] M. Chetty, S. Sundaresan, S. Muckaden, N. Feamster,
and E. Calandro. Measuring broadband performance
in South Africa. In Proceedings of the 4th Annual
Symposium on Computing for Development, ACM
DEV-4 ’13, pages 1:1–1:10, New York, NY, USA,
2013. ACM.
[8] Dennis Weller. Blurring Boundaries: Global and
Regional IP Interconnection. ITU GSR Discussion
Paper, 2012.
[9] B. Du, M. J. Demmer, and E. A. Brewer. Analysis of
WWW traffic in Cambodia and Ghana. In Proceedings
of the 15th international conference on World Wide
Web, WWW 2006, Edinburgh, Scotland, UK, May
23-26, 2006, pages 771–780, 2006.
[10] T. Flach, N. Dukkipati, A. Terzis, B. Raghavan,
N. Cardwell, Y. Cheng, A. Jain, S. Hao,
E. Katz-Bassett, and R. Govindan. Reducing web
latency: The virtue of gentle aggression. In
Proceedings of the ACM sigcomm 2013 conference on
sigcomm, SIGCOMM ’13, pages 159–170, New York,
NY, USA, 2013. ACM.
[11] GhanaWeb B.V. Mobile internet subscription hits
10.56m.
http: // www. ghanaweb. com/ GhanaHomePage/
NewsArchive/ artikel. php? ID= 291200 , 6 Nov. 2013.
[12] S. Guo, M. H. Falaki, E. A. Oliver, S. U. Rahman,
A. Seth, M. A. Zaharia, and S. Keshav. Very low-cost
internet access using KioskNet. Computer
Communication Review, 37(5):95–100, 2007.
[13] International Telecommunication Union. Measuring
the Information Society Report 2014 Executive
Summary. 2014.
[14] Z. Koradia, G. Mannava, A. Raman, G. Aggarwal,
V. J. Ribeiro, A. Seth, S. Ardon, A. Mahanti, and
S. Triukose. First impressions on the state of cellular
data connectivity in India. In Annual Symposium on
Computing for Development, ACM DEV-4, Cape
Town, South Africa - December 06 - 07, 2013, page 3,
2013.
[15] L. D. Li and J. Chen. TroTro: Web Browsing and
User Interfaces in Rural Ghana. In Proceedings of the
Sixth International Conference on Information and
Communication Technologies and Development: Full
Papers - Volume 1, ICTD ’13, pages 185–194, New
York, NY, USA, 2013. ACM.
[16] J. Liang, J. Jiang, H. Duan, K. Li, and J. Wu.
Measuring Query Latency of Top Level DNS Servers.
In Proceedings of the 14th International Conference on
Passive and Active Measurement, PAM’13, pages
145–154, Berlin, Heidelberg, 2013. Springer-Verlag.
[17] Modern Ghana Media Communication Ltd. Mobile
penetration cross 100% in Ghana.
http: // www. modernghana. com/ news/ 443415/ 1/
mobile-penetration-cross-100-in-ghana. html , 4
Feb. 2013.
[18] Popular Science. Fiber-optic transatlantic cable could
save milliseconds, millions by speeding data to stock
traders.
http: // www. popsci. com/ technology/ article/
2011-04/ new-transatlantic-cable-will-speed-
information-exchange-price , 25 Nov. 2011.
[19] N. Sambasivan, P. Lee, G. Hecht, P. M. Aoki, M.-I.
Carrera, J. Chen, D. P. Cohn, P. Kruskall,
E. Wetchler, M. Youssefmir, and A. T. Larssen. Chale,
How Much it Cost to Browse?: Results from a Mobile
Data Price Transparency Trial in Ghana. In
