Energy and Cost Considerations for Single Board Computers Usage in

Citizen Science Scenarios

Pedro Verdugo, Joaqu

ın Salvach

ua and Gabriel Huecas

Grupo de Internet de Nueva Generaci

on, DIT, ETSIT, Universidad Polit

ecnica de Madrid, Madrid, Spain

Keywords:

Citizen Science, Green Computing, Energy Efﬁciency, Single Board Computers.

Abstract:

The rich availability of single board computers as an expansion of the traditional embedded system provides a

low cost, easily managed, execution ready and scalable entry-level infrastructure, resulting in a new comput-

ing paradigm termed as Fog Computing. Simultaneously, the current expansion of Citizen Science initiatives

along with the generalization of IOT projects precludes the need for individually managed, low complexity

data processing systems, generating a new user-oriented ecosystem that has come to be deﬁned as the Edge

Cloud. In this document the authors will brieﬂy study the adequacy of the current user-grade SBC hardware

offerings to cover the needs set by the Citizen Science paradigm, starting by studying current Citizen Science

projects in order to deﬁne the aforementioned set of speciﬁc requirements, and subsequently analyzing the

hardware selection and provisioning process given the current non-enterprise user-grade supply for SBC com-

puters, taking special consideration to power usage and total cost of ownership factors from a green computing

perspective.

1 INTRODUCTION

The main goal of Citizen Science is to achieve an

increase of citizen participation and involvement in

scientiﬁc research. This user-based research needs to

be supported by a cost-driven, energy-conscious and

ﬂexible hardware infrastructure. A raise in the pop-

ularity and availability of Single Board Computers

(from now on, SBCs) sets these systems as possible

candidates to host Citizen Science projects. In order

to verify the adequacy of the SBC infrastructure ap-

proach to Citizen Science projects, we will begin by

brieﬂy studying the history and current state of the

Citizen Science paradigm, then we will identify the

required needs for a common Citizen Science project,

and ﬁnally we will study the current SBC hardware

offerings capacity to fulﬁll the discovered needs.

1.1 The Citizen Science Paradigm

Widely understood as the collaboration of volunteers

in the scientiﬁc process, since Silvertowns’ seminal

paper (Silvertown, 2009), Citizen Science has been

proven a valuable method to deal with otherwise

unattainable wide-scoped projects.

The number of samples and observations required

to obtain signiﬁcant results for some areas of knowl-

edge, such as the biology(Kelling et al., 2015) and

biochemistry (Kawrykow et al., 2012) ﬁelds, makes it

impossible to resort to traditional expert data gather-

ing and categorizing techniques, whereas motivated

and loosely trained private individuals can provide

sufﬁcient research data with quality standards com-

parable to those attained by trained experts.

Illustrating the previous point, most common Citi-

zen Science projects to date have been focused on data

collection via Crowdsourcing, deﬁned as the gather-

ing of data and interactions from a wide variety of

users and sources in a distributed fashion(Panchariya

et al., 2015), leaving the processing side to ﬁeld ex-

perts and data scientists. This traditional data chain

for Citizen Science projects (Newman et al., 2012)

sets typical collaborator tasks as data collection and

data transcription whereas typical data scientist tasks

include problem deﬁnition, data cleanup, data analy-

sis, data processing and results dissemination. Taking

into consideration the previous point, it is not strange

that the most commonplace current concern in Cit-

izen Science projects is the increase of the involve-

ment and motivation of the individual volunteer in all

aspects of the scientiﬁc process, widening the collab-

orator’s scope of inﬂuence, interaction and participa-

tion for all the deﬁned stages. This participation can-

not be unbound; to warrant observation quality there

Verdugo, P., Salvachúa, J. and Huecas, G.

Energy and Cost Considerations for Single Board Computers Usage in Citizen Science Scenarios.

In Proceedings of the 7th International Joint Conference on Pervasive and Embedded Computing and Communication Systems (PECCS 2017), pages 35-40

ISBN: 978-989-758-266-0

are available widely studied and deﬁned participation

protocols, as seen in (Sprinks et al., 2017).

