scieee Science in your language
[en] (orig)

Building a flexible web caching system.

Abstract

Web caching is a technology that has demonstrated to improve traffic on the Internet. To find out how to implement a Web caching architecture that assures improvements is not an easy task. The problem is more difficult when we are interested in deploying a distributed and cooperative Web caching system. We have found that some cooperative Web caching architectures could be unviable when changes on the network environment appear. This situation suggests that a cooperative Web caching system could get worst access to Web objects. However in this paper we present an architecture that combines the best of several Web caching configurations that we have previously analyzed. Our architecture gives basic ideas for implementing a cooperative Web caching system using groups of HTTP proxy servers which can improve access to remote Web objects regardless of the changes that might occur on the network environment (changes that could produce modifications in Web object validation policies and/or types of caching communication).

Read accessible full text

Building a flexible web caching system.

Author: Sosa Sosa, Víctor Jesús,González Serna, Juan Gabriel,Navarro Moldes, Leandro
Publisher: IEEE
Year: 2003
Source: https://upcommons.upc.edu/bitstream/2117/2277/1/sosa.PDF
Building a Flexible Web Caching Sys em
Víc o J. Sosa Sosa, Gab iel González S.,
Cen o Nacional de In es igación y Desa ollo
Tecnológico
In e io In e nado Palmi a S/N, Cue na aca,
Mo elos, México. 62490
{ jsosa, gab iel}@cenide .edu.mx
Leand o Na a o
Uni e si a Poli ècnica de Ca alunya (UPC)
Jo di Gi ona, 1-3, D6-105, Campus No d .
Ba celona, Spain. E-08034
[email p o ec ed]
Abs ac
Web caching is a echnology ha has demons a ed o
imp o e a ic on he In e ne . To ind ou how o
implemen a Web caching a chi ec u e ha assu es
imp o emen s is no an easy ask. The p oblem is mo e
di icul when we a e in e es ed in deploying a dis ibu ed
and coope a i e Web caching sys em. We ha e ound ha
some coope a i e Web caching a chi ec u es could be
un iable when changes on he ne wo k en i onmen
appea . This si ua ion sugges s ha a coope a i e Web
caching sys em could ge wo s access o Web objec s.
Howe e in his pape we p esen an a chi ec u e ha
combines he bes o se e al Web caching con igu a ions
ha we ha e p e iously analyzed. Ou a chi ec u e gi es
basic ideas o implemen ing a coope a i e Web caching
sys em using g oups o HTTP p oxy se e s which can
imp o e access o emo e Web objec s ega dless o he
changes ha migh occu on he ne wo k en i onmen
(changes ha could p oduce modi ica ions in Web objec
alida ion policies and/o ypes o caching
communica ion).
1. In oduc ion
The idea behind Web caching consis s in ge ing Web
objec s close o clien s a a low cos . Coope a ing p oxy
caches a e a g oup o caches ha sha e cached objec s and
collabo a e wi h each o he o do he same wo k as a single
Web cache. The bene i s o ha ing a coope a i e caching
sys em has been analyzed in [12],[6],[1]. Basically, he
cons uc ion o a coope a i e Web cache sys em equi es
he analysis o ou majo opics. They a e b ie ly
desc ibed nex :
Coope a i e caching sys em o ganiza ion: How o de ine
a cache opology.
Hie a chy: Caches a e loca ed a di e en ne wo k le els.
In mos cases i is assumed ha in e io le els in he
hie a chy ha e be e quali y o se ice. They ha e a
pa en -son ela ionship. A son cache is loca ed a in e io
le els in he hie a chy, and when a son cache needs a Web
objec , he son cache asks i s pa en cache o i . The
eques goes up in he hie a chy un il inding he Web
objec needed in a pa en cache o in he o iginal Web
se e .
Mesh (dis ibu ed): The e a e no in e media e caches
de ined by le els, a he he e is a single le el o caches
whe e hey can coope a e o se e he eques s gene a ed
by clien s.
Hyb id (mesh/hie a chy): A combina ion o hie a chy and
mesh.
Web caching communica ion: How caches a e going o
communica e each o he . We conside h ee p ocesses ha
a e in ol ed in Web caching communica ion: disco e y,
deli e y, and dissemina ion.