Proceedings of the Sixth International Conference on
Information and Communication Technologies and
Development: Full Papers - Volume 1, ICTD ’13,
pages 13–23, New York, NY, USA, 2013. ACM.
[20] A. Seth, D. Kroeker, M. Zaharia, S. Guo, and
S. Keshav. Low-cost Communication for Rural
Internet Kiosks Using Mechanical Backhaul. In
Proceedings of the 12th Annual International
Conference on Mobile Computing and Networking,
MobiCom ’06, pages 334–345, New York, NY, USA,
2006. ACM.
[21] H. Shang and C. E. Wills. Piggybacking related
domain names to improve DNS performance. Comput.
Netw., 50(11):1733–1748, Aug. 2006.
[22] S. Sundaresan, W. de Donato, N. Feamster,
R. Teixeira, S. Crawford, and A. Pescap`e. Broadband
internet performance: A view from the gateway. In
Proceedings of the ACM SIGCOMM 2011 Conference
on Applications, Technologies, Architectures, and
Protocols for Computer Communications, Toronto,
ON, Canada, August 15-19, 2011, pages 134–145,
2011.
[23] S. Sundaresan, N. Feamster, R. Teixeira, and
N. Magharei. Measuring and mitigating web
performance bottlenecks in broadband access
networks. In Internet Measurement Conference,
IMC’13, Barcelona, Spain, October 23-25, 2013, pages
213–226, 2013.
[24] S. Sundaresan, N. Magharei, N. Feamster, and
R. Teixeira. Characterizing and Mitigating Web
Performance Bottlenecks in Broadband Access
Networks. Proceedings of the ACM
SIGMETRICS/international conference on
Measurement and modeling of computer systems, 2013.
[25] Surfline Communications Limited. http: // www.
surflinegh. com/ en/ about-surfline/ company/ .
[26] United Nations. The Millennium Development Goals
Report 2014. http:
// www. undp. org/ content/ dam/ undp/ library/
MDG/ english/ UNDP_ MDGReport_ EN_ 2014Final1. pdf ,
page 52, 2014.
[27] X. S. Wang, A. Balasubramanian, A. Krishnamurthy,
and D. Wetherall. Demystifying Page Load
Performance with WProf. In Proceedings of the 10th
USENIX Symposium on Networked Systems Design
and Implementation, NSDI 2013, Lombard, IL, USA,
April 2-5, 2013, pages 473–485, 2013.
[28] X. S. Wang, A. Krishnamurthy, and D. Wetherall.
How Much Can We Micro-Cache Web Pages? In
Proceedings of the 2014 Conference on Internet
Measurement Conference, IMC 2014, Vancouver, BC,
Canada, November 5-7, 2014, pages 249–256, 2014.
[29] Xiao Sophia Wang and Aruna Balasubramanian and
Arvind Krishnamurthy and David Wetherall. How
Speedy is SPDY? In 11th USENIX Symposium on
Networked Systems Design and Implementation (NSDI
14), pages 387–399, Seattle, WA, Apr. 2014. USENIX
Association.
[30] Y. Zaki, J. Chen, T. P
¨
otsch, T. Ahmad, and
L. Subramanian. Dissecting web latency in Ghana. In
Proceedings of the 2014 Conference on Internet
Measurement, IMC ’14, pages 241–248, New York,
NY, USA, 2014. ACM.
Web Connectivity in Ghana: A survey
Kai Lüke
Institute of Computer Science
Department of Mathematics and Computer Science
Freie Universität Berlin
Berlin, Germany