In order to set the desired characteristics for a sus-

tainable citizen-driven hardware infrastructure, in the

next section we will summarize the most popular cur-

rent Citizen Science initiatives.

1.2 Current Citizen Science Projects

The Berkeley Open Infrastructure for Network

Computing (BOINC) is a widespread middleware

system for distributed processing, which hosts a va-

riety of projects and enjoying a long lifetime (Ander-

son, 2004). Some of the disadvantages of BOINC in-

clude the need to port code to run in heterogeneous ar-

chitectures, as well as the common dependency man-

agement and user authentication issues.

The BOINC based SETI@home project is the

most long lived of the herein presented (Korpela

et al., 2015), but continues to grow and refresh

its experiment set as shown by the CASPER child

project (Werthimer, 2015), where FPGAs and dedi-

cated GPUs are used to increase processing power.

Also based on the aforementioned BOINC

project, Asteroids@home (

Durech et al., 2015) aims

for the reconstruction of asteroid shapes based on vol-

unteer processing power, and, as commented above,

must include binaries with native support for all com-

mon operating systems.

The Folding@home project has also had a long

running trajectory (Beberg et al., 2009), and presents

a robust infrastructure based on machines with dedi-

cated roles. However, it seems to have decreased in

growth, at least in the academic environment related

to our context of study.

The GROMACS project (Abraham et al., 2015),

is an active and impressive attempt at generating par-

allellizable algorithms for chemistry application in a

local environment. With over 2 million lines of code,

it requires a compilation toolchain for supported plat-

forms.

A mix of available resources from GROMACS

and Folding@home is possible as shown in (Lawrenz

et al., 2015), combining fast local resources with pos-

sibly abundant remote ones.

We can also note the existence of independent

Crowd Computing efforts, as illustrated in (Kovacs

and Lovas, 2014), also commonly based on the

BOINC platform and thus suffering from its limita-

tions.

Once the advantages and shortcomings of the

given projects are known, in the next section we will

abstract from them a set of desirable characteristics

for a common Citizen Science infrastructure.

Table 1: Single Board Systems.

Name Processor Price ($)

ODroid-C1+ Cortex-A5 37

ODroid-U3 Cortex-A9 59

Beaglebone Black Cortex-A8 55

Banana-Pi BPI M2 Cortex-A7 39.50

Banana pi BPI-M3 Cortex-A7 75

Hummingboard-i1 Freescale GC880 70

Minnowboard Max Intel Atom 99

Raspberry Pi 2B Cortex-A7 35

1.3 Detected Citizen Science Hardware

Requirements

The focus of our current work is the generation of

a simple, cheap and ﬂexible enough environment

and infrastructure to allow individuals to set up, use

and obtain immediate results for their own science

projects, while also enabling the sharing of data or

processing power in distributed endeavors. The study

of the initiatives presented in the previous section lets

us detect a series of common deﬁning factors needed

to achieve our goal of a user-level autonomous hard-

ware environment, that will be detailed below:

• Initial Cost: The environment component cost

should be ﬂexible enough to warrant any desired

level of initial user economical environment.

• Efﬁciency: The components should be as less

energy wasteful as possible, following the green

computing paradigm. Thermal considerations and

power usage will be of paramount importance,

and will also reﬂect on the accumulated running

cost.

• Simplicity. The integration of the environment

components should be as simple as possible, not

requiring complex hardware setups or custom

equipment. For that purpose, when given the

choice we will opt for common solutions.

• Flexibility. The system should be as scalable as

required for the user needs from a physical view

as well as price-wise. Related to the previous

point, the system components should be easily in-

terchangeable and of widespread availability.

2 HARDWARE

CONSIDERATIONS

Taking into consideration the resulting factors from

the previous section, the use of single board systems

PEC 2017 - International Conference on Pervasive and Embedded Computing

Table 2: Hardware Speciﬁcations.

