=Paper=
{{Paper
|id=None
|storemode=property
|title=Assessing Applicability of Power-Efficient Embedded Devices for Micro-Cloud Computing
|pdfUrl=https://ceur-ws.org/Vol-1422/17.pdf
|volume=Vol-1422
|dblpUrl=https://dblp.org/rec/conf/itat/KrulisSYZ15
}}
==Assessing Applicability of Power-Efficient Embedded Devices for Micro-Cloud Computing==
https://ceur-ws.org/Vol-1422/17.pdf
J. Yaghob (Ed.): ITAT 2015 pp. 17–22
Charles University in Prague, Prague, 2015
Assessing Applicability of Power-Efficient Embedded Devices for Micro-Cloud
Computing
Martin Kruliš, Petr Stefan, Jakub Yaghob, and Filip Zavoral
Parallel Architectures/Algorithms/Applications Research Group
Faculty of Mathematics and Physics, Charles University in Prague
Malostranské nám. 25, Prague, Czech Republic
{krulis,yaghob,zavoral}@ksi.mff.cuni.cz, ptr.stef@gmail.com
Abstract: Distributed computing and cloud phenomenon With the growth of the cloud infrastructure, the power
have become an intensively studied topic in the past efficiency become a more and more important problem.
decade. These technologies have been enveloped with at- Despite the fact that the cloud services utilize underlying
tractive business models, where the customer pays only hardware more efficiently than it could have been used
for the resources or services which have been actually by individual users, the pressure to reduce power con-
utilized. Even though this popularity lead to rapid de- sumption of this infrastructure is raising steadily. One of
velopment of distributed algorithms, virtualization plat- the possibilities is to utilize more efficient hardware that
forms, and various cloud services, many issues are still requires less power to perform the same task. A quite
waiting to be solved. One of these issues is the question promising platform are the ARM CPUs which are cur-
of power efficiency. In this paper, we investigate possibil- rently utilized in mainstream mobile and other handheld
ities of applying single-board computers as platform for devices. However, the majority of enterprise servers and
distributed systems and cloud computing. These small de- professional solutions use CPUs based on x86 architec-
vices (such as Raspberry Pi) are quite power efficient and ture which have more computational power. Neverthe-
relatively cheap, so they may reduce the overall cost for less, these solutions are considered less efficient at least
cloud services. Furthermore, they may be employed to for some tasks.
create small clusters that could replace traditional enter-
prise servers and achieve lower cost and better robustness Some specialized problems cannot utilize cloud solu-
for some tasks. tions for various reasons such as security or domain-
Keywords: reliability, distributed systems, cloud comput- specific constraints, hence they must be hosted on priva-
ing, micro cloud, power efficiency, Raspberry PI tized clusters. Beside the power efficiency issues, small
clusters may benefit from small ARM-based devices in
other ways. For instance, utilizing many single-board
1 Introduction computers instead of a few enterprise server may be
cheaper. Furthermore, using many devices allows more
Distributed computing has been an intensively studied fine-grained performance scaling.
topic since the dawn of computer science. The idea of uti-
lizing multiple ordinary devices instead of a single pow- In this paper, we study issues of power efficiency in dis-
erful one brought many advantages, such as much easier tributed systems, clouds, and micro-cloud solutions. We
scaling, possibly higher robustness, or better utilization of have selected the Raspberry Pi single-board computer as a
spare hardware. On the other hand, distributed comput- representative of power efficient hardware based on ARM
ing is encumbered with many challenges that include the platform. We have tested performance of this device using
question of efficiency, communication and synchroniza- our own application benchmark and compare the results
tion overhead, or the necessity of handling failures of in- with a commodity desktop PC and an enterprise server to
dividual nodes. determine the power-to-performance ratio and relative ap-
In combination with modern technologies and hardware plicability for various problems. Even though the results
virtualization, the distributed computing lead to the incep- are only approximate, the Raspberry Pi seems to be a vi-
tion of the cloud phenomenon, where large complex sys- able candidate for green micro-cloud solutions.
