Cache Management for Mobile Databases: Design and Evaluation*

Boris 1 ~Chan
. ;\ntonio Si Hong V. Lcorig
DepartInerit of Computing
The Hong Iiong Polytechnic University
Hong Kong

Abstract cast paradigm has heen addressed in [ I , 2, 8, 12 The

C o m m u n i c a t i o n between mobile clients ~ 7 1 ddata- performance nf mobile caching will be evaluateJ.\ .'l a a
base servers in a mobile computing enuironment i s tiin detail simulation model.
wireless channels with low banduizdth and low rrliabil- The remainder of this paprr is organized as fol-
i t y . A mobile client coiild coche i t s frrquenthj a c c r s e d lows. In Stxtiori 2. wr survey previous work on caching
database itPr1i.s wit0 tt.5 local storage t n order to i m mechairisiris. Tht, design and implementation of var-
proi~cperfomiance uf databasr qiierirs and ouailability ious cache rnanagenient issues for niohile caching are
of database i t e m s for per?, processriiq r l i ~ r - z n gdisroii- dc:scrilieil i n Scction 3. In Scxtiori 4. we present the dc3-
nertion. In this paper. we describe o mobile c a r h i y l sign i j f n u r siuiiii.itiori inoilel. Sectinn 3 presents simw
mechani,sm f o r a mobile f!nVl~~J7i7~ierLt I d i ~ Z Z i 7 i gpuirit- nf our reprrscntativc: experimental ri'sults. Finally, A'?
to-point i:urnmunzcation paradigm. hi particulor. IUC offer brief i:nncluding remarks i n Section 6.
investigate issiies on cachirig gra7ii~larity, mherence
strategy, and replacement policy of mobile caching. 2 Related Work
V i a a detail simulation model. we compare oiir pro- Caching mechanisnis in ronvrntional client-server
posed caching mechanixrn with cuniientionnl ones and environment arc usually page-hased [GI. primarily lie-
discoiirr that o i i r mobile caching mechani.sm outper- cause the ovrrhrad for transmitting one item or a pagr
f o r m s eonventiunal ones in must sttautioris. is similar in ronventional client-srrvrr environment.
Pagr-hnseil caching mechanisms require a high dcgrre
1 Introduction of locality nuinrig the items within a page to he etfec-
In a mobile environnient. a set of datahase%srrwrs tivr [-I].
disseniinat,rs rlataliase inforinatioli v i a rvirclcx chan- In practicr. database itriris requested by different
nels t o multiple mobile clienh. Drpmdiiig on the mohile rlieius via drdicated channc~lswill iliffrr much
affinity of irrdivir1u;il ilat.al~;tsritrms. itcms uf intrr- in a point-til-point nioliile ~ n ~ i r o n r r i t ~ut.herwise
rit: thr
Cst t o must niohilc ciicrits s h i i i i ~ dbr t)ro;idcast from mort' rffrct i r r itnil scalable broa11i:ast paradigm should
a ilatatiasr scrvi'r to niiiltiplr clicnts a-hile itrms (if be employid t o liroailcast itrnis uf common intrr-
interest to single client shni11i1 tlr rlissrmiiiated ovpr est, [!), 151. .I ~iliysicalnrganization that favors the,
dedicated charmeis uii dernarrd [ O ] . Sincp ii wireless locality exhiliit~dliy m e rlicrlt might rrsult in poor
channel siitkrs from a low handwidtli of 19.2 Khps prr loralit? for another. Datahase itrms Xvithin a pagc at
channel and is also vulnerable tu friquent disconnec- r i database servrr thus 1,arely i~xhiliit.a n y degrre of
tion, it is important to cache frequently accessed itmis locality. Furthrrmore, mobile clients are powered by
(hot spot) into a mobile client's loc;il storage. This short-life batteries [9]. Caching a page will result in
improves the perforniance of database queries and the wasting of energy when t h e degree of locality is low.
availability of database items for qucry processing dur- The overhead of transmitting a page nver a low band-
ing discunnection. widt.h wireless channel would be too expensive to be
A caching mechanism is characterized by its caching justified. It is, therefore, necessary t.o consider caching
granularity, cache coherence strategy, and cache re- at a smaller granularity in this context.
placement policy. Conventional caching usually re- Cached items will become uut-dated when the base
quires a quite stable network; a reasonably high trans- items (copies residing at the database server) are up-
mission bandwidth. and a high degree of locality dated. A cache coherence strategy must be provided
among database itenis residing within a d a t a page at to update the cached items at each client. Conven-
the database server [6]. These rnnflict with t,he char- tional cache coherence strategies require the server to
acteristics of a mobile en\-ironment. notify all relevant clients whenevrr a n itern is updated.
In this paper. we irivcstigatr t h r ;ihiive t h r w issurs Each mobile client. however. connects or disconnects
of a caching mechanism in a moliile rnvironmait utiliz- from the wireless netwnrk freely arid frequently. It is,
ing poirit-to-point communication paradigm. rrfrrrril t.herefore. not feasible for t h e servrr to keep track of all
to as mobile caching. Caching rriech;uiisni for hroad- cached copies nf individual it,rms. A clirnt should take
a more active role i n maintaining the coherence of its
'This research is supported i n part by t h v Ilorrg K o n g Poly- cached itmis and determining if a particular cached
technic linivrrsiry Rrscarch G r a n t n u m b e r ,liO/570 item should be invalidated.