Board CPU DMIPS RAM Net Power

Name Family Cores GHz Single Total Mb Mhz Mbps W-Full W-Idle

ODroid-C1+ Cortex-A5 4 1.5 1570 6280 1024 900 1000 2.31 0.73

ODroid-U3 Cortex-A9 4 1.7 1780 7120 2048 900 1000 4.25 1.75

Beaglebone Black Cortex-A8 1 1.0 2000 2000 512 600 100 2.44 0.42

Banana-Pi BPI M2 Cortex-A7 2 1.0 1270 2540 1024 432 100 2.01 0.43

Banana pi BPI-M3 Cortex-A7 4 2.0 1270 5080 2048 480 100 2.35 1.32

Raspberry Pi 2B Cortex-A7 2 0.9 1180 2360 1024 900 100 2.25 1.15

is obviously a reasonable choice to increase the sim-

plicity and ﬂexibility of our setup, as shown on the

excellent performance comparison given in (Lencse

and R

as, 2015). In Table 1 we will update the men-

tioned table to detail some of the current user grade

SBC offerings.

Table 1 moves us to consider the idea of using

ARM (Advanced RISC Machine) based architecture,

which has recently been proven to be a cost-oriented

viable alternative to the traditional x86 architecture as

shown in Keipert (Keipert et al., 2015). These ideas

have already been successfully taken to practice in the

OLPC project as illustrated in Gahire (Gaihre, 2015).

Restricting our selection to ARM architectures, we

can now proceed to compare individual board charac-

teristics, always taking into consideration the restric-

tions for our particular goal.

2.1 CPU Comparison

For starters, table 2 details the number of cores and

work frequencies along with the different ARM fam-

ilies for each of our SBCs of interest.

These different values assert the need to compare

the CPU performance for each board across different

processor families and frequencies. For that purpose,

we will use the Dhrystone tool, which lets us obtain an

standardized MIPS (Millions of Instructions Per Sec-

ond) value independently from the processor family

or architecture, allowing simpler inter-system com-

parison.

The results shown in table 2 are mostly self-

explanatory, but we can note how the Cortex-A9 sys-

tem seems to have the better score at raw processing,

surprisingly closely followed by the older A5 archi-

tecture.

2.2 RAM Comparison

RAM size and I/O throughput is of high importance

in data-oriented environments. This may probably be

the biggest ﬂaw of the current board offerings. As OS

and applications will have to share the meager mem-

ory space, there will constantly be a lack of fast access

RAM, and maybe pagination issues.

Table 2 shows how the Odroid-U3 system seems

to be the best technical offering, noting the use of

LPDDR3 technology, a standard in mobile applica-

tions.

2.3 Network Comparison

Given the common use cases for our system we can

expect very high network I/O throughput between lo-

cal machines and also with remote systems, making

cabled ethernet a desirable item.

Table 2 shows that the ODroid family takes the

lead here with an integrated 1Gbps ethernet NIC.

2.4 Storage Comparison

In table 3 we will very brieﬂy consider the storage ca-

pacities of the different boards, with the parameter c

corresponding to the class of the used microSD card.

This class corresponds directly to the speedup over

a reference 1MBps transfer speed, and given the cur-

rent exponential cost of higher than 10-class microSD

cards will be the main practical limitation to our data

transfer capabilities.

Table 3: Storage Speciﬁcations (Mbps).

Name Interface Read Write

ODroid-C1+ microSD c c

- eMMC 250 90

ODroid-U3 microSD c c

- eMMC 250 90

Beaglebone Black microSD c c

Banana-Pi BPI M2 SD c c

- SATAII 300 300

Banana pi BPI-M3 SD c c

- SATAII 300 300

Raspberry Pi 2B microSD c c

There is also room to notice the eMMC interface

provided by the ODroid family, again inherited from

the mobile industry and comparable in transfer speeds

Energy and Cost Considerations for Single Board Computers Usage in Citizen Science Scenarios

ODroid C1

ODroid U3

BeagleBone

Banana M2

Banana M3

Raspberry 2B

0.000

0.005

0.010

0.015

0.020

0.025

0.030

0.035

DPJ (Mb/J)

ODroid C1

ODroid U3

BeagleBone

Banana M2

Banana M3

Raspberry 2B

0.00

0.01

0.02

0.03

0.04

0.05

0.06

DPS (Mb/s)

ODroid C1

ODroid U3

BeagleBone

Banana M2

Banana M3

Raspberry 2B

50000

100000

150000

200000

250000

EDP (s/Mb)

Figure 1: DPJ, DPS and EDP.

to SSD technology. The high cost and dedicated na-

ture of such a solution, precludes us from further con-

sideration.