kai.lueke@fu-berlin.de
ABSTRACT
Starting from the research “Dissecting web latency in Ghana”
of Zaki et al. the field of Internet connectivity in Ghana is
explored and the different aspects like latency, DNS resolution times, complexity of web sites, caching and prefetching, TLS/SSL, TCP and cellular network performance, role
of ISPs, Internet Exchange Points, are brought in relation.
Measurement approaches and tailored solutions for areas
with bad connectivity are listed. Possible ways of improvement are discussed.

But the complexity of web site construction and data amount
per page load has grown faster than the Internet connectivity in developing countries, thus leading to slow page loading times while bandwidth has increased in Ghana [30, 15].
These are not only frustration in usability but can also eliminate certain use cases. The following overview will show
that growing bandwidth alone can not solve the problem
due to the impact of latency and package loss on common
transport protocols and the nested dependencies of requests
during page load.

Categories and Subject Descriptors

2.

C.2.6 [Internetworking]: Standards; C.4 [Performance
of Systems]: Reliability, availability, and serviceability

Keywords
Developing Countries, Internet, Latency, Performance, TCP,
Mobile, Cellular

1.

INTRODUCTION

The Millennium Development Goals, which were supposed
to be reached at the end of this year, also include measurements in target 8F for “mak[ing] available the benefits of new
technologies, especially information and communications”.
By these numbers almost 40% of the world population have
Internet access [26]. Despite the big gap between global south
and global north in a variety of aspects, the huge number of
6.9 billion mobile-cellular subscriptions could be utilised to
raise Internet connectivity quickly. As in Ghana the population number was outnumbered [17], because people tend
to have more than one subscription, the distribution might
not be equal, but has led to 40.7% mobile Internet penetration in 2013 [11]. Concerning the ICT Development Index
Ghana ranks 6th among African countries [13] and has direct connection to major undersea cables [30]. Recently LTE
is rolled out as 4G cellular standard and the licence requires
coverage of half the country within five years [25].

SPEED OF DATA TRANSMISSION

Round trip times to servers on the other half of the world can
not be faster than the theoretical limit which is the speed of
light through glass fiber. By example taking s = 12767 km
as distance from Accra to Seoul and maximum speed because
c
≈ 200000 km
the
of the refractive index for glass is v = 1.5
s
round trip takes always more than t = 2·s
≈
128
ms.
v
The importance of low latency when interacting with web
sites is shown at the example of usage reduction for search
engines after introducing an artificial delay as shortly described in the detailed survey of Briscoe et al. on reducing
latency [3].
The links between Europe and North America are quite close
to the theoretical limit and even almost half a billion dollar is
spend for few millisecond gains by establishing new undersea
cables [18]. Developing countries are not close to optimum
and so improvements are possible in many ways. But still
it must be questioned if an access outside the continent is
justified or whether the content and services can be located
at infrastructure nearby.
While throughput can be increased by adding more links,
the performance of TCP is affected by latency. Also the
recursive dependencies in requests while loading web pages
result, as described later, in connectedness of latency and
throughput.

3.

METHODS OF DATA COLLECTION

Statistical data and specially detailed recordings can be taken
at various places. In 2006 the operator of a main Internet
cafe in Accra allowed analysis of logs containing 13.7 million requests [9]. Also browser access to hundreds of websites ranked popular by Alexa has been profiled at Accra,
Kumawu and Hohoe (all in the south) in 2012 and 2014 by
Zaki et al. [30] using a similar approach to [27]. These ap-

proaches help in tracing down how the slow loading times experienced by users are composed by the single parts. But as
they happen on application layer, the reasons for decreased
throughput and higher round trip times for packages can
not be examined. Regarding to Pötsch the charts in [30] do
also only include web pages which were able to be received
fully which was not the case for all websites. However the
paper focuses rather on DNS latency and solution space in
the browser. Other works referenced in the chapter here on
TCP contain research for these incidents.
Others showed how measurement devices for broadband and
cellular connections can be deployed at diverse places in India and South Africa [14, 7]. The special benchmark routers
contribute to appropriate long term data and also eliminate
possible disturbance to accuracy when the line is shared
with other users or due to the local network setup (ibid.
and [22]). Also applications on smartphones gain interest
as they provide an easy way to reach data sources [7, 19].
All these works also note the faced challenges like unstable
power supply and absence of connectivity or suspiciousness
towards the smartphone application.
Measurement can help in empowering users which constantly
see a much slower connection speed than advertised by the
provider [7].

4.

INTERNET SERVICE PROVIDERS AND
INTERNET EXCHANGE POINTS

The contrast between urban and rural areas regarding (fast)
internet access is not only present in developing countries,
but can therefore significantly determine whether a certain
limit is exceeded which makes usage of demanding services
not worthwhile any more. Particular for Ghana is also the
gradient from the south to the overall less prosperous north.
As urban inhabitants tend to be a better income source,
providers can be inveigled to privilege them in resource allocation e.g. as it was observed by which gateways were given
to urban and rural cellular clients in India [14].
Studies where measurements were realised for more than one
provider indicate differences not only in bandwidth and high
load during popular usage hours but also latency, specially
last mile latency, which is the time to get to the first hop in
the provider’s IP infrastructure [22]. It might even be that
just network configurations bring betterment for cellular access [14]. In addition to correlation of the previous findings
[7] also notes that the interconnection of ISPs in the region
including neighbour countries plays a huge role when routes
outside of the continent can be avoided.
Internet exchange points bring not only ISPs together but
also provide a place where content can be made available.
The GIX in Accra operates for a decade now and contributed
in reducing access times for those organisations who are interconnected there. On average the traffic is around 500
MB/s. Specially the common content distribution networks
and popular web services need to take part by locating their
servers there. With the now abandoned AIX there used to
be an other IX in Accra run by different members. While
basis is laid, further steps are to be taken [4, 2, 8].

