Design T ade-o s o C owdsou ced Web Access
in Communi y Ne wo ks
Emmanouil Dimoge on akis∗, Roc Mesegue ∗, Leand o Na a o∗, Se gio Ochoa†and Luís Veiga‡
∗Uni e si a Poli ècnica de Ca alunya, Ba celona, Spain
{edimoge ,jlne o,mesegue ,leand o}@ac.upc.edu
†Uni e sidad de Chile, San iago, Chile
[email p o ec ed]
‡Tecnico Lisboa/INESC-ID Lisboa, Lisboa, Po ugal
luis. [email protected]
Abs ac —In e ne access has become a equi emen o pa ic-
ipa e in socie y; howe e , he majo i y o he wo ld’s popula ion
is no ye online. Ci izens can sel -o ganize coope a i ely o
c owdsou ce communi y ne wo k in as uc u es and achie e
In e ne access. In o de o help add ess ha challenge, his
pape p o ides an analysis o a c owdsou ced In e ne access
mechanism: he dis ibu ed Web p oxy se ice in one o he
la ges communi y ne wo ks in he wo ld. Se e al pe spec i es
we e conside ed in his analysis, e.g., da a a ic, ne wo king
issues, and p oxies esponsi eness. The e alua ion esul s show
how he cu en manual p oxy choice, based on social clues,
becomes a popula se ice plagued wi h ho spo s and ine i-
ciencies, which opens se e al oppo uni ies o imp o ing hese
in as uc u es. By aking ad an age o i , ou esea ch shows ha
he ade-o s be ween in o med p oxy selec ion and admission
con ol in p oxies, could alle ia e imbalances and unce ain y,
and also imp o e he se ice wi h li le addi ional bu den. This
ep esen s an explici and di ec mechanism o imp o ing he
se ice p o ided by hese communi y ne wo ks, and a clea
bene i o i s membe s.
I. INTRODUCTION
In e ne access has become a equi emen o pa icipa e
in socie y; o ins ance, o access public se ices, educa ion
ma e ial, social media and also o suppo e e yday wo k
o millions o o ganiza ions. Howe e , he majo i y o he
wo ld’s popula ion is no online [1] ye , a om he ision o
“uni e sal se ice”. This si ua ion agg a a es he digi al di ide
be ween se e al communi ies/ egions/coun ies, and he es o
he wo ld. The cos and a ailabili y o his se ice seem o be
he main limi a ion o becoming In e ne access a gene ally
a ailable se ice in ou socie y. Ne wo k in as uc u es ha
p o ide hese se ices a e, in mos cases, unde con ol o
o me monopolies, now elecom incumben s. These en i ies
con ol he o e and ha e he s eng h o in luence egula ion
and discou age compe i o s. Excep in de eloped u ban a eas,
he ypical si ua ion is a lack o compe i i e o e s, de ined
as “ma ke ailu e” [2]. This nega i ely a ec s people by
educing hei capabili y o imp o ing hei digi al li e acy
and inc easing he digi al di ide wi h hem. Ru al and poo
communi ies, and also age b acke s (like he olde adul s) a e
pa icula ly ulne able o his si ua ion.
As a way o mi iga e his challenge, in many egions
wo ldwide he ci izens sel -o ganize o explo e al e na i e
models o ge ing In e ne access unde easonable condi ions.
An example o i a e he communi y ne wo ks (CN), ha a e
c owdsou ced da a ne wo k in as uc u es buil by ci izens and
o ganisa ions, who pool hei esou ces and coo dina e hei
e o s [3] o p o ide an In e ne communi y access se ice o
hei membe s. These communi ies a e open, ee, and neu al.
They a e open since e e yone has he igh o know how hey
a e buil . They a e ee because he ne wo k access is d i en
by he non-disc imina o y p inciple; hus, hey a e uni e sal.
Mo eo e , hey a e neu al in e ms o echnology solu ions o
ex end he ne wo k, and neu al o suppo ing da a ans e s.
The communi y ne wo ks a e qui e new, and hey ep esen
an al e na i e pa adigm o de eloping ne wo k in as uc u es
and se ices in a b oad sense. Communi ies can p opose
locally adap ed sel -o ganized coope a i e schemes o de el-
oping sel -p o ided da a ne wo king solu ions, sha ing wi eless
links and spec um, op ical ib e, and In e ne ga eways; and
e en sha ing In e ne connec i i y wi h o he membe s o he
communi y. When hese undamen al p inciples a e applied
o an in as uc u e, hey o en esul in ne wo ks ha a e
collec i e goods,socially p oduced, and go e ned as common-
pool esou ces (CPR). Na u al CPR, also called commons
(such as, communal pas u es, ishe ies, o es s), we e s udied
in dep h by E. Os om [4]. Acco ding o ha we use he e m
ne wo k in as uc u e commons [2]. These in as uc u es de-
eloped coope a i ely become egional IP ne wo ks ha enable
inexpensi e in e ac ion and access o local digi al con en and
se ices. In addi ion, he e exis s he issue o access o he
global In e ne , ha can be eached h ough In e ne Se ice
P o ide s (ISP) in hese egional ne wo k in as uc u es.