tems are presented to users not in a form o a distributed
system, but as virtual hardware, programming platform, The paper is organized as follows. More detailed
or even specialized services. In this form, the user is overview of distributed systems and cloud solutions is pro-
completely shielded from tedious details of system design. vided in Section 2. Section 3 revises related work on
Furthermore, the concept of cloud allows much more effi- micro-cloud systems. In Section 4, we present details
cient allocation of hardware resources, from which bene- about our tested platform – the Raspberry Pi device. Sec-
fits both the cloud providers (since they have less hardware tion 5 summarizes our empirical evaluation, Section 6 out-
to buy and maintain) and cloud customers (who pay only lines possible applicability of these technologies, and Sec-
for resources they really utilize). tion 7 concludes the paper.
18 M. Kruliš, P. Stefan, J. Yaghob, F. Zavoral
2 Distributed Applications and Cloud 3 Micro-Cloud Solutions
Solutions
The recent introduction of the Raspberry Pi, a low-cost,
Among the most important computing technologies that low-power single-board computer, has made the construc-
are in use nowadays are Distributed Systems and Cloud tion of miniature green cloud systems more affordable.
Computing Systems. Distributed system [1] is a collec- Glasgow Raspberry Pi Cloud [5] is a model of a micro-
tion of computers that work together and appear as one cloud solution composed of clusters of Raspberry Pi de-
large computer. These computers cooperate to solve usu- vices. The PiCloud emulates every layer of a cloud stack,
ally complex tasks; they are mutually interconnected to ranging from resource virtualisation to network behaviour,
provide a massive computing power. providing a full-featured cloud computing research and
The basic advantages of distributed systems are: educational environment.
Iridis-pi [6] cluster consists of 64 Raspberry Pi Model
• High performance B nodes each equipped with a 700 MHz ARM processor,
256 Mbit of RAM and a 16 GiB SD card for local storage.
• Transparency The cluster has a number of advantages that are typical
for micro-clouds, such as low total power consumption,
• Resource sharing easy portability due to its small size and weight, and pas-
sive, ambient cooling. These attributes make Iridis-Pi ide-
• Reliability and availability ally suited to educational applications, where it provides
a low-cost starting point to inspire and enable students to
• Incremental extensibility understand principles of high-performance computing.
Sher.ly [7] builds a network-attached storage (NAS) de-
On the other hand, the disadvantages that we may face vice, the Sherlybox, that comes with its own peer-to-peer
in distributed systems are complexity, software develop- virtual private network and file server. The Sherylbox is
ment difficulties, networking problems, and security is- built around the Raspberry Pi Model B computer. It comes
sues. with 512 MB of RAM, two USB 2.0 ports, 802.11n Wi-
Contemporary cloud solutions has evolved from the ear- FI, and a 100mb Ethernet port. Instead of just the naked
lier distributed systems. Cloud computing (despite the board, the Sherylbox comes with a case, a 4GB eMMC
term has no exact definition) can be considered as a spe- flash drive, and an optional 1 TB hard-drive. The com-
cialized form of distributed computing where virtualized pany claims that with external USB drives, it can support
resources are available as a service over the internet. These up to 127 USB drives.
services usually include infrastructure, platform, applica- Tonido [8] offer a compelling alternative to public cloud
tions, storage space and many other vendor-specific mod- file services allowing consumers to leverage their existing
ules, libraries and frameworks. The users pay only for computers or IT infrastructures to keep control over their
the services or resources they actually use. The under- own data. It is available for a wide list of operating sys-
lying resources, such as storage, processors, memory, are tems running on different hardware including Raspberry
completely abstracted from the consumer. The vendor of Pi using Raspbian or Raspbmc OS. Nimbus [9] is another
the cloud service is responsible for the reliability, perfor- example of a micro-cloud solution.