5.

WEB SITE ARCHITECTURE, LATENCY
AND THE LOADING PROCESS

The work of Zaki et al. is the centre of this section. It
gives insights on how assumptions within design decisions
of web services on conditions much different from the realities of Internet access in developing regions cause avoidable round trips leading to huge delays. Web pages become
more complex and featureful and the dynamically loaded
content through AJAX requests implies that even more requests need to be done until the page is in a complete state.
Utilising browser automation through the Selenium framework access to hundreds of Alexa’s TOP sites (global and
local) has been recorded to HAR reports (HTTP Archive),
allowing a deeper look on how loading times are constituted
and which dependencies are in the critical path through the
Wprof tool. In 80% of the cases an initial delay is incorporated because of HTTP redirects (like from name.tld to
www.name.tld ) [30].
Similar work has been done by Sundaresan et al. to identify
bottlenecks [23, 24]. Like in [27] we can see in Zaki et al.
how blocking - as all TCP connections are in use - and computation time have their share. But specially the time until
the connection is set up - which is name resolution and TCP
handshake - stands out with DNS having around 40% of the
time in 2012 and up to 14-24% in 2014 (although mind that
the amount to be downloaded increased in total). Time to
connect contributed 5% of page loading time in 2012 and
15-20% in 2014. All data according to the two clients in
Accra via campus WiFi and 3G as well as one 3G client in
a more remote area.
While it might be interesting trying to reduce the blocking
and speeding up receiving time by allowing more concurrent
TCP connections it is also observed that competing TCP
connections can have negative impact on each other [5].
Often it might be worth to reflect on which part of the web
site has the highest value to the user. Blocking of images or
JavaScript as well as using text based browsers will reduce
load time of media rich sites. It was found that filtering of
advertisement for products which are anyway not available
in developing countries should be considered [9]. There is
also a difference in browsers on the ability to render half
received HTML sites to get quick feedback. But as they
- like xlinks2 - are rather simple compared to established
browsers, they might just be used in addition. In contrast
the usage of lightweight REST APIs through AJAX could
also end up in faster interaction because it is not any more
the whole HTML site which has to be send by the server.
As web interfaces for mail services are commonly used Du et
al. suggest usage of SMTP because of advantages in saving
data transfer [9]. In addition it would also avoid slower interaction through the browser and prevent loss of the written
mail in case of transmission failure.
HTTP 1.1 pipelining allows to fetch more than one resource
through a TCP connection. Similarly, SPDY brings even
more flexible multiplexing of requests and the possibility
to mark priorities, but also adds in-built compression and
works exclusively over encrypted channels (i.e. to outperform HTTP without TLS the gain of SPDY must amortise

the increased start up time). The interesting parts for connections with high round trip times are server push and
server hint, so that the server can continue transferring the
probably next wanted content without the need of a client
request or just indicate that the client should probably start
requesting some content which is often requested soon by
other clients. This technique is handy as the manual flattening of the dependencies to avoid too many recursive requests
(and shorten a long critical path) is not necessary any more.
On the downside usage of only one TCP connection results
in an entire delay under conditions of high package loss [29].
Yet in the tests in Ghana an overall speed up of 40% was
observed and further refinements might be possible [30].

6.

INTERNET ACCESS WITH LOW OR INTERMITTENT CONNECTIVITY

The global digital divide has also lead to solutions paying
attention to the conditions in the global south. Aggressive
caching and constructing an index for offline browsing take
care that benefits of Internet access are also available during
times of an unstable or non-existing connection and even
setups are used with the absence of a direct line and data
is physically transfered. As the approach on information
exchange is quite different, challenges of these mechanical
backhaul systems include how to break down the process of
web searches to be one round only as the “latency” is now
in the order of hours while much data can be transferred at
once on storage devices [6, 12, 20].
A more conventional system which serves a customised interface utilising a proxy was tested in Hohoe [15]. To make
best usage of the connection, requests are queued and can
even be scheduled for less busy times. It also uses filters to
reduce page size and a shared cache which can be accessed
through an offline index.1
The other popular Internet access method which is raising
analogical as in the global north is through second and third
generation mobile Internet subscriptions. Due to speed and
data volume costs being the regulative factors Sambasivan
et al. developed an application enabling mobile users to
be informed about the site prize (as by download size) and
gives short links/bookmarks instead of going through search
engines to reduce delay. They noted on the conditions of
mobile Internet usage and show observations with their custom infrastructure which was used to monitor page size by
a non-caching proxy [19].