The e a e many examples o communi y ne wo ks ha
can i in his scheme. In [5] we ou line 18 cases, wi h 9
desc ibed in de ail, and 267 po en ial cases in 41 coun ies.
The e a e also se e al s udies ha conside s uc u al [6], [7],
[8], echnological [9], [10], [11] and o ganisa ional [3], [12],
[5] poin s o iew o hese ne wo ks.
In his pape we ocus ou s udy on he gui i.ne com-
muni y ne wo k, one o he la ges wo ldwide. Pa icula ly
we analyze c owdsou ced p o ision o In e ne access using
a pool o sha ed Web/In e ne p oxies, dis ibu ed o e nodes
in a egional ne wo k. This is an inclusi e and cos e ec i e
model o p o ide limi ed In e ne (Web) access, complemen-
a y o comme cial o e ings. Howe e , c owdsou cing equi es
mo i a ion and incen i es o he pa icipan s, egula ion o
con ibu ions and consump ion o achie e ope a ional and
sus ainable ou comes, such as in [13].
The con ibu ion o his pape is he analysis o he eal-
wo ld dis ibu ed In e ne access se ice, implemen ed as a se
o Web p oxy se e s in he gui i.ne communi y ne wo k. This
analysis allows us o iden i y po en ial weaknesses ha limi
he se ice p o ision, i s quali y and he way in which he com-
muni y e ol es. This s udy conside s se e al da a inpu s; e.g.,
he pa e ns o usage om se ice logs, he design choices and
implica ions (conside ing clien and p oxy choices) acco ding
o pa e ns o usage, and he ela i e loca ion o use s and
p oxies in he ne wo k opology. The esul s show he key
me ics, he design space o coope a i e choices, he in ol ed
ade-o s, and he e ec s on he se ice cos and pe o mance.
Nex , we i s in oduce he p oxy se ice in he gui i.ne
CN. Sec ion III looks a he beha io and clus e ing o use s
acco ding o con en and ne wo k locali y, and i also analyzes
he impac on he c i e ia o p oxy selec ion. We p esen an
analysis o he cu en scena io, limi a ions and po en ial o
imp o emen om he pe spec i e o he access ne wo k in
Sec ion IV, p oxies in Sec ion V and use s in Sec ion VI.
Sec ion VII p esen s he conclusions and he u u e wo k.
II. THE WEB THROUGH A PROXY SERVICE
Global access o In e ne o e e ybody equi es no only o
inc ease he se ice a ailabili y, bu also a d ama ic educ ion
o i s cos , especially in geog aphies and popula ions wi h low
pene a ion [14]. This cos educ ion can be achie ed by sha -
ing; e.g., a la ge popula ion o C clien s can b owse he web
aking ad an age o he agg ega ed capaci y o a pool o P web
p oxies, wi h C P, o e a egional ne wo k in as uc u e,
a a ac ion o he cos o C In e ne connec ions.
The p o ide s o hese In e ne connec ions can be qui e di-
e se, such as comme cial ISP, coope a i es o associa ions o
use s sha ing cos s [15], con en o se ice p o ide s p omo ing
hei o e [16], ci izens sha ing hei unused In e ne access
capaci y wi h neighbo s and iends [17], public o p i a e
o ganiza ions sponso ing In e ne access o complimen a y
in e es s. Howe e , elecom egula ion au ho i ies in many
coun ies limi publicly subsidized In e ne access o p ese e
ma ke compe i ion. In e ne access h ough web p oxies is
clea ly a limi ed se ice compa ed o an IP unnel, as he
se ice is usually es ic ed o a se o p o ocols/po s; howe e ,
i can help enhance p i acy as he o igin IP add esses may be
hidden. The mos popula applica ion in communi y ne wo ks
is Web access. Many ci izens, p i a e and public o ganiza ions
in ol ed in communi y ne wo ks, such as ei unk.ne o
gui i.ne , ha e chosen o p o ide ha se ice wi hin hei com-
muni y ne wo k. Using Web p oxies h ough local ne wo king
in as uc u es (e.g., communi y ne wo ks) ha p o ide local
o egional connec i i y, he ci izens can each In e ne con en
and se ices a no addi ional cos .