Disco e y. How do caches ind he Web objec s? The e
a e h ee majo app oaches: Exhaus ed que y: asking o a
eques ed objec o all sibling caches using a p o ocol like
ICP (In e cache Communica ion P o ocol). Using diges :
Diges can be in e changed using me hodologies such as:
Pee o pee o by a hie a chy. Using hashing: cache
objec s can be loca ed in p oxy caches de ined by a hash
unc ion.
Deli e y. How do caches deli e pages o clien s? I could
be using di ec connec ion be ween he cache con aining
he page and he clien , deli e ing copies using a cache
hie a chy, o deli e ing copies using a cache mesh.
Dissemina ion. Deli e y o Web objec s ini ia ed by
o iginal se e s.
Consis ency s a egies: How caches keep “ esh” cached
objec s.
Expi e. Using p ede ined expi a ion da es on Web pages.
TTL0. Ve i ying consis ency e e y ime a hi occu s ( ime
o li e is ze o, l=0).
TTLA. Time o li e is assigned based on he elapsed ime
since he las eques .
In alida ion. O iginal se e sends an in alida ion
message o caches when a Web objec upda e occu s.
Usually his messages is sen ia mul icas (in alida ion
mul icas ). Some a ian s o mul icas in alida ion o a
mix o all p e ious s a egies.
P oceedings o he Fou h Mexican In e na ional Con e ence on Compu e Science (ENC’03)
0-7695-1915-6/03 $17.00 © 2003 IEEE
Au ho ized licensed use limi ed o: IEEE Xplo e Cus ome . Downloaded on Oc obe 7, 2008 a 8:19 om IEEE Xplo e. Res ic ions apply.
Wo kload beha io : Beha io o wo kload can be ea ed
depending on he pa icipa ing elemen : Clien s, P oxy-
Caches and Se e s. In his case we ocus ou a en ion on
p oxy-caches wo kload.
The goal o his wo k is o show many op ions o building
a iable Web caching sys em based on many analyses o
di e en Web caching con igu a ions. We also p opose a
Web caching a chi ec u e ha akes he bes o se e al
Web caching con igu a ion. The p incipal goal o his Web
caching a chi ec u e is o a oid p oblems ha could make
i un iable when he ne wo k beha io changes.
2. A chi ec u e desc ip ion
This a chi ec u e is based on ou p e ious esea ch
[8],[9],[10], which encompasses se e al s udies o Web
cache sys ems (wo kloads, consis ency s a egies, Web
caching communica ion). The main cha ac e is ics o ou
a chi ec u e a e:
1) Web caching o ganiza ion. Conside ing he ob ained
esul s in [8],[9],[10], we p opose a caching o ganiza ion
o hyb id ype ( h ee le el hie a chy in long la encies,
whe e each le el has a mesh, Fig. 1). I has been
demons a ed in [8],[6],[12] ha a hyb id o ganiza ion
o e s be e esul s as o access ime, bandwid h
consump ion in a wide a ea ne wo k (WAN), and scaling
suppo . Likewise, a h ee le el hie a chy has been
widely accep ed [4],[3],[7],[5] because i does no
emphasize he cos o s o e-and- o wa d p ocess.
2) Caching communica ion. As we ha e men ioned, he
communica ion in a coope a i e caching sys em akes
places in se e al p ocesses (disco e y, deli e y, and
dissemina ion).
Disco e y p ocess: This p ocess allows o he loca ing
o eques ed Web objec s wi hin he coope a i e caching
sys em. A e y ypical mechanism is he exhaus i e que y
by using ICP messages. Howe e , exhaus i e que y will
no be used in ou a chi ec u e, due o he high bandwid h
consump ion ha i equi es [2]. A mechanism o he
cons uc ion o di ec o ies (me ada a) will be used ins ead.
The cons uc ion o such di ec o ies will be done
dynamically. When a pa en cache ecei es a eques o a
Web objec , immedia ely a e he eques ed Web objec
has been sen o i s eques e cache, a mul icas message
is sen o all o he pa en cache descenden s o no i y hem
he Web objec s ha ha e been eques ed by he eques e
cache. This le s he lowe le els o he hie a chy know he
loca ion o hese Web objec s o u he eques s. The
caches ha will ecei e such no i ica ion a e only hose
caches ha belong o ha pa icula b anch in he lowe
le els o he hie a chy.