7.

Not much applicable to cellular access, caching and prefetching in home routers has various aspects. From obvious DNS
caching it reaches to caching of TCP connections and even
limited content caching. By also introducing popularity
based prefetching for DNS, TCP and HTTP requests the
extensive studies of Sundaresan et al. end in cutting down
page loading times to the half [23, 24].
Yet cellular access already has a longer path until reaching the point where such methods can be deployed and also
non-cacheable encrypted traffic in the “post-Snowden era”
is increasing and shipped by major web services. In 2014
15% of the Alexa sites accessed at the research of Zaki et al.
where using TLS/SSL [30]. Mid of this year the initiative
Let’s encrypt 2 plans to give free of charge TLS certificates
in an automatical manner and is backed by Mozilla, Cisco,
Akamai and IdenTrust SSL CA. This will influence not only
start up time for connections as already written in the study
but also the effectiveness of caches.

8.

It should be considered to include access to a Wikipedia
dump as well like provided by the Evopedia application.

LATENCY IN DNS RESOLUTION AND
POSSIBLE IMPROVEMENTS

The bad performance of DNS request in Ghana was the outstanding part of the recent study of Zaki et al.. Specially
when a page load incorporates 11 name resolutions, as observed during their research, the round trip times slow down
the page load tremendously by making up 72% of the time.
Thus they do not only propose caching3 but also find that a
better DNS infrastructure is needed in Ghana. Concerning
caching is has to be considered that DNS resolution is also
used for load balancing and localisation.
The distance to root servers which are mainly located in
North America and Europe is one experienced nuisance. But
also the often occurring recursive requests to the local DNS
server, root server, top level domain server to sub domain
servers in combination with the high round trip times sum
up to a big delay. It was shown that by assuming a fictive
zero latency for root servers in Ghana the access times could
meet with the ones from Europe and North America. But
not only placing root servers closer is giving a solution, also
routes have to be optimally chosen:

CACHING AND PREFETCHING

Caching can play a major role in speeding up page loading
time under the here discussed conditions [9]. The emersion
of CDNs and common JavaScript libraries allow already received content to be used between different web sites. Serving the common HTTP redirections from a cache resulted in
20% faster page load in the test set of Zaki et al. [30]. But a
caching mechanism which is more intrusive than the internal
browser cache may result to unwanted side effects. Also in
contrast to static pages dynamic web pages are altered and
personalised according to the situation of the request and
1

caching them is be not of much use because of their limited
life time. New approaches like micro-caching appear to be
promising as style and programme code of a web site are
highly cacheable [28].

“We found that nearly all of the traceroutes’ first
hops were outside Ghana and somewhere in Europe (Switzerland, UK, etc.), from where, the
route sometimes diverts to the US or Asia. None
of the routes went to DNS root servers that are
geographically nearby.” [30]

Others say that “ISPs should be more proactive to deploy
local root anycast instances to improve their DNS query
2

https://letsencrypt.org/
In the Accra WiFi case caching is done on university campus which leads to improvement from 2012 to 2014.
3

performance” [16]. Coming back to the issue of prefetching
quite a while ago it was proposed to practise piggybacking
and putting answers to queries for related domain names in
the additional answer field of a DNS response [21].

9.