mance, scalability and security of the service. Although all of the abovementioned solutions are in-
Cloud computing has many benefits, but cases exist tended especially to personal or educational use (and a ma-
where some data cannot be moved to the cloud for various jority of scientific papers expect such use-cases), we claim
reasons. In some cases, data may be generated at rates that that, under certain conditions, there may exist a wider
are too big to move or at rates that exceed transfer capacity, range of possible applications. Some of them are dis-
for example in surveillance, operations in remote areas, cussed in Section 6.
and telemetry applications. In other cases, security con-
cerns or regulatory compliance requirements might limit
the use of the cloud. 4 Single-board Computers
Green computing [2] [3] refers to the environmentally
responsible use of computers and any other technology Single-board computers constitute a special brand of com-
related resources. Green computing includes the imple- putational devices which aim for compactness and power
mentation of best practices, such as energy efficiency cen- efficiency. These devices have various applications in
tral processing units (CPUs), peripherals and servers [4]. robotics, intelligent household devices, smart monitoring
Green Cloud is a computing facility that is entirely built, stations, and many other domains. Even though their per-
managed and operated on green computing principles. It formance cannot compete with mainstream desktop PC
provides the same features and capabilities of a typical and servers, they may achieve better power to performance
cloud solution but uses less energy and space, and its de- and power to cost ratios. In this section, we present a few
sign and operation are environmentally friendly. examples of compact single-board devices and revise the
Assessing Applicability of Power-Efficient Embedded Devices for Micro-Cloud Computing 19
properties of Raspberry Pi device, which was selected as quickly adopted for various applications, such as embed-
a representative for our research. ded devices, simple audio and video players, etc.
At present, there are several configurations available
(models A, B, and B+) and a new version called Raspberry
4.1 Computer Examples Pi 2 was introduced to the market. In this work, we present
(and measure) the properties of Raspberry Pi model B+,
Arandale Board [10] is a single-board computer powered
which is the newest revision of the original Raspberry Pi
by Samsung Exynos 5, which is an ARM CPU. The board
(before its second version was released).
is equipped with 2GB of RAM and various common pe-
The device is powered by Broadcom BCM2835 CPU,
ripherals such as USB 3.0, WiFi, GPS module, or inter-
which is an ARMv6 processor clocked at 700 MHz. The
face for LCD display. The board is mainly designed for
graphics is rendered by VideoCore IV GPU clocked at
tablets and embedded computers; however, it also provide
250 MHz. The GPU is capable of decoding a full-HD
adequate performance for cloud computing. On the other
video in real time; however, there is currently no API (such
hand, most of its peripherals are undesired for a solely
as OpenCL) provided for computations. The system holds
computational solution and they may increase the overall
512 MB of DDR2 RAM, which is shared both by CPU and
cost.
GPU. Persistent memory is not integrated on the board, but
AMD presented a Gizmo 2 board [11], which is also
it contains an interface for memory cards. We have used
a single-board computer that is compatible with x86 ar-
commodity 32 GB Kingston MicroSDHC card (class 10)
chitecture. It comprises specialized double core APU
as the persistent data storage.
(clocked at 1 GHz), which is a single chip that integrates
Raspberry Pi has many external interfaces. Beside tradi-
power efficient CPU and Radeon GPU, and 1 GB of DDR3
tional USB or HDMI connector, it also holds custom GPIO
RAM. The board is designed to provide all-in-one PC so-
port or I2C bus, which make the device suitable as a high-
lution, so it is equipped with traditional interfaces such
level controller for many electronic devices. The most
as USB, Gigabit Ethernet, or HDMI. The integrated GPU
important interface for our intentions is the 100 Mb Eth-
may provide excellent performance (with respect to power
ernet. Unfortunately, the Ethernet interface is internally
consumption); however, the price of the board is rather
connected via USB 2.0 bus. This bridged solution does
high in comparison to similar devices.