Deli e y p ocess: Clien s ob ain a copy o he eques ed
Web objec h ough he mesh (caches o he same
hie a chy le el –b anch-). Those copies ha e a i ed o he
mesh ollowing a ou e om he o iginal se e o he
cache h ough he hie a chy le els.
Dissemina ion p ocess: This p ocess akes ad an age o
he mul icas in alida ion mechanism o push upda ed
Web objec s ha ha e been p e iously eques ed by he
lowe le el caches in he hie a chy. This p ocess will be
ully explained nex .
3) Consis ency mechanism. The consis ency mechanism
o be used in his a chi ec u e is mul icas in alida ion, and
i uses wha we call li e signals. The pa en cache sends a
message (mul icas message -li e signal-) e e y minu e o
he lowe le el caches o indica e ha i is ali e (on line).
I a Web objec is modi ied in he o iginal se e (one o
hose Web objec s p e iously eques ed), a message is sen
o he lowe le el caches wi hin he same mesh, o le hem
know ha such Web objec is no longe alid, and ha an
upda ed copy o he Web objec should be ob ained he
nex ime he Web objec is eques ed. The same li e signal
is used o in o m and p e en hose caches ha did no
con ain he Web objec om ying o access such a Web
objec om an in alid copy.
Dissemina ion s a egy (pushing) h ough he
consis ency mechanism. We ha e a dissemina ion s a egy
(pushing) whe e he se e s (o pa en caches in a
hie a chy) mul icas he mos popula Web objec s ( ha
ha e been upda ed by he o iginal se e ) o he caches,
be o e being eques ed once again. In his a chi ec u e, no
global decision is aken conce ning he mul icas ing o a
Web objec h ough he sys em. Con e sely, each cache
and he o iginal se e make hei own decision on
whe he he dissemina ion should be made o no . To do
his, each cache and he o iginal se e keep an access
coun e (Xp) o each Web objec which in he beginning
is se o 0, along wi h a dissemina ion bi ha will indica e
whe he he Web objec mus be dissemina ed o no . I he
dissemina ion bi is se o 1, ha will indica e ha he
cache has o dissemina e he upda ed Web objec when he
cache issues a li e signal o i s lowe le el caches. The
heu is ics o his s a egy use h ee posi i e a iables:
X,E,G. I a cache ecei es an in alida ion message o page
P, hen he nex calcula ion akes place: XP = XP - E. i he
cache ecei es a eques o page P, hen we calcula e: XP
= XP + G. I XP exceeds a h eshold X (XP > X), hen he
dissemina ion bi is se o 1, o he wise a ze o is assigned.
Addi ionally, he lowe le el caches send a message o
hei pa en caches each ime he page is ead. This is done
o keep a knowledge abou all ead pages wi hin ha
b anch o hie a chy. The ixed alues used wi hin ou
a chi ec u e a e X=8, E=1, y G=2. The app oach ha we
ollowed o aking a 1 o each in alida ion and o adding
a 2 o each eques , is due o he popula i y s udies made
wi h some p oxi-caches logs ha a e s ongly close o he
Zip dis ibu ion. The Zip dis ibu ion says ha he
numbe o eques s o a page (R) is ela ed o i s
popula i y in his way: :
R(i) = ----------
iD
P oceedings o he Fou h Mexican In e na ional Con e ence on Compu e Science (ENC’03)
0-7695-1915-6/03 $17.00 © 2003 IEEE
Au ho ized licensed use limi ed o: IEEE Xplo e Cus ome . Downloaded on Oc obe 7, 2008 a 8:19 om IEEE Xplo e. Res ic ions apply.
whe e he exponen D e lec s he popula i y skew be ween
one Web objec and ano he , and he cons an : de ines an
app oxima ion o he numbe o eques s o he mos
popula Web objec ep esen ed by i= 1. Fo a be e
unde s anding, le s assume ha D = 1 and ha : = 100,
hen he mos popula Web objec (in o he wo ds, he one
wi h he posi ion i= 1) will ha e 100 eques s, he Web
objec nex o i in popula i y will ha e 50 (i = 2), and he
nex (i= 3) will be close o 30 and so o h. All his gi es
us an idea ha he popula i y o Web objec s ollows a
sequence o 2 o 1 o each posi ion in he popula i y
index. This helps us o de ine cons an s E and G. The X
alue comes om s udies made in his wo k, along wi h
heu is ic analysis made in [13].