PERFORMANCE OF TCP UNDER CONDITIONS OF HIGH LATENCY AND PACKAGE LOSS

The TCP slow start behaviour for congestion control is well
known and the three-way handshake for opening a connection and the way reliability is achieved by receiving ACKs
has constraints on throughput (e.g. through head-of-line
blocking) when round trip times are high and/or package
loss occurs. A study claims that under their conditions loss
makes web page latency 5 times slower, transfers take 5
times longer on average and 77% of losses were recovered
by timeout instead of fast recovery. It is up to the algorithms to reduce these effects and contributions of the proposed methods to the Linux kernel have been done in 2013
[10]. Also RFC 6298 wants to improve the situation, but
also spreading of newer software often needs longer time.
While the focus lays on speed and latency, stability of TCP
connections is a challenge as well. Traditionally package
drop is considered to come from congestion of a router. But
this assumption is not accurate for unstable links and wireless transfers. Instead of reducing sending rate because of
timeouts it might even help to retransmit quickly to prevent further collapse of the connection. This is also the case
when competing connections share one link and some flows
take the whole bandwidth while others do not recover and
experience, thus leading to unfairness. A middlebox solution by Chen et al. has been tested to reduce these issues by
preventing timeouts through scheduling of TCP connections
and tactical package drops [5]. An other aspect impacting
TCP latency is the bufferbloat in cellular networks which is
discussed in the last section.
Adjustments which, like changing initial TCP window size
to 10 segments, might have helped in certain situations, but
research in general solutions by more robust algorithms performing in any corner of the world appears to be more desirable.

9.1

Alternatives to TCP

Other protocols like SCTP could in fact rule out the shortcomings of TCP which dates back to beginnings in 1974.
But changing away from such an established protocol is a
long way and promoting improvements to TCP instead of
new protocol are adopted.
But in the context of Internet connectivity in developing
countries SCTP4 has some properties to note on. On one
hand the four-way handshake which should prevent DOS
attacks seems to go contrary to the efforts to reduce start
up time. On the other the flow control algorithms might
match better with the conditions and thus reducing loading
times - but this is to be investigated. Clearly the features of
taking multiple paths and having multiple addresses (multi
home) do fit to the unsteady nature of connectivity in remote
4

http://tools.ietf.org/html/rfc7053

areas.

9.2

SSL/TLS

Normal start up has almost 10 round trips and it is an ongoing field to search for optimisations. False start was one
attempt, now caching of server parameters - called fast-track
- is promising [3, 30].

10.

THE MOBILE CASE, RELIABILITY AND
LAST MILE LATENCY

Much can be learned from the studies by Koradia et al.
on Indian cellular networks in urban and rural areas [14].
The gap between urban and rural network performance is
not necessarily having technical reasons but also reflects
provider decisions of resource allocation. Also influence of
network configuration at the gateway on latency of 3G links
and throughput has been observed. The occurrence of stalled
connections has been investigated and it is related to bursty
transmission and the number of packages in flight, but some
questions remain open. Some packets were delayed by 30
seconds. It was seen that in general connections without
stall were 1.25 to 3 times faster, 2G connections were affected 40%, 3G 30%. The 5 minute transfers spend 15%
of the time in this state. The effects could be reduced by
changing the receive window size, but on cost of throughput.
Bufferbloat is a phenomenon in cellular network where packets are not directly forwarded but kept in a per client buffer
which can have impact on TCP performance [1, 22]. Under
these conditions the congestion control algorithms NewReno,
Westwood+ and CUBIC performed differently when more
then one connection is in use, specially with mixed short
and long staying connections.
When new cellular networks are installed it is thus not enough
to only look on the higher bandwidth but also efforts must
be taken towards optimal latency. As there are situations
where TCP recovery does not work well, providers have to
make sure that they rule out reasons contributing to unstable TCP connections.

11.

CONCLUSIONS

The aspects of slow Internet access in Ghana are diverse.
The total latency in web requests consists of the various
parts taking their share. This also includes that there are
many ways to improve the situation, be it on the international level down to even improving the behaviour of one’s
own device. Yet it needs people working on these issues,
be it in the non-profit GIX, in rural communities and school
LANs which are not connected to the Internet or in developing robust and efficient communication protocols. But after
all the digital divide is connected to the economical situation and relation of the global south and north towards each
other.

12.