not reduce the overall throughput, but slightly increases
Intel entered the domain of power efficient single-
the communication latency.
board devices with Galileo development board [12]. It is
The system is designed mainly for Linux operating sys-
equipped with Intel Quark X1000 CPU, which is a single-
tem, but it can accomodate virtually any system that can
core Pentium-based 32-bit processor clocked at 400 MHz,
run on ARM CPU (e.g., RiscOS). For the convenience of
and 256 MB of DRAM. The board is compatible with Ar-
the users, the community has prepared modified distribu-
duino [13] device specification, which allows it to share
tion of Debian Linux called Raspbian and some other dis-
peripherals and extensions designed for this platform.
tributions based on Ubuntu or Fedora are also available.
Another similar platform is Intel Edison. It also con-
We have used the Raspbian in our experiments, since it is
tains Intel Quark CPU, but the Edison platform aims
the recommended system.
mainly at wearable devices and extensive miniaturization.
The Parallela board [14] is a relatively novel accom-
plishment in the field of efficient parallel hardware. Un- 5 Experimental Results
like many other devices, Parallela was designed by a small
company Adapteva. It is equipped with ARM Cortex-9 We have subjected Raspberry Pi to a custom set of perfor-
CPU, FPGA, and a Epiphany coprocessor. The coproces- mance tests to assess its applicability for distributed com-
sor is perhaps the most intriguing part of this hardware, puting and cloud applications. The performance results are
since it is a specialized power-efficient parallel processing compared with results from a desktop PC and commodity
unit which organizes the cores in a 2D grid. This device server in the perspective of the power consumption. Let us
may be the most promising alternative for a Raspberry Pi emphasize that the results measured for Raspberry Pi and
in the terms of power efficiency and total performance. On for full-sized computers are not directly comparable and
the other hand, Parallela is approximately 3× more expen- provide only approximate comparison since our measure-
sive than Raspberry Pi. ments of power consumption does not use same method-
ology and our benchmark is only single-threaded.
4.2 Raspberry Pi
5.1 Experimental Setup
Raspberry Pi [15] is one of the first low-cost devices that is
capable of running a traditional operating system (in this The parameters of Raspberry Pi are detailed in Section 4.2.
case Linux), so it can be used as a modest desktop PC. It The referential desktop PC is equipped with Intel Core
was originally created as a cheap platform that would al- i7 870 CPU, which has four physical (8 logical) cores
low children to learn basics of programming, but it was clocked at 2, 93 GHz, and 16 GB DDR3-1600 RAM. The
20 M. Kruliš, P. Stefan, J. Yaghob, F. Zavoral
persistent storage is represented by two 100 GB SSD disks • hash – Simulation of database hash-join operation us-
connected in RAID 1. ing integer keys.
The referential server is Dell PowerEdge M910. It is
4-way cache-coherent NUMA1 system, where each node • merge – Simulation of database merge-join on sorted
has 8 physical (16 logical) cores clocked at 2 GHz. Each data streams using integer keys.
node manages 32 GB RAM – i.e., the whole system com-
• levenshtein – Wagner-Fischer dynamic programming
prises 64 logical cores and 128 GB of internal memory.
algorithm [19] that computes Levenshtein edit dis-
The server was connected to Infortrend ESDS 3060 disk
tance
array comprising two 400 GB SSD disks and 14 magnetic
disks of 4 TB each. Both desktop PC and server are run- • multiply – Naïve (O(N 3 )) algorithm for matrix multi-
ning Red Hat Enterprise Linux 7 as an operating system. plication on float numbers.