To unde s and he impac o hese p oxies in he beha io
o hese ne wo ks, we ocus ou s udy on he dis ibu ed web
p oxy se ice o he gui i.ne communi y ne wo k; a ee,
neu al and open access CN wi h mo e han 32,000 nodes
mos ly in Spain [9]. The se ice has mo e han 300 Web p oxy
se e s, howe e his s udy is based on measu emen s o 30
days on a ne wo k sub-zone ha consis s o 4 p oxies sha ed
among mo e han 500 use s.
Wi hou access o one o hese p oxies o a gui i.ne
connec ed ISP, communi y membe s can s ill sha e con en s
and access applica ions wi hin he same communi y ne wo k,
bu no o ex e nal esou ces. In o de o ge Web access, he
clien s manually speci y a lis o p oxies, he main p oxy and
he seconda y ones. Access con ol o he p oxy is pe o med
h ough ede a ed au hen ica ion c eden ials. In case a p oxy
does no espond ( imeou ) o ejec s he connec ion, he clien
au oma ically swi ches o he nex p oxy in he lis . The
choice o p oxies is manual and he lis usually comes om
acquain ances in he communi y o pe sonal expe ience.
By conside ing he a ailable in o ma ion (anonymized
p oxy log iles, and opology and link da ase s o he gui i.ne
communi y1), he nex sec ion p esen s he me hodology used
o iden i y pa e ns o usage o he Web p oxy se ice, which
could be le e aged o imp o e such a se ice.
III. CLUSTERING OF USERS
We s a ed he s udy explo ing da a conce ning he se ice
usage, in o de o g oup use s acco ding o hei beha io .
Then, we iden i ied he g aph communi ies ha exis in he
ne wo k o analyze he ac o o ne wo k locali y. Thus, we
ied o unde s and he ade-o s o g ouping use s acco ding
o simila i ies in hei beha io and/o acco ding o hei
loca ion in he ne wo k.
A. Clus e ing acco ding o usage
Fo he analysis o se ice usage acco ding o pa e ns o
da a a ic, and based on [18], we conside ed h ee di e en
ypes o clus e ing algo i hms: K-means, sui able o gene ic
applica ions, DBSCAN and Wa d’s hie a chical clus e ing
(HC) ha can ace complex pa e ns. The inpu used by he
algo i hms was he o al da a ans e ed pe use in by es,
as well as he co esponding amoun o a ic o con en s
ha cons i u e a la ge amoun o he o al se ice a ic, like
ideo (20%), image (6%) and HTML (2%). We expe imen ed
wi h a ious clus e sizes o K-means and Wa d’s HC, includ-
ing well-known empi ical es ima ion me hods like he ’elbow
me hod’, as well as many pa ame e s o DBSCAN. Table I
p esen s he op imal esul s o each me hod in e ms o clus e
alida ion. Fo he alida ion we used he coe icien Shiloue e
sco e ha has alues in he [-1,1] ange. As desc ibed in
Table I, o all he cases he e is a big clus e o 450-480
use s wi h a Shiloue e sco e o 0.9, indica ing a e y s ong
clus e densi y. None heless, he es o he use s belong o
o e lapping clus e s, wi h sco es close o 0. A e manually
e iewing o he esul s o he algo i hms o ge ing a be e
insigh , since i is he s anda d p ocess in hese cases, we chose
Wa d’s HC me hod wi h 3 clus e s, ha pa i ions he use s in
one la ge consis en clus e and wo small o e lapping clus e s,
minimizing hus o e lapping elemen s.
Table II p esen s he cha ac e is ics o he clus e s, as
o med using Wa d’s HC o 3 clus e s. We ind wo consis en
clus e s o use s wi h dis inc p ope ies. The Figu e 1 depic s
he compa ison o he clus e s in e ms o a ic and size
(numbe o use s). Clus e 1 o ligh use s, includes he
majo i y o use s and hei p o ile consis s o gene a ing e y
low a ic, as low as 1% o he maximum no iced pe use
1Da ase s a ailable a : h p://dsg.ac.upc.edu/p oxy-gui i
Table I. RESULTS FROM CLUSTERING ALGORITHMS ON USAGE
Me hod Clus e s # Clus e s Size Clus e s Shiloue es
DBSCAN 2 7, 499 0.03, 0.90
2 33, 473 -0.04, 0.89
Wa d’s 3 4, 29, 473 0.30, 0.01, 0.87
2 21, 485 0.09, 0.89
K-Means 3 10, 44, 452 0.07, -0.04, 0.88
Table II. USERS BEHAVIOR CLUSTERS DESCRIPTION (WARD’S)
ID Size Shiloue e Cha ac e is ics Alias
1 473 0.87 Low o al a ic Ligh
2 29 0.01 Medium o al and ideo/images
a ic
Medium
3 4 0.30 High o al and ideo/images
a ic
Hea y
alue, mos ly HTML b owsing. Clus e 3, hea y use s, consis s
o only 4 use s and i is cha ac e ized by high o al a ic,
whe e mos o i is spen on downloading ideo and images.