3 POWER CONSIDERATIONS

The Thermal Design Power (TDP), or thermal de-

sign point, has been historically used to generate

power usage metrics(Hennessy and Patterson, 2012).

Being deﬁned as the maximum amount of heat gener-

ated during typical computer operation (Gough et al.,

2015), this concept results inadequate as a measure of

a full system processing power.

The work of Hennessy clearly states that any use-

ful metric for computer power must necessarily be

tied to energy usage, and introduces the following

terms:

• Data Processed Per Second (DPS): data pro-

cessed by the system in a given time (in our case,

one second).

• Data Processed Per Joule (DPJ): amount of data

processable by the system with a given energy

budget of 1 Joule.

• Energy-Delay Product (EDP): As introduced by

Horowitz(Horowitz et al., 1994) in the transistor

performance environment, and adapted by Laros

III(Laros III et al., 2013), deﬁnes the time taken

by the system to output a given amount of data for

a ﬁxed energy budget.

Thus, from a thermal and power usage standpoint, we

can refer to the equations provided by Malik (Malik

and Homayoun, 2015), used to transform the power

values in table 2 into the more useful metrics given in

Figure 1.

In Figure 1 we can appreciate how the ODroid

family has the highest DPJ rate, mainly due to the

high core speeds. These speeds, along with the ele-

vated MIPS previously remarked, are also responsible

for the DPS measurement. The EDP graph from Fig-

ure 1 shows us the price to pay for both previous high

ﬁgures: an elevated EDP signifying that the ODroid

family is not as energy-efﬁcient as its lower-power

counterparts.

4 COST CONSIDERATIONS

The most common tool for complete cost calculation

is the Total Cost of Ownership (TCO) metric, as pre-

sented in Martens (Martens et al., 2012). For the sys-

tem under study, we will reﬁne the previous formulas

with multinode extended considerations as developed

by Barroso (Barroso et al., 2013).

To begin with, the terms of interest will be de-

ﬁned:

• n: Number of boards (7)

• t: Runtime (5 years)

• C

: Provisioning Cost per node ($37)

• C

: Total Electricity Cost per node ($ To be de-

termined)

• C

: Electricity Cost per hour (0.11 KWh)

• P

: Full Power Usage per node (2.3W)

• P

: Idle Power Usage per node (0.73 W)

• U : Usage Factor (0.15% (default) / 0.95% (high))

The main cost formula is obtainable by means of the

following equation:

TCO =

∑

i=1

(Cpi ∗Cei) (1)

Where C

can be expanded as follows:

Ce = t ∗Ch ∗ (U ∗ P f + (1 −U ) ∗ Pi) (2)

PEC 2017 - International Conference on Pervasive and Embedded Computing

0 1 2 3 4 5

Years

100

150

200

250

300

TCO ($) U=0.15

Figure 2: Accumulated Results per Year for Regular Usage

Factor.

0 1 2 3 4 5

Years

100

150

200

250

300

350

TCO ($) U=0.95

Figure 3: Accumulated Results per Year for High Usage

Factor.

The results obtained from applying the given val-

ues to the previous formulas can be seen in ﬁgures

2 and 3, and show that even for a low number of

boards and very high usage factors the provisioning

costs (red) are constantly higher than the yearly accu-

mulated costs (green), indicating that for the present

selection a 100% full time use would still be proﬁtable

in the long run.

5 CONCLUSIONS AND FUTURE

WORK

Based on all the previous points, it seems clear that

the ODroid family provides superior CPU and net-

work performance, with medium RAM and storage

technology. Between both studied offerings, the U3

board is undoubtedly the best technical choice from

a mere energy-based standpoint. Nonetheless, taking

into consideration the initial provisioning cost objec-

tives, the C1+, at half the cost of the U3 board, covers

all our requirements in a more reasonable fashion.