To asses the performance, we measure the real execu-
tion time of prepared tests. All tests are executed on the • strassen – Strassen algorithm for matrix multiplica-
same data inputs and the size of the input is selected so tion on float numbers.
that the test takes reasonable time on Raspberry Pi and at
least a few seconds on desktop PC and server. Each test • quicksort – Quicksort [20] in memory sorting algo-
was repeated 10× and the average time is presented as rithm implemented in C++ std::sort routine ap-
the final result. The values were processed by statistical plied on integers.
methods to remove outliers (times tainted with errors of
measurement). • zlib – DEFLATE [21] compression algorithm imple-
The power consumption of the Raspberry Pi was deter- mented in Zlib.
mined by KCX-017 device, which measures voltage and
current on an USB power cord, since Raspberry Pi is pow- Beside these application tests, we have performed ad-
ered via USB. The power consumption is equal to voltage ditional tests designed to determine the speed of internal
times current (P = UI) and we employ additional correc- memory, effectivity of its CPU caches, and performance
tion factor of 1/0.8, which simulates loss on power source of the persitent storage (i.e., the SD flash card). However,
with efficiency of 80%. The power consumption was be- we do not present detailed results of all these tests for the
tween 1.2 W (idle device) to 1.7 W (performing crypto- sake of the scope.
graphical tests).
The power of our server was measured on its power con- 5.3 Results
troller embedded in server chassis. We also include esti-
mated partial consumption of the chassis itself and addi- The application benchmark results are presented in Fig-
tional equipment (such as cooling infrastructure), hence ure 1. The results depict computational power efficiency
we will operate with aggregated approximate consump- normalized relatively to Raspberry Pi (individually for
tion of 500 W. The power consumption of the desktop PC each algorithm) – i.e., higher value means greater power
was calculated from the component specifications since consumption with respect to computational performance.
we were not able to measure this value with reasonable Hence, we can directly determine, which platform is bet-
effort. For our purposes and intentions, we will operate ter and which is worse for a particular problem. Let us
with the value 250 W. note that we have adjusted the results so that they take
the multi-core and multi-processor nature of the desktop
5.2 Tests PC and the server, since our benchmark is only single
threaded. The performance of the full-sized computers
The performance experiments were design to test various were multiplied by the number of their physical cores.
aspects of the device. Since we are trying to determine The results indicate that Raspberry Pi is quite efficient
applicability of Raspberry Pi as a platform for distributed for memory-intensive tasks. For some tests (especially
system and cloud infrastructure, we have selected algo- database merge joins), the Raspberry Pi even outperforms
rithms that cover many different domains: both desktop PC and server. On the other hand, number
crunching operations (such as the matrix multiplication on
• aes – The Rijndael (Advanced Encryption Standard) float numbers) are more suitable for x86 architecture, since
algorithm [16] for symmetric cryptography. it may employ SIMD instructions. We have performed ad-
• scrypt – Computing scrypt [17] hash function. ditional synthetic memory-oriented experiments and they
have confirmed this observation.
• sha256 – Computing SHA256 hash function. In addition to application tests, we have measured per-
formance of the persistent storage. The throughput of in-
• dijkstra – Finding shortest path in a sparse graph us-
dividual operations is presented in Table 1. Let us empha-
ing Dijkstra algorithm [18] with regular heaps.
size that the Raspberry Pi has only a commodity SD card,
1 Nonuniform Memory Architecture while the server uses enterprise disk array.
Assessing Applicability of Power-Efficient Embedded Devices for Micro-Cloud Computing 21
5 The heat dissipation presents a challenging problem for
relative power consumption / performance
RPi B+
desktop
modern servers as most powerful x86 processors easily
4 server produce over a hundred watts of thermal power. Hence,
the servers, their chassis, and the server racks employ so-
3 phisticated cooling mechanism to drive the undesired heat
out off the server room. In case of smaller devices, the
2 produced heat has much lower watt per area ratio, so it is
much easier to cool these devices.