Clus e 2, medium use s, p esen s an in e media e beha io ;
ne e heless, ollowing he pa e ns o he hea y use s. Medium
use s c ea e a signi ican po ion o he o al a ic, a ound
20% o he maximum alue, which hey consume mos ly on
ideos and images. This clus e has low consis ency, wi h use s
p esen ing a beha io simila o clus e 3, bu wi h a ic le el
close o clus e 1.
Clus e 1 Clus e 2 Clus e 3
0.0
0.2
0.4
0.6
0.8
1.0
Clus e T a icPe cen age
Clus e Use sPe cen age
Figu e 1. T a ic and Use s Pe cen age pe Clus e
B. Clus e ing acco ding o Ne wo k Locali y
Fo he analysis o use g oups acco ding o ne wo k
locali y, we use g aph communi y de ec ion echniques. Based
on [19], we choose h ee o he mos p ominen de ec ion
algo i hms: Spinglass, Mul ile el and In omap. The da a inpu
o he algo i hms is he backbone g aph, consis ing o 48
nodes. Mo eo e , since he s udied gui i.ne zone has a small
well-connec ed backbone, wi h many clien s connec ed o he
ou e s o he backbone, we used he numbe o clien s using
hose ou e s o es ablish he g aph weigh o he In oMap
algo i hm. The weigh o each link is de ined as he a e age
ime o ans e a single by e acco ding o ou opology da ase .
The esul s o he di e en algo i hms can be seen in Table III.
We compa e he algo i hms using he modula i y sco e, which
lies in he ange [-1/2, 1), whe e he highe he alue, he mo e
consis en he communi y. Expe imen ing wi h he algo i hms
we no iced ha he node size a gumen o he In omap does no
a ec signi ican ly he ou pu , hus In omap does no o e any
addi ional in o ma ion. The e o e, we choose he Mul i-le el
Algo i hm ha has he highes modula i y sco e and smalle
numbe o clus e s, conside ing he small backbone.
Table III. COMPARISON OF COMMUNITY DETECTION ALGORITHMS
In omap Mul ile el Spinglass
Modula i y 0.699 0.712 0.702
Clus e s 12 9 15
Figu e 2 shows he esul ing g aph o he Mul i-le el
algo i hm. The squa es ep esen he ou e s ha ope a e also
as p oxies. As depic ed, he p oxies a e no well posi ioned
ela i ely o he ne wo k clus e s, conside ing ha mos o
he clus e s ha e no p oxies, while one o he clus e s has
wo p oxies. Addi ionally, we obse e ha he e a e clus e s
poo ly connec ed o hei neighbou ing clus e s, esul ing in
an in as uc u e a om ideal. Fo he es o his wo k we
assume ha all he clien s o a ou e belong o he clus e o
ha ou e .
Figu e 2. Mul i-le el Communi y De ec ion o he backbone ne wo k (colo s)
C. In luence o he c i e ia o p oxy selec ion
Acco ding o ou clus e ing analysis, we p esen simula-
ions ha exploi he wo clus e ing echniques in algo i hms
o p oxy selec ion, in o de o p o ide al e na i es o he
cu en manual p oxy selec ion. The objec i e is o demons a e
he impac o ne wo k locali y and use a ic beha io on he
pe o mance o he p oxy se ice and use expe ience. Thus,
we show how hey can be used o in o m he design o an
imp o ed se ice.
Nex , we p esen an ini ial e alua ion o he men ioned
echniques unde he pe spec i es o he ne wo k, he p oxies
and he use s. I is impo an o cla i y ha ou algo i hms
implemen one o se e al ways o use he in o ma ion om
use beha io clus e ing and communi y de ec ion. The i s
algo i hm we implemen ed, e e ed as da a_clus e , uses he
clus e ing o use beha io o assign equi alen use load o
each p oxy by equally dis ibu ing he use s o each clus e . In
he cases whe e a new use has o be assigned o a p oxy and
all exis ing assignmen s om he clus e s a e equally balanced,
he algo i hm selec s a p oxy andomly. The second algo i hm,
1071081091010 1011
To alBy esPe Link
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
ECDF
manual
da a_clus e
ne wo k_clus e
da a+ne wo k
Figu e 3. Compa ison o To al Links By es Pe S a egy ECDF
e e ed as ne wo k_clus e , uses g aph communi y de ec ion
o assign use s o p oxies acco ding o he p oximi y o hei
communi y. Fo ins ance, a use wi h an a ailable p oxy
in his communi y will be assigned o his p oxy, while in
he opposi e case, i will be assigned o he p oxy ha is
loca ed in he closes communi y. In case o equal p oximi y,
he p oxy selec ion is andom. Finally, we implemen ed an
algo i hm ha combines bo h solu ions in one o he possible
ways. The algo i hm da a +ne wo k is mainly based on
he da a_clus e algo i hm, bu in case i encoun e s equal
assignmen s, i uses he ne wo k_clus e algo i hm o decide.