REFERENCES

[1] S. Alfredsson, G. D. Giudice, J. Garcia, A. Brunstrom,
L. D. Cicco, and S. Mascolo. Impact of TCP
congestion control on bufferbloat in cellular networks.
In IEEE 14th International Symposium on ”A World
of Wireless, Mobile and Multimedia Networks”,

[2]
[3]

[4]

[5]

[6]

[7]

[8]

[9]

[10]

[11]

[12]

[13]

[14]

WoWMoM 2013, Madrid, Spain, June 4-7, 2013,
pages 1–7, 2013.
Ayitey Bulley. The Peering Scene in Ghana. Africa
Peering and Interconnect Forum Accra, Aug. 2011.
Bob Briscoe and Anna Brunstrom and Andreas
Petlund and David Hayes and David Ros and Ing-Jyh
Tsang and Stein Gjessing and Gorry Fairhurst and
Carsten Griwodz and Michael Welzl. Reducing
Internet Latency: A Survey of Techniques and their
Merits. IEEE Commun. Surveys Tuts., 2014.
D. Charles Amega-Selorm, Muriuki Mureithi. Impact
of IXPs - A review of the experiences of Ghana, Kenya
and South Africa. Technical report, Open Society
Institute, Aug. 2009.
J. Chen, L. Subramanian, J. R. Iyengar, and B. Ford.
TAQ: enhancing fairness and performance
predictability in small packet regimes. In Ninth
Eurosys Conference 2014, EuroSys 2014, Amsterdam,
The Netherlands, April 13-16, 2014, page 7, 2014.
J. Chen, L. Subramanian, and J. Li. RuralCafe: Web
Search in the Rural Developing World. In Proceedings
of the 18th International Conference on World Wide
Web, WWW 2009, Madrid, Spain, April 20-24, 2009,
pages 411–420, 2009.
M. Chetty, S. Sundaresan, S. Muckaden, N. Feamster,
and E. Calandro. Measuring broadband performance
in South Africa. In Proceedings of the 4th Annual
Symposium on Computing for Development, ACM
DEV-4 ’13, pages 1:1–1:10, New York, NY, USA,
2013. ACM.
Dennis Weller. Blurring Boundaries: Global and
Regional IP Interconnection. ITU GSR Discussion
Paper, 2012.
B. Du, M. J. Demmer, and E. A. Brewer. Analysis of
WWW traffic in Cambodia and Ghana. In Proceedings
of the 15th international conference on World Wide
Web, WWW 2006, Edinburgh, Scotland, UK, May
23-26, 2006, pages 771–780, 2006.
T. Flach, N. Dukkipati, A. Terzis, B. Raghavan,
N. Cardwell, Y. Cheng, A. Jain, S. Hao,
E. Katz-Bassett, and R. Govindan. Reducing web
latency: The virtue of gentle aggression. In
Proceedings of the ACM sigcomm 2013 conference on
sigcomm, SIGCOMM ’13, pages 159–170, New York,
NY, USA, 2013. ACM.
GhanaWeb B.V. Mobile internet subscription hits
10.56m.
http: // www. ghanaweb. com/ GhanaHomePage/
NewsArchive/ artikel. php? ID= 291200 , 6 Nov. 2013.
S. Guo, M. H. Falaki, E. A. Oliver, S. U. Rahman,
A. Seth, M. A. Zaharia, and S. Keshav. Very low-cost
internet access using KioskNet. Computer
Communication Review, 37(5):95–100, 2007.
International Telecommunication Union. Measuring
the Information Society Report 2014 Executive
Summary. 2014.
Z. Koradia, G. Mannava, A. Raman, G. Aggarwal,
V. J. Ribeiro, A. Seth, S. Ardon, A. Mahanti, and
S. Triukose. First impressions on the state of cellular
data connectivity in India. In Annual Symposium on
Computing for Development, ACM DEV-4, Cape
Town, South Africa - December 06 - 07, 2013, page 3,