In the authors’ consideration the system choice as

presented accomplishes the goals set in the introduc-

tion, as detailed below:

• Cost Oriented: Technical preferences are given a

secondary plane in order to focus on ﬂexible user

budgets.

• Simple: Minimal software/hardware knowledge

is needed to deploy the infrastructure. As a ﬁnal

user, and depending on the project of choice, no

previous expertise is required.

• Scalable: Growth complexities are reduced by

ensuring the use of lightweight infrastructure

choices.

• Efﬁcient: As shown in the previous power and

cost studies, the proposed hardware infrastructure

warrants near-optimal resource usage for a mini-

mal initial inversion. For high workload scenar-

ios, efﬁciency increases.

A note for future deployment could be that even if

the current software offerings simplify the develop-

ment of collaborative infrastructures, the worldwide

expansion of Citizen Science still needs a more com-

monplace point of entry for the casual private user.

Mobile platforms seem to be optimally placed for that

role(Panchariya et al., 2015), and as such, should be

given the utmost importance when considering the

given infrastructure interactions.

To summarize the innovations introduced by the

presented proposal:

• It adopts an inclusive view from several seem-

ingly distant state of the art knowledge ﬁelds,

as current Cloud Computing offerings are given

the main focus in industry and academic-oriented

works(Penzel et al., 2015)

• It subsequently reasons that the main point of in-

terest for the individual Citizen Science user can

be more clearly focused in IOT, Edge Computing

initiatives(Kido and Swan, 2014) providing a di-

rect application and immediate results.

• Finally, it expresses an integral solution oriented

to give the individual enthusiast, volunteer or ca-

sual user a simple opportunity to take active par-

ticipation in the scientiﬁc process with very low

entry-point costs.

REFERENCES

Abraham, M. J., Murtola, T., Schulz, R., P

all, S., Smith,

J. C., Hess, B., and Lindahl, E. (2015). GROMACS:

High performance molecular simulations through

multi-level parallelism from laptops to supercomput-

ers. SoftwareX, 1:19–25.

Anderson, D. P. (2004). Boinc: A system for public-

resource computing and storage. In Grid Comput-

Energy and Cost Considerations for Single Board Computers Usage in Citizen Science Scenarios

ing, 2004. Proceedings. Fifth IEEE/ACM Interna-

tional Workshop on, pages 4–10. IEEE.

Barroso, L. A., Clidaras, J., and H

olzle, U. (2013). The dat-

acenter as a computer: An introduction to the design

of warehouse-scale machines. Synthesis lectures on

computer architecture, 8(3):1–154. 01105.

Beberg, A. L., Ensign, D. L., Jayachandran, G., Khaliq, S.,

and Pande, V. S. (2009). Folding@home: Lessons

from eight years of volunteer distributed computing.

In IEEE International Symposium on Parallel Dis-

tributed Processing, 2009. IPDPS 2009, pages 1–8.

Durech, J., Hanu

s, J., and Van

co, R. (2015). Asteroids@

home—A BOINC distributed computing project for

asteroid shape reconstruction. Astronomy and Com-

puting, 13:80–84.

Gaihre, K. P. (2015). ARM Based Computing Technol-

ogy for Sustainable Development (Performance Anal-

ysis on E-Learning). In Proceedings of IOE Graduate

Conference, pages 298–304.

Gough, C., Steiner, I., and Saunders, W. (2015). CPU Power

Management. In Energy Efﬁcient Servers, pages 21–

70. Apress.

Hennessy, J. L. and Patterson, D. A. (2012). Computer Ar-

chitecture: A Quantitative Approach. Elsevier.

Horowitz, M., Indermaur, T., and Gonzalez, R. (1994).

Low-power digital design. In Low Power Electronics,

1994. Digest of Technical Papers., IEEE Symposium,

pages 8–11. IEEE. 00362.

Kawrykow, A., Roumanis, G., Kam, A., Kwak, D., Leung,

C., Wu, C., Zarour, E., Sarmenta, L., Blanchette, M.,

Waldisp

uhl, J., and others (2012). Phylo: a citizen sci-

ence approach for improving multiple sequence align-

ment. PloS one, 7(3):e31362.