All hese algo i hms a e compa ed wi h he manual manual
se ice selec ion.
IV. NETWORK PERSPECTIVE
The impac on he ne wo k is s udied acco ding o he
o al by es ans e ed h ough each link du ing he simula ion.
We do no ake in o accoun possible e ansmissions, and we
assume ha he links canno be sa u a ed and ha e always he
same pe o mance, e en ac oss di e en links.
As shown in Figu e 3, he ne wo k_clus e algo i hm
ou pe o ms signi ican ly he o he algo i hms in dis ibu ing
he load in he links. I main ains he o al a ic o 50%
o he links, one o de o magni ude lowe han he o he
algo i hms wi hou compensa ing ha by o e loading a ew
links, as we would expec o he links ha connec he clus e s.
The o he algo i hms p esen a simila , bu shi ed, dis ibu ion.
Mo eo e , conside ing ha each algo i hm is using di e en
numbe o links o send he a ic, i is wo h men ioning
ha ne wo k_clus e ans e s he lowes o al amoun o
by es, 1 Te aby e, while da a_clus e is he mos expensi e
ans e ing 1.7 Te aby es. We also ind ha da a +ne wo k
lies be ween ne wo k_clus e and da a_clus e , wi h 1.4
Te aby es, while manual ans e s 1.3 Te aby es.
O e all, we obse e ha ne wo k locali y plays a signi ican
ole in dis ibu ing he load on he ne wo k. E en in he case
o exis ing communi ies wi hou p oxies, like he s udied case,
a locali y-awa e se ice can educe i s ne wo k impac .
V. PROXY PERSPECTIVE
F om he pe spec i e o he p oxies, i is impo an o bo h
he se ice pe o mance and he use ’s expe ience o dis ibu e
he load acco ding o he capaci y and pe o mance o each
p oxy.
P oxy1 P oxy2 P oxy3 P oxy4
1010
1011
1012
To alBy es
manual
clus e
1
2
3
P oxy1 P oxy2 P oxy3 P oxy4
1010
1011
1012
To alBy es
da a_clus e
clus e
1
2
3
P oxy1 P oxy2 P oxy3 P oxy4
1010
1011
1012
To alBy es
ne wo k_clus e
clus e
1
2
3
P oxy1 P oxy2 P oxy3 P oxy4
1010
1011
1012
To alBy es
da a+ne wo k
clus e
1
2
3
Figu e 4. Compa ison o S a egies: T a ic pe P oxy and Clus e ing
da a+ne wo k da a_clus e manual ne wo k_clus e
0
1
2
3
4
5
6
Rela i eP oxiesT a icVa iance
Figu e 5. P oxies pe second Rela i e T a ic Va iance
In ou simula ions we s a by assuming ha all p oxies
ha e in ini e capaci y and he same p ocessing pe o mance
(i.e., unlimi ed h oughpu ). We e alua e he di e en algo-
i hms by he o al amoun o by es sen o each p oxy pe
s a egy, wi h in o ma ion o he co esponding clus e s, as
seen in Figu e 4. We ini ially obse ed ha he hea y use s
occupy an impo an pe cen age o he a ic, e en hough
hey a e nea ly he 1% o he o al use s. None heless, ligh
use s gene a e he majo i y o a ic despi e he ac ha each
o hem use he se ice compa a i ely much less. The e o e, as
a esul o manual selec ion he p oxy load is e y unbalanced,
bu he da a_clus e and da a+ne wo k algo i hms succeed
in balancing he a ic. The ne wo k_clus e app oach can
esul o an imbalance in he load among p oxies, due o sub-
ne wo ks wi h an une en numbe o clien s and he p oxies
incon enien ly placed wi h espec o he clien s. I is wo h
no ing ha , om he p oxy pe spec i e, he da a +ne wo k
algo i hm achie es i s goal e y success ully since i is mainly
based on he da a_clus e algo i hm; howe e , i also achie es
a be e pe o mance han da a_clus e om he ne wo k pe -
spec i e. Mo eo e , as shown in Figu e 5, he sum o dis ances
o a ic alues o each p oxy o he mean a each ins an is
clea ly smalle o he da a_clus e o da a+ne wo k. This
small a iabili y implies ha hese algo i hms wo k well o e
bo h sho and long e m pe iods. We can he e o e deduce ha
an algo i hm ha combines bo h, use clus e ing and ne wo k
g aph communi y de ec ion, can be used o uning he ade-
o o impac o une en p oxy load and excessi e ne wo k
impac , due o long ne wo k pa hs. This lesson is applicable
o se e selec ion in a decen alized se ice.