3. How his web caching a chi ec u e was
de ined
This Web caching a chi ec u e was subs an ially ob ained
using a simula o de eloped du ing he ime o his wo k
[10]. This simula o is mainly an ex ension o he Ne wo k
Simula o (NS) [11]. The undamen al esea ch ha helps
us o de ine his a chi ec u e could be ound in ou wo k
[8], which basically consis ed in analyzing he mos
s udied ways o o ganize coope a i e Web caches like:
hie a chies, mesh and hyb ids, combined wi h he mos
s udied Web caching consis ency s a egies. In his wo k
we ied he ollowing consis ency mechanisms: TTL0,
TTLA, M, and some a ian s o mul icas in alida ion
mechanism: PSM and APM. PSM (pushing selec i e
mul icas ) only pushes he upda ed Web objec when he
numbe o eques s exceeds a h eshold de ined by X (as
men ioned in he p e ious sec ion). APM (always pushing
mul icas ) pushes a Web objec whene e i is upda ed.
The con igu a ions men ioned abo e we e implemen ed in
ou simula o . We use wo kloads ha we e ob ained om
a cache loca ed a he Spanish academic ne wo k backbone
managed by he Cen e o Supe Compu ing o Ca alonia
(CeSCa). The simula ion s ops when i inishes eplaying
all he wo kload. Table 1 desc ibes he wo kload. The
Web caching a chi ec u es ha gi e us he bes esul s in
[8] we e he base o ou new Web caching a chi ec u e
discussed he e .
Table 1. Wo kload cha ac e is ics.
Numbe o eques s 3,089,592
Numbe o Web objec s 212,352
Numbe o clien s (app ox.) 11,765
A e age eques s pe second 21
T ans e ed by es 30GB
Du a ion 2 days
4. How his a chi ec u e wo ks
This sec ion desc ibes how his a chi ec u e wo ks. The
Web caching sys em s a s when a clien sends a eques o
i s Web p oxy cache ( he i s cache ha is con ac ed by a
clien is called clien -cache). The clien -cache e i ies i i
can espond o he eques . I so, he clien -cache sends he
eques ed Web objec o he clien and he ope a ion is
inished. O he wise, he clien -cache checks i he
eques ed objec is in a sibling cache. A sibling cache is a
cache connec ed by a mesh a he same b anch in a cache
hie a chy (see Fig. 1). E e y cache sibling has a diges
whe e he cached objec s in a hie a chy b anch a e
egis e ed. I he eques ed Web objec is in a sibling
cache (a cache knows ha by looking a i s diges ), he
cache o wa ds he eques o he cache sibling. A copy o
he eques ed Web objec a els o he end clien h ough
he clien -cache which keeps a copy o he Web objec . I
he eques ed Web objec is no in any cache sibling hen
he eques is o wa ded o he pa en cache. The same
p ocess is epea ed a each hie a chy le el un il he
eques ed Web objec is ound in a pa en cache o in he
o iginal se e (i he eques ed Web objec is no ound in
he Web caching in as uc u e). E e y pa en cache keeps
a lis o pai s { eques e _son_cache,
eques ed_Web_objec } which is sen o all i s son caches
piggybacking he li e signal used in he consis ency
mechanism explained be o e. E e y cache ecei ing ha
lis includes i in i s diges . In ha way e e y son cache is
building i s own diges . The lis is emo ed om he
pa en cache igh a e he lis has been sen . This educes
con ol o e head in pa en caches. As we ha e said be o e,
e e y cache diges only has e e ences o Web objec
copies s o ed a he same b anch (mesh which he cache
belongs o) in he hie a chy. Likewise, when a Web objec
is modi ied in he o iginal se e , he se e keeps a lis o
modi ied_Web_objec s o be sen piggybacking he li e
signal used in he consis ency mechanism. Figu e 1 shows
an example o ou a chi ec u e. We can see in Figu e 1
wi h a double line he pa h ha he li e signal will ake.