2013.
[15] L. D. Li and J. Chen. TroTro: Web Browsing and
User Interfaces in Rural Ghana. In Proceedings of the
Sixth International Conference on Information and
Communication Technologies and Development: Full
Papers - Volume 1, ICTD ’13, pages 185–194, New
York, NY, USA, 2013. ACM.
[16] J. Liang, J. Jiang, H. Duan, K. Li, and J. Wu.
Measuring Query Latency of Top Level DNS Servers.
In Proceedings of the 14th International Conference on
Passive and Active Measurement, PAM’13, pages
145–154, Berlin, Heidelberg, 2013. Springer-Verlag.
[17] Modern Ghana Media Communication Ltd. Mobile
penetration cross 100% in Ghana.
http: // www. modernghana. com/ news/ 443415/ 1/
mobile-penetration-cross-100-in-ghana. html , 4
Feb. 2013.
[18] Popular Science. Fiber-optic transatlantic cable could
save milliseconds, millions by speeding data to stock
traders.
http: // www. popsci. com/ technology/ article/
2011-04/ new-transatlantic-cable-will-speedinformation-exchange-price , 25 Nov. 2011.
[19] N. Sambasivan, P. Lee, G. Hecht, P. M. Aoki, M.-I.
Carrera, J. Chen, D. P. Cohn, P. Kruskall,
E. Wetchler, M. Youssefmir, and A. T. Larssen. Chale,
How Much it Cost to Browse?: Results from a Mobile
Data Price Transparency Trial in Ghana. In
Proceedings of the Sixth International Conference on
Information and Communication Technologies and
Development: Full Papers - Volume 1, ICTD ’13,
pages 13–23, New York, NY, USA, 2013. ACM.
[20] A. Seth, D. Kroeker, M. Zaharia, S. Guo, and
S. Keshav. Low-cost Communication for Rural
Internet Kiosks Using Mechanical Backhaul. In
Proceedings of the 12th Annual International
Conference on Mobile Computing and Networking,
MobiCom ’06, pages 334–345, New York, NY, USA,
2006. ACM.
[21] H. Shang and C. E. Wills. Piggybacking related
domain names to improve DNS performance. Comput.
Netw., 50(11):1733–1748, Aug. 2006.
[22] S. Sundaresan, W. de Donato, N. Feamster,
R. Teixeira, S. Crawford, and A. Pescapè. Broadband
internet performance: A view from the gateway. In
Proceedings of the ACM SIGCOMM 2011 Conference
on Applications, Technologies, Architectures, and
Protocols for Computer Communications, Toronto,
ON, Canada, August 15-19, 2011, pages 134–145,
2011.
[23] S. Sundaresan, N. Feamster, R. Teixeira, and
N. Magharei. Measuring and mitigating web
performance bottlenecks in broadband access
networks. In Internet Measurement Conference,
IMC’13, Barcelona, Spain, October 23-25, 2013, pages
213–226, 2013.
[24] S. Sundaresan, N. Magharei, N. Feamster, and
R. Teixeira. Characterizing and Mitigating Web
Performance Bottlenecks in Broadband Access
Networks. Proceedings of the ACM
SIGMETRICS/international conference on
Measurement and modeling of computer systems, 2013.

[25] Surfline Communications Limited. http: // www.
surflinegh. com/ en/ about-surfline/ company/ .
[26] United Nations. The Millennium Development Goals
Report 2014. http:
// www. undp. org/ content/ dam/ undp/ library/
MDG/ english/ UNDP_ MDGReport_ EN_ 2014Final1. pdf ,
page 52, 2014.
[27] X. S. Wang, A. Balasubramanian, A. Krishnamurthy,
and D. Wetherall. Demystifying Page Load
Performance with WProf. In Proceedings of the 10th
USENIX Symposium on Networked Systems Design
and Implementation, NSDI 2013, Lombard, IL, USA,
April 2-5, 2013, pages 473–485, 2013.
[28] X. S. Wang, A. Krishnamurthy, and D. Wetherall.
How Much Can We Micro-Cache Web Pages? In
Proceedings of the 2014 Conference on Internet
Measurement Conference, IMC 2014, Vancouver, BC,
Canada, November 5-7, 2014, pages 249–256, 2014.
[29] Xiao Sophia Wang and Aruna Balasubramanian and
Arvind Krishnamurthy and David Wetherall. How
Speedy is SPDY? In 11th USENIX Symposium on
Networked Systems Design and Implementation (NSDI
14), pages 387–399, Seattle, WA, Apr. 2014. USENIX
Association.
[30] Y. Zaki, J. Chen, T. Pötsch, T. Ahmad, and
L. Subramanian. Dissecting web latency in Ghana. In
Proceedings of the 2014 Conference on Internet
Measurement, IMC ’14, pages 241–248, New York,
NY, USA, 2014. ACM.