1
6.1 Replacing Tradional Servers
0
quicksort
scrypt
aes
sha256
hash1
hash2
merge1
merge2
zlib
dijkstra
multiply4
multiply8
strassen4
strassen8
levenshtein
A direct applicability of a Raspberry Pi cluster could be to
replace traditional enterprise servers. Based on the scale,
this solution could work for a small cluster within one
Figure 1: Relative efficiency of application tests server room or as a large distributed system that provides
cloud services. In any case, the main advantage of such
Rasbperry Pi desktop PC server solution is the more evenly distributed heat output. Hence,
the system does not to have a server room with powerful
rand. read 0.7 9.3 5.3 cooling system.
seq. read 17.3 171.9 404.5 It may even be considered to place most of the hard-
seq. write 1.7 58.1 86.9 ware outside of a server room and integrate the single
board computers into the infrastructure of a building or
Table 1: Persistent storage performance (MB/s) into regular rooms (offices, etc.). The Raspberry Pi does
not require a cooling fan, hence such solution would not
increase background noise inside the building. Further-
The results indicate that the performance of the Rasp- more, the heat produced by the devices may be used as
berry Pi is approximately 10-100× worse than the perfor- part of internal heating system and the I/O ports (USB or
mance of other two platforms. On the other hand, if the GPIO) could be used to operate building sensors.
data are distributed evenly among the devices, each Rasp-
berry Pi has to handle two orders of magnitude smaller
amount of data, so the performance is comparable. Fur- 6.2 Outdoor Micro-Clouds
thermore, the devices may also utilize external disk array
connected via 100 Mbit ethernet, which should provide The compactness and low consumption of single-board
data transfers around 5 MB/s. computers may be utilized in many applications which
could be characterized as outside the server room projects.
Such projects would include robotics, autonomous vehi-
6 Applicability cles and aircraft, probes and intelligent exploration de-
vices, etc. A micro-cloud solution could increase robust-
In this section, we would like to outline possible applica- ness of these devices, which could be important since their
bility of single-board devices for various problems. Be- hardware is subjected to much harsh physical conditions
sides the obvious cost issue, the presented solutions are than hardware located in a server room or in an office.
expected to take advantage of two greatest benefits over Let us use an autonomous car (which is a domain that
traditional servers or desktop PCs: spawned an intensive research in the past few years) as
an example of such outdoor device that required nontrivial
• increased robustness computational power. A cluster of single-board comput-
• and better heat dissipation. ers may provide much scalable hardware for navigation
computations. For instance, when the car is driving on a
The robustness is one of the expected properties of straight road in an unpopulated area, it requires much less
many distributed systems. However, when one server fails, computational power to track and analyse surrounding en-
the total drop of performance could be significant, espe- vironment. Hence, it may shut down most of the devices in
cially in case of smaller and mid-sized clusters. When the cluster to save energy. On the other hand, when driv-
small devices such as Raspberry Pi are used, the failure ing inside a city, it may turn on the whole cluster to get
of a single device is hardly noticable on the overall per- necessary computational power. Finally, the decentralized
formance and the faulty hardware could be replaced more nature of the hardware may provide enough computational
quickly. Furthermore, small devices permit more fine- power even in extreme cases, such as when part of the ve-
grained redundancy in the system. hicle is compromised in a car crash.
22 M. Kruliš, P. Stefan, J. Yaghob, F. Zavoral
7 Conclusions [8] Tonido, “Turn your Raspberry PI into your personal
cloud.” [Online]. Available: http://www.tonido.com
In this paper, we have addressed the issue of power ef- [9] Cloudnimbus, “Nimbus - Personal Cloud for Raspberry
ficiency in cloud systems. Many systems would benefit Pi.” [Online]. Available: www.cloudnimbus.org
greatly from a hardware that provide less computational [10] “Arandale board.” [Online]. Available:
power, but which is more power efficient and has lower ini- http://www.arndaleboard.org/
tial and maintenance costs. We have designed an applica- [11] Gizmosphere, “Amd gizmo 2.” [Online]. Available:
tion benchmark for small devices that tests various known http://www.gizmosphere.org/products/gizmo-2/
algorithms. The benchmark was applied on the Raspberry [12] Intel, “Galileo Gen 2 Development Board,
Pi, which is one of the first single-board computers. The url=http://www.intel.com/content/www/us/en/do-it-
Raspberry Pi is very power efficient and cost around $30, yourself/galileo-maker-quark-board.html,.”