Figu e 1. A possible scena io using he Web caching
a chi ec u e
The li e signal is an indica o o son caches ha he pa en
cache is on line (ali e). I a son cache does no ecei e i e
consecu i e li e signals, hen he son cache p esumes ha
Se e
Mes
h
Mes
h
Mes
h
Mes
h
Mes
h
Mes
hMes
h
Mes
h
Mes
hMes
h
Mes
h
Mesh
L1
L2
L3
Clien Caches
Mes
h
Mes
h
Mesh
Mesh
Mes
h
Siblings
Siblings
Siblings
L1’s
pa en s
L2’s pa en s
P oceedings o he Fou h Mexican In e na ional Con e ence on Compu e Science (ENC’03)
0-7695-1915-6/03 $17.00 © 2003 IEEE
Au ho ized licensed use limi ed o: IEEE Xplo e Cus ome . Downloaded on Oc obe 7, 2008 a 8:19 om IEEE Xplo e. Res ic ions apply.
i s pa en cache is dead. A e ha , he son cache
in alida es all he Web objec s ecei ed by i s pa en
cache. We allow i e missed messages because we wan ed
o conside some high ne wo k conges ion. E e y pa en
cache (a e e y hie a chy le el) will o wa d (mul icas )
he in alida ion message wi h he li e signal i one o i s
son caches has he in alida ed Web objec , o he wise he
pa en cache will o wa d only he li e signal. I a Web
objec is modi ied in he o iginal se e and he push
mechanism is ac i a ed (pushing bi = 1), hen he new
Web objec will be piggybacked on he li e signal. I a
pa en cache has no been eques ed wi h ha Web
objec (s), he pa en cache will no o wa d he Web
objec (s) bu only he li e signal.
5. Analysis o his a chi ec u e
This sec ion shows some compa a i e analyses o he Web
caching a chi ec u e sugges ed in his pape . Table 2
summa izes he simula ion scena ios. These simula ion
scena ios use he wo kload desc ibed in able 1.
Table 2. Scena ios o compa ing Web caching
a chi ec u es
Iden i ie s Desc ip ion
JTTLA0,
JTTLA, JM,
JPSM, JAPM
Hie a chical coope a ion a chi ec u e (J) using
he ollowing consis ency mechanism:
TTLA0, TTLA, M, PSM, APM
DTTLA0,
DTTLA, DM,
DPSM, DAPM
Dis ibu ed coope a ion a chi ec u e (D) using
he ollowing consis ency mechanisms:
TTLA0, TTLA, M, PSM, APM
H1TTLA0,
H1TTLA,
H1M, H1PSM,
H1APM
Hyb id coope a ion a chi ec u e 1 (H1) using
he ollowing consis ency mechanism:
TTLA0, TTLA, M, PSM, APM
H2TTLA0,
H2TTLA,
H2M, H2PSM,
H2APM
Hyb id coope a ion a chi ec u e 2 (H2) using
he ollowing consis ency mechanism:
TTLA0, TTLA, M, PSM, APM
ADDTTA0,
ADDTTLA,
ADDM,ADDP
SM, ADDPM
Web Objec s Dis ibu ion a chi ec u e (ADD),
sugges ed in his pape , using he ollowing
consis ency mechanisms: TTLA0, TTLA, M,
PSM, APM
The i s ha we can see in his analysis is he pe cei ed
esponse ime by clien s. Figu e 2 shows how he
consis ency mechanisms a ec e e y coope a i e caching
o ganiza ion in se e al ways along he wo kload on his
simula ion scena io (CESCA’s coope a i e Web caching
sys em). We can see ha dis ibu ed caching coope a ion
a chi ec u e in his con ex p esen s he bes esul s. I is
impo an o no ice ha sligh ly lowe in his a chi ec u e
we can see he ADD a chi ec u e which shows be e
esul s han he o he s. I was no ele an which
consis ency mechanism was used. I is in e es ing o no ice
ha when we ha e TTLA as a consis ency mechanism
implemen ed on a hyb id coope a i e Web caching
a chi ec u e (H2), esponse imes a e good. Howe e , i
he consis ency mechanism is changed in his a chi ec u e,
he esponse ime inc eases conside ably..