I we ake in o conside a ion he limi ed capaci y and
h oughpu in p oxies, hen balancing he a ic ac oss hem
acco ding o he capaci y o each p oxy becomes a key issue.
Fo example, in he case o a la ge numbe o use s he
clus e ing in o ma ion could be used o pe o m admission
con ol and he e o e conges ion con ol in he p oxy.
In he cu en scena io p oxies ha e a ough admission con-
ol based exhaus ion o limi s, and hey do no on conges ion
con ol acco ding o load o pe o mance. P oxies ake new
eques s based on a maximum numbe o concu en clien s,
e en when he p oxy se ice is al eady unde -pe o ming o
ongoing esponses. This esul s in poo pe o mance du ing
peaks o la ge eques s ha cause conges ion o a se ice
imeou . In ou decen alized scheme, clien s ha e a lis o
se e al p oxy choices. Clien s make an ini ial choice, p oxies
can ejec connec ions, and clien s can jus make a new local
choice, anspa en ly e y and con inue om he e, wi h no
majo isible e ec o he use . The combina ion o clien s
using a lis o p oxy choices, p oxy admission con ol, and
ne wo k ou ing choices esul s in a simple, decen alized and
coope a i e egula ion scheme ha equi es li le coo dina ion.
Admission con ol is impo an in la ge use popula ions,
e.g., wide-a ea ne wo ks wi h many p oxies, since p oxies ha e
a limi ed In e ne access capaci y. Any ho spo o imbalance
in a massi e sys em can easily lead o conges ion, ei he in
he access ne wo k, any p oxy o he In e ne access, esul ing
in a d ama ic educ ion o se ice h oughpu o many use s
o ha p oxy.
In addi ion o he local choices a each clien and p oxy,
he e is po en ial o global op imiza ion in balancing global
choices, ac oss all p oxies, by combining he use a ic
beha iou , use p oxy choices, and p oxy capaci ies. Thus, we
can help a oid globally imbalanced scena ios, whe e a p oxy
is sa u a ed o p o iding low h oughpu , while a he same
ime ano he p oxy is unde u ilized.
VI. USERS PERSPECTIVE
The e alua ion o impac on se ice pe o mance om he
use pe spec i e is he mos complex, as use s ha e di e en
me ics o assess hei se ice acco ding o hei di e se usage
habi s. While explo ing hese me ics is u u e wo k, he e
we p esen a i s simple cos model o es ima e how use s
pe cei e he impac o he p esen ed algo i hms. We assume
ha use s y o minimize he ans e ime in he local ne wo k,
combined wi h he p ocessing ime in he p oxy se e .
As a as he ne wo k is conce ned, we de ine as cl he cos
o he link l, in e ms o ime, o ans e one by e, assuming
ha he links ha e in ini e capaci y, al hough we plan o s udy
mo e sophis ica ed models in he u u e.
Fo each use we calcula e he o al cos o he ne wo k
ans e as Pn
l=0 cl∗bu, lLu, whe e Luis he se o links
and bu he o al numbe o by es a ibu ed o use u.
The use s’ pe cep ion o he p oxy pe o mance is modeled
simila ly o he ne wo k pe o mance. We de ine cpas he cos
o p oxy p o p ocess one by e, om he ime i ecei es he
eques om he use , un il i sends he las by e. We calcula e
he cos cpo each p oxy psepa a ely o e e y s a egy as
/ Pbu, uUp, whe e is he o al measu emen ime, bu he
o al numbe o by es sen by use uand Up he se o use s
o p oxy p. Based on ha , he p oxy pe cei ed cos o each
use is: cp∗bu.
min_hop ne wo k_clus e manual da a+ne wo k da a_clus e andom
105
106
107
108
109
To alTimepe Use (s)
Figu e 6. Cos pe use ECDF
Conside ing ha he cos s a e linea and independen , we
can assume ha he o e all cos pe cei ed by a use uis:
Cu=Pn
l=0 cl∗bu+cp∗ u, lLu. Hence, he objec i e o use
uwould be o minimize Cu. Figu e 6 p esen s he dis ibu ion
o he use s’ cos s o each o he p esen ed s a egies.
While he dis ibu ions ha e e y simila beha io , we can
obse e ha o 80% o he use s, he ne wo k communi y
de ec ion s a egy pe o ms sligh ly be e han he cu en
si ua ion, and he es o he s a egies ollow. The communi y
s a egy achie es equi alen esul s o a min −hop s a egy,
only di e ing when p oxies a e no in he cen e o i s
zone. The andom s a egy achie es equi alen esul s o
clus e , as he la e only ca es abou con en s and none abou
in as uc u al aspec s.
