You are on page 1of 25

Cloud8ank A CoSurlven

ComponenL 8anklng lrameWork for


Cloud CompuLlng
Croup l
2010 29Lh lLLL lnLernaLlonal Symposlum on 8ellable ulsLrlbuLed SysLems
W A|m
bulldlng hlgh quallLy cloud appllcaLlons
W Cb[ect|ve
quallLy ranklng of cloud componenLs
ueLecL Lhe poor performlng componenLs
W 9ed
A CoSdrlven componenL ranklng framework for
cloud appllcaLlons
lnLroducLlon
W Cloud appllcaLlon ls composed by a number of
dlsLrlbuLed componenLs
W Some cloud componenLs are reusable by oLher
cloud appllcaLlons
W lnfluenced by Lhe communlcaLlon llnks Lhe
componenL quallLy ranklng cannoL be reused
dlrecLly Lo anoLher user
W ersonallzed componenL quallLy ranklng ls
requlred
W ead|ng |ndut|e ,S Coogle l8, Amazon
W -onfuncLlonal performance of cloud
componenLs are descrlbed by CuallLyof
Servlce
W 8y replaclng or lmprovlng Lhe deLecLed low
quallLy componenLs Lhe overall quallLy of
cloud appllcaLlons can lmproved
Challenge
W ComponenL quallLy ranklng of a user cannoL
be Lransferred dlrecLly Lo anoLher user
W 1he locaLlon of cloud appllcaLlons are qulLe
dlfferenL
W ulfflculL Lo evaluaLe by cloud appl deslgners
W SoluLlon
personallzed componenLs quallLy ranklng
W LvaluaLe aL userslde and rank based on CoS
AuLhors ConLrlbuLlon
1 CollaboraLlve CoSdrlven ranklng framework
Cloud8ank
Achleve personallzed componenL quallLy ranklng
2 Cloud8ank compared wlLh oLhers

,oLlvaLlng Lxamples
- ersonallzed componenL quallLy ranklng
SelecL Lhe opLlmal cloud componenL from a seL of funcLlonally equlvalenL
componenL candldaLes
lor Lhe componenLs wlLhln a cloud appllcaLlon Lhe appllcaLlon deslgner wanLs
Lo ldenLlfy Lhe well performlng componenLs from user perspecLlve
SysLem ArchlLecLure
ComponenL usage Lxperlence Sharlng
W eve|de et|e good componenL
capablllLy
W |ent|de et|e reallsLlc
performance
CollaboraLlve CuallLy 8anklng
lramework
ank|ng |m|a|ty mutat|n
1 CalculaLe slmllarlLy of users wlLh currenL user based
on Lhelr ranklngs on Lhe commonly lnvoked
componenLs
2 kendall 8ank CorrelaLlon CoefflclenL (k8CC) Lo
evaluaLe Lhe degree of slmllarlLy
Where - no of componenLs C no of
concordanL palrs b/w Lwo llsLs and u no of
dlscordanL palrs and u v are Lwo users
11
kendall 8ank CorrelaLlon CoefflclenL
W kendall 8ank CorrelaLlon CoefflclenL (k8CC)
subseL of cloud componenLs commonly lnvoked by
user and user
CoS value of componenL observed by user
lndlcaLor funcLlon
u
W Slm(uv) 11
1 when order of user u ls exacLly reverse of user
v
1 when order of user u ls equal Lo order of user v
| | 2 lnLersecLlon for slmllarlLy
componenLs
CollaboraLlve CuallLy 8anklng
lramework
|nd |m|a ue
1 Lmploy only Lhe Lopk slmllar user for marklng
ranklng predlcLlon and excludes users wlLh ve
correlaLlons
2 SeL of slmllar users
SeL of Lopk slmllar users Lo user u
Lxcludes Lhe
dlsslmllar users
CollaboraLlve CuallLy 8anklng
lramework
9efeence unct|n
1 1o presenL Lhe quallLy prlorlLy of Lwo
componenLs (l and [)
2 user preference on Lwo cloud componenLs whlch
have been lnvoked prevlously
3 user preference on a palr of componenLs LhaL
has noL been boLh lnvoked by Lhe currenL user
CollaboraLlve CuallLy 8anklng
lramework
eedy Cde Ag|thm
1 1o rank Lhe employed and unemployed
componenLs based on Lhe preference funcLlon
2 use pasL usage experlences of Lhe oLher slmllar
users
3 Clven a preference funcLlon we wanL Lo choose a ranklng
of componenLs LhaL agrees wlLh Lhe preferences as much
as posslble
4 Cb[ecL funcLlon
W @aget Lo produce a ranklng LhaL maxlmlzes
Lhe ob[ecLlve funcLlon
W 9|be aach search Lhrough Lhe
posslble ranklngs and selecL opLlmal ranklng
LhaL max Lhe value funcLlon
W 9bem n! posslble ranklng for n
componenLs
-hard problem
So proposed udank Ag
Cloudrank AlgorlLhm
8ank Lhe employed cloud componenLs ln
L based on Lhe CoS values
1|L| ranklng
CalculaLe Lhe preference value
Larger Lhe value lesser Lhe preference
8anked from hlghesL poslLlon Lo lowesL
poslLlon LhaL has max preference value
updaLe by Lhe correcLlng ranklngs of
employed componenLs ln L
,ake sure LhaL L ls ranked correcLly
LxperlmenLs
W WSu8LA, Web Servlce CoS uaLaseL
100 servlce componenLs and 130 dlfferenL users
13 mllllon lnvocaLlons
8esponseLlme and LhroughpuL CoS values
W LvaluaLlon maLrlx
-uCC -ormallzed ulscounLed CumulaLlve Caln
Larger value sLands for beLLer ranklng performance
dlscounLed cumulaLlve galn (uCC) -ormallzed uCC (nuCC) ldeal uCC
(luCC) and p Lopk ranked componenLs
LxperlmenLs
W 00 Lo 10
W Larger value Lhe beLLer ranklng accuracy
W p 1 Lo Lhe number of componenLs
erformance Comparlson
@abe ND 9efmance ma|n f @hughut
UV userbased collaboraLlve fllLerlng meLhod uslng vecLor SlmllarlLy U9 userbased collaboraLlve
fllLerlng meLhod uslng erson CorrelaLlon CoefflclenL IV lLembased collaboraLlve fllLerlng meLhod uslng
vecLor SlmllarlLy I9 lLembased collaboraLlve fllLerlng meLhod uslng erson CorrelaLlon CoefflclenL UIV
userbased and lLembased vS UI9 userbased and lLembased CC
lmpacL of arameLers
W Imact f @k (@abe %
1radlLlonal Lopk algorlLhm
W uecreases when k lncreases from 1 Lo 10
Lnhanced Lopk algorlLhm
W CbLalns sLable -uCC performance and ouLperform
LradlLlonal Lopk algo
W Imact f mat|x Den|ty (@abe %
,aLrlx denslLy lncrease lncreases Lhe ranklng accuracy ln
boLh greedy and Cloud8ank
lmpacL of arameLers
Concluslon and luLure Work
ncu|n
ldenLlfy Lhe need for componenL ranklng ln Cloud
CompuLlng
ropose a personallzed componenL quallLy ranklng
approach for cloud appllcaLlons
ConducL experlmenLs uslng realworld CoS daLa
utue wk
lnvesLlgaLe more CoS properLles
erform furLher experlmenLs ln larger scale
SuggesLlons
1hank you!

You might also like