0
50
100
150
200
250
300
350
400
450
TTLA0 TTLA M PSM APM
Consis ency Mechanism
Hou s
Hie a chy
Dis ibu ed
H1
H2
ADD
Figu e 2. To al esponse ime pe cei ed by clien s
Figu e 3 shows he bandwid h (BW) consump ion o he
in e -cache ne wo k and a ic gene a ed o he wide a ea
ne wo k (WAN). I is impo an o see bo h a ics in a
sepa a e way because we can see whe e he bo lenecks a e
(i any), and o be awa e i ou HTTP a ic is yielding
some p oblems o ano he ype o a ic inside and ou side
he in e -cache ne wo k.
Figu e 3. Bandwid h (BW) consump ion inside and
ou side o he in e -cache ne wo k
Dis ibu ed a chi ec u es ha e he highes in e -cache
bandwid h consump ion. Con e sely, hie a chies ha e he
lowes in e -cache bandwid h consump ion independen ly
o which consis ency mechanism is being used. We can
see in Figu e 3 ha TTLA0 is he consis ency mechanism
which consumes less in e -cache a ic. TTLA0 has a
s ong dependence on Web se e s ou side in e -cache
ne wo k because i alida es e e y eques ha caches
ecei e o a s o ed Web objec . TTLA0 p oduces high
WAN a ic. WAN a ic is impo an because i could be
he eason o a iabili y in clien -pe cei ed esponse ime.
I we gene a e less impac in WAN a ic hen i is
BW consump ion in in e -caches ne wo k
0
2
4
6
8
10
12
TTLA0 TTLA M PSM APM
Consis ency mechanisms
GB
Hie a chy
Dis ibu ed
H1
H2
ADD
BW consump ion in WAN
0
5
10
15
20
TTLA0 TTLA M PSM APM
Consis ency mechanisms
GB
Hie a chy
Dis ibu ed
H1
H2
ADD
P oceedings o he Fou h Mexican In e na ional Con e ence on Compu e Science (ENC’03)
0-7695-1915-6/03 $17.00 © 2003 IEEE
Au ho ized licensed use limi ed o: IEEE Xplo e Cus ome . Downloaded on Oc obe 7, 2008 a 8:19 om IEEE Xplo e. Res ic ions apply.
PNC
possible o p e en esponse ime peaks, p e en ing clien s
om pe cei ing pa hological esponse ime. Figu e 3
shows ADD a chi ec u e wi h compe i i e bandwid h
consump ion o bo h in e -cache links and WAN links.
E e y consis ency mechanism implemen ed in e e y
coope a i e Web caching a chi ec u e was con igu ed
acco ding o men ioned pa ame e s in e e y o iginal
coope a i e Web caching a chi ec u e p oposal. In Figu e
4 he numbe o ecei ed “s ale” Web objec s by clien s is
compa ed. As we can see, i ou in e es is ge ing s ong
consis ency, he consis ency mechanism o be chosen is
TTLA0 (i is no impo an which ype o cache
o ganiza ion is implemen ed). PSM will be he nex
op ion.
0
500
1000
1500
2000
2500
3000
3500
4000
4500
TTLA0 TTLA M PSM APM
Consis ency mechanisms
Hie a chy
Dis ibu ed
H1
H2
ADD
Figu e 4. Numbe o “s ale” Web objec s ecei ed by
clien s
We measu e he bene i deg ee (i any) ha is p oduced by
he coope a i e caching a chi ec u e and i s consis ency
mechanism compa ed wi h a sys em ha does no use
caches. I was aken he sum o he esponse imes
ob ained ep oducing he ace ile in each one o he
sys ems as a measu e o compa ison.