which makes it a good candidate to be a worker in a green [13] M. Banzi, D. Cuartielles, T. Igoe, G. Martino,
cluster or a micro cloud. The benchmark results indicate and D. Mellis, “Arduino uno.” [Online]. Available:
that current version of Raspberry Pi is competitive with http://www.arduino.cc/
desktop PC as well as an enterprise server in tasks that can [14] Adapteva, “Parallela.” [Online]. Available:
be idealy distributed. https://www.parallella.org/
In our future work, we would like to test other similar [15] Raspberry Pi Foundation, “Raspberry Pi single-board com-
devices, especially the second version Raspberry Pi and puter.” [Online]. Available: https://www.raspberrypi.org/
the Parallela board with Epiphany coprocessor. Further- [16] Gladman, B.: A specification for rijndael, the aes
more, we are planning to build a small cluster from these algorithm. at fp. gladman. plus. com/cryptogra-
devices to measure the total consumption more precisely phy_technology/rijndael/aes. spec 311 (2001), 18–19
and to determine the communication overhead of various [17] Percival, C.: Stronger key derivation via sequential
distributed algorithms. memory-hard functions. Proceedings BSD Canada, 2009
[18] Dijkstra, E. W.: A note on two problems in connexion with
graphs. Numerische Mathematik 1 (1) (1959), 269–271
Acknowledgment
[19] Wagner, R. A., Fischer, M. J.: The string-to-string correc-
This paper was supported by Czech Science Foundation tion problem. Journal of the ACM (JACM) 21 (1) (1974),
168–173
(GACR) projects P103/13/08195 and P103/14/14292P and
by SVV-2015-260222. [20] Hoare, C. A.: Quicksort. The Computer Journal 5 (1)
(1962), 10–16
[21] Deutsch, L. P.: Deflate compressed data format specifica-
References tion version 1.3, 1996
[1] Tanenbaum, A. S., Van Steen, M.: Distributed systems:
principles and paradigms (2nd Ed.). Prentice-Hall, 2006
[2] Priya, B., Pilli, E., Joshi, R.: A survey on energy and power
consumption models for greener cloud. In: Advance Com-
puting Conference (IACC), 2013 IEEE 3rd International,
2013, 76–82
[3] Kumar, S., Buyya, R.: Green cloud computing and environ-
mental sustainability. In: Harnessing Green It, John Wiley
and Sons, Ltd, UK, 2012, 315–339
[4] Yanovskaya, O., Yanovsky, M., Kharchenko, V.: The con-
cept of green cloud infrastructure based on distributed com-
puting and hardware accelerator within fpga as a service.
In: Design Test Symposium (EWDTS), 2014 East-West,
2014, 1–4
[5] Tso, F. P., White, D. R., Jouet, S., Singer, J., Pezaros, D. P.:
The glasgow raspberry pi cloud: A scale model for cloud
computing infrastructures. In: Distributed Computing Sys-
tems Workshops (ICDCSW), 2013 IEEE 33rd International
Conference on, IEEE, 2013, 108–112
[6] Cox, S. J., Cox, J. T., Boardman, R. P., Johnston, S. J.,
Scott, M., O’Brien, N. S.: Iridis-pi: a low-cost, compact
demonstration cluster. In: Cluster Computing 17 (2014),
349–358
[7] sher.ly, “Private Cloud Solution for Sensitive Data
Sharing with Secure Access Control.” [Online]. Available:
http://sher.ly/