The ne wo k e iciency o communi y-based p oxy selec-
ion, and he e o e he impac o ne wo k locali y, appea s as
an impo an ac o . S udying he indi idual cos s we obse e
ha he ne wo k ans e ime cos is, in a e age, signi ican ly
highe han he p oxy p ocessing cos . This ac explains why
he communi y solu ion pe o ms be e o e all, e en hough
i is an ine icien op ion o load dis ibu ion in he p oxies.
The (clus e ing acco ding o) use beha io appea s o ha e
an in luence on he use pe cei ed pe o mance (cos ), since
i p esen s a di e en ia ed beha io om he cu en si ua ion
(manual p oxy selec ion). Howe e , he simplici y o he model
does no allow us o d aw mo e conclusions.
In con as , he cu en si ua ion is ha clien s (Web
b owse s) ha e a lis o p oxy se e s manually de ined o

adjus ed. The ini ial con igu a ion is based on hin s om o he
nea by use s, o by downloading he lis om a local gui i.ne
o um. The adjus men s come om simila sou ces, pe sonal
usage expe ience, hin s om o he use s o news abou new
p oxies being o e ed. Web b owse s swi ch o ano he p oxy
se e jus when a p oxy ails o espond and do no p o-
ide load balancing, o mo e e ec i e choices conside ing o
deg ada ion, conges ion signals o ela i e pe o mance. These
models enables us o design a se ice selec ion algo i hm ha
akes in o accoun he cha ac e is ics o he use s and he
local ne wo k, con on ing hus he ine iciencies caused in
he se ice and he use expe ience by he manual s a ic p oxy
selec ion.
VII. CONCLUSIONS AND FUTURE WORK
The pape p esen s an analysis o how c owdsou ced ne -
wo k in as uc u es can p o ide e y cos e ec i e ways o
access he In e ne . We look a a dis ibu ed p oxy se ice
in gui i.ne , one o he la ges communi y ne wo ks in he
wo ld. The analysis o se ice logs shows pa e ns o usage
and ne wo k opology g ouping use s and p oxies, ha can
in luence he c i e ia o p oxy selec ion. The cu en ly manual
and no well-in o med choice o p oxies by clien s wo k a he
well o i s use s, bu i esul in ine iciencies ha a ec
he se ice cos and shows episodes o deg aded pe o mance.
Conside ing ha si ua ion, his pape explo es al e na i es o
cos educ ion and se ice imp o emen when going om a
simple bu igid mapping be ween use s and p oxies, owa ds
coo dina ed in o med choices based on se e al me ics. Design
ade-o s lie in conside ing in as uc u al aspec s (e.g., educe
ne wo k cos , a oid ne wo k and p oxy conges ion) and se ice
aspec s (e.g., good esponse ime o QoE).
The combina ion o se e al e na i es in clien s, ine g ain
p oxy admission con ol, and he unde lying ne wo k ou -
ing decisions esul in a decen alized coope a i e egula ion
scheme ha can p o ide a c owdsou ced p oxy se ice, wi h
good pe o mance and equi ing li le coo dina ion. Mo eo e ,
ha scheme allows scaling up he ne wo k o la ge sizes. As
pa o he u u e wo k, we plan o e alua e hese s a egies in
de ail wi h expe imen s unde se e al eal condi ions, whe e
clien s pe o m mo e in o med choices and p oxies pe o m
mo e ine g ained admission con ol.
ACKNOWLEDGMENTS
This wo k was pa ially suppo ed by he E asmus Mundus
Join Doc o a e in Dis ibu ed Compu ing (EMJD-DC) unded
by he Eu opean Commission (EACEA) (FPA 2012-0030), he
EU Ho izon 2020 F amewo k P og am p ojec ne Commons
(H2020-688768), he Spanish go e nmen unde con ac
TIN2016-77836-C2-2-R, he Gene ali a de Ca alunya as
Consolida ed Resea ch G oup 2014-SGR-881, he Chilean
Fondecy g an 1150252, and Po uguese unds h ough
Fundação pa a a Ciência e a Tecnologia wi h e e ence
UID/CEC/50021/2013. Ou special hanks o Roge Baig and
se e al gui i.ne membe s o he in o ma ion, suppo and
logs ha ha e made his wo k possible.
REFERENCES
[1] In e ne Socie y, “Global in e ne epo 2015,” In e ne Socie y, Tech.
Rep., Oc obe 2015. [Online]. A ailable: h p://www.in e ne socie y.
o g/globalin e ne epo /
[2] L. Na a o, F. F ei ag, R. Baig, and R. Roca, Communi y Connec i i y:
Building he In e ne om Sc a ch. In e ne Go e nance Fo um, 2016,
ch. A commons-o ien ed amewo k o Communi y Ne wo ks, pp. 25–
71.
[3] R. Baig, R. Roca, F. F ei ag, and L. Na a o, “gui i. ne , a c owdsou ced
ne wo k in as uc u e held in common,” Compu e Ne wo ks, ol. 90,
pp. 150–165, 2015.