Keipert, K., Mitra, G., Sunriyal, V., Leang, S. S., Sosonk-

ina, M., Rendell, A. P., and Gordon, M. S. (2015).

Energy-Efﬁcient Computational Chemistry: Compar-

ison of x86 and ARM Systems. Journal of chemical

theory and computation, 11(11):5055–5061. 00001.

Kelling, S., Fink, D., La Sorte, F. A., Johnston, A., Bruns,

N. E., and Hochachka, W. M. (2015). Taking a ‘Big

Data’approach to data quality in a citizen science

project. Ambio, 44(4):601–611.

Kido, T. and Swan, M. (2014). Know thyself: Data

driven self-awareness for understanding our uncon-

scious behaviors-Citizen Science Genetics for sleep

and wellness research.

Korpela, E. J., Siemion, A. P., Werthimer, D., Lebofsky,

M., Cobb, J., Croft, S., and Anderson, D. (2015). The

next phases of SETI@ home. In SPIE Optical Engi-

neering+ Applications, pages 96060B–96060B. Inter-

national Society for Optics and Photonics.

Kovacs, J. and Lovas, R. (2014). Crowd computing to

support EGI scientists. In 2014 37th International

Convention on Information and Communication Tech-

nology, Electronics and Microelectronics (MIPRO),

pages 325–331.

Laros III, J. H., Pedretti, K., Kelly, S. M., Shu, W., Fer-

reira, K., Vandyke, J., and Vaughan, C. (2013). Energy

delay product. In Energy-Efﬁcient High Performance

Computing, pages 51–55. Springer. 00005.

Lawrenz, M., Shukla, D., and Pande, V. S. (2015). Cloud

computing approaches for prediction of ligand bind-

ing poses and pathways. Scientiﬁc reports, 5.

Lencse, G. and R

as, S. (2015). Method for benchmarking

single board computers for building a mini supercom-

puter for simulation of telecommunication systems.

In Telecommunications and Signal Processing (TSP),

2015 38th International Conference on, pages 246–

251. IEEE.

Malik, M. and Homayoun, H. (2015). Big data on low

power cores: Are low power embedded processors a

good ﬁt for the big data workloads? In Computer

Design (ICCD), 2015 33rd IEEE International Con-

ference on, pages 379–382. IEEE. 00001.

Martens, B., Walterbusch, M., and Teuteberg, F. (2012).

Costing of cloud computing services: A total cost

of ownership approach. In System Science (HICSS),

2012 45th Hawaii International Conference on, pages

1563–1572. IEEE. 00050.

Newman, G., Wiggins, A., Crall, A., Graham, E., Newman,

S., and Crowston, K. (2012). The future of citizen sci-

ence: emerging technologies and shifting paradigms.

Frontiers in Ecology and the Environment, 10(6):298–

304.

Panchariya, N. S., DeStefano, A. J., Nimbagal, V., Ragu-

pathy, R., Yavuz, S., Herbert, K. G., Hill, E., and

Fails, J. A. (2015). Current Developments in Big Data

and Sustainability Sciences in Mobile Citizen Science

Applications. In 2015 IEEE First International Con-

ference on Big Data Computing Service and Applica-

tions (BigDataService), pages 202–212.

Penzel, D., Kryvinska, N., Strauss, C., and Gregu, M.

(2015). The Future of Cloud Computing: A SWOT

Analysis and Predictions of Development. In 2015 3rd

International Conference on Future Internet of Things

and Cloud (FiCloud), pages 391–397.

Silvertown, J. (2009). A new dawn for citizen science.

Trends in ecology & evolution, 24(9):467–471.

Sprinks, J., Wardlaw, J., Houghton, R., Bamford, S., and

Morley, J. (2017). Task Workﬂow Design and its im-

pact on performance and volunteers’ subjective pref-

erence in Virtual Citizen Science. International Jour-

nal of Human-Computer Studies, 104:50–63.

Werthimer, D. (2015). What’s new in SETI and CASPER?

In Radio Science Meeting (Joint with AP-S Sympo-

sium), 2015 USNC-URSI, pages 293–293.

PEC 2017 - International Conference on Pervasive and Embedded Computing