0
1
2
3
4
TTLA0 TTLA M PSM APM
Consis ency mechanism
Speedup
Hie a chy
Dis ibu ed
H1
H2
ADD
Figu e 5. Response ime speedup o di e en
con igu a ions (PNC: P oxy No Cache
Figu e 5 shows he deg ee o gain (speedup) ob ained by
each coope a i e a chi ec u e and i s espec i e
consis ency mechanism. We ha e called he a chi ec u e
ha does no use caches “P oxies No Caches” (PNC). A
speedup equal o 1 indica es ha he coope a i e caching
a chi ec u e and he alida ion mechanism ha i uses is
no gene a ing gains. Speedups g ea e han 1 mean he
coope a i e caching a chi ec u e ob ains some gains. As
we can see in Figu e 5, he bes esul s a e in dis ibu ed
and ADD a chi ec u es, no ma e which consis ency
mechanism is implemen ed. They ha e speedups highe
han 2.5 and 3.5. In Figu e 6 all he a chi ec u es e iewed
in his wo k appea con as ing speedups wi h he
bandwid h consump ion ha hey gene a e. The p oximi y
o he ideal ma k ep esen s g ea e bene i wi h smalle
cos . ADDPSM, ADDAPM, DAPM, DTTLA y H2M
con igu a ions a e he bes al e na i es when he in e es is
ocused on he ne wo k in as uc u e. We ha e done mo e
expe imen s scaling all caching a chi ec u es o mo e
caches ( o mo e han 20 caches, hese expe imen s a e no
included in his pape ). I we scale (including mo e
caches) he coope a i e caching a chi ec u e, we can see
ha dis ibu ed caching a chi ec u e was he mos a ec ed
in e ms o high esponse ime and dec easing speedup.
When a dis ibu ed caching a chi ec u e g ows, he
numbe o in e -cache messages g ows linea ly. Tha
means ha he bandwid h consump ion g ows as well, and
he in e -cache pe o mance goes down.
Figu e 6. Speedup s. BW consump ion
6. Conclusions
This pape p esen s a Web objec dis ibu ion a chi ec u e
which is esilien . Tha means, i he ne wo k beha io
changes, and he e is a need o make some changes o he
Web caching a chi ec u e o make up he si ua ion, he
changes could be done and Web caching a chi ec u e s ill
o e s bene i s wi hou conside ably a ec ing he ne wo k
esou ces. The app oach ha we ha e implemen ed was o
e alua e he mos popula coope a i e cache a chi ec u es
combined wi h he mos used consis ency mechanisms.
A e analyzing he esul s o hese e alua ions, we ha e
buil a Web caching a chi ec u e ha we called ADD. This
caching a chi ec u e akes ad an age o he bes ea u es
de ec ed in o he caching a chi ec u es. Tha is why his
caching a chi ec u e ob ains be e pe o mances e en i
some changes in con igu a ion ha e o be done. ADD
a chi ec u e is based on a coope a i e cache hyb id
o ganiza ion including an in e -cache Web objec
disco e y mechanism based on di ec o ies (diges ). The
ideal consis ency mechanism o his a chi ec u e is
in alida ion mul icas wi h li e signaling o keep he s a e
o pa en -son cache links. I some changes in ne wo k
beha io occu , i could be necessa y o change he
0
0.5
1
1.5
2
2.5
3
3.5
4
0 5 10 15 20 25
Bandwid h consump ion (GB)
Speedup
21.29807617
22.15881482
22.41816168
22.37229272
22.37229272
18.47791563
18.3567164
18.27099728
18.28478654
18.28478654
22.5085245
22.49942255
22.8167308
22.71748725
22.71978826
13.2000845
13.41183903
13.12447315
13.07397105
13.0726632
11.4811294
12.22968144
12.2376863
12.14751755
12.14953414
JTTLA0
JTTLA
JM
JPSM
JAPM
DTTLA0
DTTLA
DM
DPSM
DAPM
HTTLA0
HTTLA
HM
HPSM
HAPM
H2TTLA0
H2TTLA
H2M
H2PSM
H2APM
ADDTTLA0
ADDTTLA
ADDM
ADDPSM
ADDAPM
Ideal
0
0.5
1
1.5
2
2.5
3
3.5
4
0 5 10 15 20 25
Bandwid h consump ion (GB)
Speedup
21.29807617
22.15881482
22.41816168
22.37229272
22.37229272
18.47791563
18.3567164
18.27099728
18.28478654
18.28478654
22.5085245
22.49942255
22.8167308
22.71748725
22.71978826
13.2000845
13.41183903
13.12447315
13.07397105
13.0726632
11.4811294
12.22968144
12.2376863
12.14751755
12.14953414
JTTLA0
JTTLA
JM
JPSM
JAPM
DTTLA0
DTTLA
DM
DPSM
DAPM
HTTLA0
HTTLA
HM
HPSM
HAPM
H2TTLA0
H2TTLA
H2M
H2PSM
H2APM
ADDTTLA0
ADDTTLA
ADDM
ADDPSM
ADDAPM
Ideal
P oceedings o he Fou h Mexican In e na ional Con e ence on Compu e Science (ENC’03)
0-7695-1915-6/03 $17.00 © 2003 IEEE
Au ho ized licensed use limi ed o: IEEE Xplo e Cus ome . Downloaded on Oc obe 7, 2008 a 8:19 om IEEE Xplo e. Res ic ions apply.