[4] E. Os om, Go e ning he commons: he e olu ion o ins i u ions o
collec i e ac ion. Camb idge Uni e si y P ess, No . 1990.
[5] ne Commons, “Deli e able D1.2, Repo on he Exis -
ing CNs and hei O ganiza ion ( 2),” Sep 2016,
ech. Rep. D1.2. [Online]. A ailable: h p://ne commons.eu/?q=
con en / epo -exis ing-cns-and- hei -o ganiza ion- 2
[6] L. Ce da-Alabe n, “On he opology cha ac e iza ion o gui i.ne ,” in
P oc. In . Con . Wi eless and Mobile Compu ing, Ne wo king and
Communica ions (WiMob), Ba celona, Spain, 2012, pp. 389–396.
[7] D. Vega, L. Ce da-Alabe n, L. Na a o, and R. Mesegue , “Topology
pa e ns o a communi y ne wo k: Gui i.ne ,” in P oc. In . Con . Wi eless
and Mobile Compu ing, Ne wo king and Communica ions (WiMob),
Ba celona, Spain, 2012, pp. 612–619.
[8] L. Macca i and R. Lo Cigno, “A week in he li e o h ee la ge wi eless
communi y ne wo ks,” Ad Hoc Ne wo ks, ol. 24, Pa B, pp. 175–190,
2015.
[9] D. Vega, R. Baig, L. Ce dà-Alabe n, E. Medina, R. Mesegue , and
L. Na a o, “A echnological o e iew o he gui i. ne communi y
ne wo k,” Compu e Ne wo ks, ol. 93, Pa 2, pp. 260–278, 2015.
[10] L. Macca i, L. Baldesi, R. Lo Cigno, J. Fo coni, and A. Caiazza,
“Li e Video S eaming o Communi y Ne wo ks, Expe imen ing wi h
Pee S eame on he Ninux Communi y,” in P oc. Wo kshop on Do-i -
you sel Ne wo king: An In e disciplina y App oach (DIYNe wo king),
Flo ence, I aly, 2015, pp. 1–6.
[11] L. Baldesi, L. Macca i, and R. Lo Cigno, “Imp o ing P2P s eaming
in Wi eless Communi y Ne wo ks,” Compu e Ne wo ks, ol. 93, Pa
2, pp. 389–403, 2015.
[12] R. Lo Cigno and L. Macca i, “U ban Wi eless Communi y Ne wo ks:
Challenges and Solu ions o Sma Ci y Communica ions,” in ACM
In . Wo kshop Wi eless and Mobile Technologies o Sma Ci ies
(WiMobCi y), pa o MobiHoc, Philadelphia, PA, US, 2014, pp. 49–54.
[13] D. Vega, R. Mesegue , F. F ei ag, and S. F. Ochoa, “Mo i a ing he non-
echnical pa icipa ion in echnical communi ies,” in P oc. Compu e
Suppo ed Coope a i e Wo k in Design (CSCWD), 2015, pp. 259–264.
[14] G. WG, “Global access o he in e ne o all esea ch g oup,”
2016, [Online; accessed 14-Sep embe -2016]. [Online]. A ailable:
h ps://i .o g/gaia
[15] C. Rey-Mo eno, Z. Ro o, W. D. Tucke , M. J. Siya, N. J. Bidwell, and
J. Simo-Reigadas, “Expe iences, challenges and lessons om olling
ou a u al wi i mesh ne wo k,” in ACM Compu ing o De elopmen
(ACM-DEV), Bangalo e, India, 2013, p. 11.
[16] R. Sen, H. A. Pi zada, A. Phokee , Z. A. Fa ooq, S. Sengup a,
D. Cho nes, and K. P. Gummadi, “On he ee b idge ac oss he digi al
di ide: Assessing he quali y o acebook’s ee basics se ice,” in P oc.
ACM on In e ne Measu emen Con e ence (IMC), San a Monica, CA,
USA, 2016, pp. 127–133.
[17] E. Dimoge on akis, J. Ne o, R. Mesegue , L. Na a o, and L. Veiga,
“Clien -side ou ing-agnos ic ga eway selec ion o he e ogeneous wi e-
less mesh ne wo ks,” in IFIP/IEEE In . Symp. In eg a ed Ne wo k
Managemen , IM, Lisboa, Po ugal, 2017, pp. 1–8.
[18] P. Be khin, G ouping Mul idimensional Da a: Recen Ad ances in
Clus e ing. Sp inge , 2006, ch. A Su ey o Clus e ing Da a Mining
Techniques, pp. 25–71.
[19] A. Lancichine i and S. Fo una o, “Communi y de ec ion algo i hms:
A compa a i e analysis,” Phys. Re . E, ol. 80, p. 056117, No 2009.