consis ency mechanism. Tha si ua ion is no going o be
all ha impo an because we us he cache sys em will
s ill o e some bene i s. This is a e y impo an
ad an age o his a chi ec u e. Du ing his wo k, we ha e
de eloped a eliable simula o which is a use ul and
dependable ool o an a p io i e alua ion o Web caching
a chi ec u es and is a good con ibu ion o his pape .
Re e ences
[1]S.G. Dykes, C.L., K.A. Robbins, and C.L. Je e y, “A
Viabili y Analysis o Coope a i e P oxy Caching”, in P oc. o
he IEEE In ocom 2001
[2]"In e ne Cache P o ocol (ICP) e sion2" a ailable a :
p:// p. edi is.es/docs/ c/21xx/2186
[3]Ko ea Na ional Cache, a ailable a : h p://cache.kais .ac.k
[4]Na ional Labo a o y o Applied Ne wo k Resea ch
(NLANR), “I cache p ojec ”,
a ailable on line a h p://i cache.nlan .ne /
[5]Spanish academic ne wo k (Red I is) a ailable a :
h p://www. edi is.es/
[6]P. Rod iguez, C. Spanne , and E. W. Bie sack, “Web Caching
A chi ec u es: Hie a chical and Dis ibu ed Caching”. 4 h
In e na ional Web Caching Wo kshop, San Diego, USA. 31s
Ma ch –2nd Ap il, 1999
[7]N. G. Smi h, “The UK na ional Web Cache – The s a e o he
a ”, Compu e Ne wo ks and ISDN Sys em, 28:1407-1414,
1996
[8] V. J. Sosa, L. Na a o, “In luence o he Documen
Replica ion/Valida ion Me hods on Coope a i e Web P oxy
Caching A chi ec u es”. Communica ion Ne wo k and
Dis ibu ed Sys ems Modeling and Simula ion Con e ence
CNDS’02 in WMC’02. Pags. 238-245. ISBN: 1-56555-244-
X. San An onio, Tx. USA
[9]V.J.Sosa, L. Na a o, "A New En i onmen o Web Caching
and Replica ion S udy", Wo kshop o dis ibu ed and pa allel
sys ems (WSDP'00). Uni e si y o San iago. No . 13-18
2000. Chile. ISBN 956-7069-53-0. CD-ROM
[10]R. Tewa i, M. Dahlin, H. M. Yin, and J. S. Kay, “Design
conside a ions o dis ibu ed caching on he in e ne ”, in
P oc. o he In ’l. Con . On Dis ibu ed Compu ing Sys ems
(ICDS’99)
[11]Vi ual In e Ne wo k Tes bed P ojec . A ailable a :
h p://ne Web.usc.edu/ in /
[12]A. Wolman, G. Voelke , N. Sha ma, N. Ca dwell, A. Ka lin,
and H. Le y, “On he scale and pe o mance o
coope a i e Web P oxy caching”, in P oc. o he 17 h ACM
Symp. On Ope a ing Sys ems P inciples, Dec. 1999
[13]H. Yu, L. B eslau, and S. Shenke , "A Scalable Web Cache
Consis ency A chi ec u e". SIGCOMM99. olume 29,
numbe 4, Oc obe 1999.
h p://www.acm.o g/sigs/sigcomm/sigcomm99/pape s/session
5-1.h ml
P oceedings o he Fou h Mexican In e na ional Con e ence on Compu e Science (ENC’03)
0-7695-1915-6/03 $17.00 © 2003 IEEE
Au ho ized licensed use limi ed o: IEEE Xplo e Cus ome . Downloaded on Oc obe 7, 2008 a 8:19 om IEEE Xplo e. Res ic ions apply.