54
0-8186-8289-Zf98 510.00 Q 1998 IEEE

Authorized licensed use limited to: Universidad Nacional Autonoma De Mexico (UNAM). Downloaded on February 24,2024 at 04:59:43 UTC from IEEE Xplore. Restrictions apply.
 I
In t h e Leasri fikcaching mechanism [7], each file, nism is needed t o retain the frequently rrquest,ed data-
cached in tho h z l storage of a client, is associated hasp items in a client's local storage. The niore effec-
with a pre-specified refresh time which defines the du- tive a caching mechanism in keeping the frequent,]? a r -
ration within which the cached file could be regarded cessed items. the better a query w i l l perform and t h e
as valid in the client's local storage. When the refresh morr queries could be served during disconnection.
time expiriw the client needs t o contart t h e server for 3.1 The Cache Model
an updated file. I t is, however, difficult t o determine
an appropriate refresh duration. \Ve investigate three different levels of granular-
If a mobile client can provide unbounded disk stor- ity of caching a database item in a n ol>ject-orieated
age, it can cache all database items accessed. How- database (OODB), namely, attribute caching, object
caching. and hybrid cachzng. Intuitivel>-.in attribute
ever, the available storage for caching is often lim-
ited [13]. Furthermore, caching items that will barely caching. frequently accessed attributes of database ob-
be accessed will result in a waste of energy. h cache jects are cached in a client's local storage. In object
replacement policy is needed to retain only frequently caching. the objects themselves are cached. Finally, in
accessed items for best performance. hybrid caching, only frequently accessed attributes of
those frequently accessed database objects are cached.
In [ 5 ] , various cache replacement policies for a con- This ensures that the cached attributes of the cached
ventional database system have been examined: op- objrcts will have a high likelihood to be accessed in
timal, WORST, least recently used (LRU). CLOCK, t h r future.
and least refererire density (LRD). These policies are Several drsign issues need t o be addressrd in the
all page-based. In general, t h e performanre of indi-
vidual replacement policies is sensitive t o the rharar- implementatimi. First. a cache table is needed i n each
client. t o idrntify if a databasr item (attributr or ob-
teristics of queries initiated and the application eri-
viroiiment. A general conclusion 011 the perforinancr ject) is c;diml 111 local storagtz. Second. if a client
is coiinrctt!d 1.0 a scrver. t h r client should br ablr t o
of thc rrplarernent policies rannot hi, rrroinmrndeil. retricvr t h r r;ichcd items from the local storage and
In prarticr.. thv optimal policy is often approsiniatrd thr iincai.hr~iitems froni the srrvcr. The rlirnt will
by LRU in ronventional caching IF. 161. LRU is fur-
ther generalized into LRU-I: [14] whirli idrntifics t h r only rrtritarc' t h r c;ichrd itrnis otherwise. Third, ail
Pffrct iv1, r ( h , w i i c r strategy is needrd t o maintain thc
rrplarcmrnt victim according t o the tiinr of t,llr k I h frrshrirss o f tlic, rached items arid finally. an effrrtive
prri,iou.s accrss of it page. LRU is, thus, rquivalrnt t o r q ~ l a c c ~ m r n[idicy
t ~iredst o br identificid t,o rrtain thr,
LRU-1. In a rriobilr environment. since a client might most frrqiiimtly accrssrd items in a c h r n t ' s local stor-
change its location, the set of database items in wliicli agr.
a client is interrstrtl might change over timr as wcll.
Therefore. we nrcd t o exaniirir t h r suitability of roii-
ventional rrplaremcnt policirs in this changing accrss 3.1.1 The Cache Table
pattern and to draelop other morr suitable rrplacc-
rncnt policies for bettrr performance. Tlir rarhv t a l i l r of i i mobile client is simply a mini-
Only until recently have raching mrrhanisms hren dat:h:isr.wliirli ronld PLI managed by its own ohjrct-
t 81.
invrstigatrd in a mobile databasr c ~ ~ v i r o n i n r n12. orirntrd il;itahasr maniigcnirnt systrni. In its local
In 121, an invalidation rrport is hroadcast ovrr a win= storagr. ii r l i i s s hirrarchy rootrd a t rlass Remote and
ii rIa,\s hiw;irrliy rootrd at class Cache are niaintained.
less rhannrl to inform individnal niohilc clients ahout
the invalidation of rarhed itrms. This reqnircs a nio- For each class X a t the servrr. a corrrsponding class
bilr client t o kccp tuning into the rhannrl t o invali(liitr X is c r r a t 4 ;is sohclass of Remote and a class C X
and refrrsli its rarlicd items. .A c1isr:o~inertrdc h i t is crc,atrd i ~ ssiihrlass of Cache. Remote is usrd t o
may miss the invalidation. Iri [8 , nn itrm is citiIird iiidiratr if a i i itrm is c x l i r d n . t i i l ( 3 Cache is used t o
/"
i n a mobilr client's local storagt, i t h r nnmhrr' of riwl
operations perfurrned on t h r item is grratrr than t h r
'
hold t h i , r;irhriI valur of aii item. Colli!ct,ivcly. thcz
1iirr;irrhy rootid a t Remote funrtioris as a rache tahle.
number of write operations performed, as is usually For riirli attritmtr. u. of class X at tlir srrver, a nsrr
tlir case. \Ve helirve that a n item shonld hr carhrd drfiiir(l nitthoil \vit,ti t h r sanie name. 1 1 . will br created
iii thr local storage if it is frequently arcessed rcgard- for X i n t h r rlirnt's local storage. Thc definition of
less of the access operations. Rather than disallowing inrthod (1 ir-ill t)r drscxihed hclow. In addition. an
an item to b r cached if it is frequently updated. t h r atrritmtr w i t h namr c.n is crratcd for class C X . with
caching nicchanisni should adapt t o tlir situation. t h r s i i m i ~d d i i i i t i o i i as attribute a i n t h r svrver.
To illiistr;~t(~.consider an Adiiarrced Tratieler
I n f o n n o t w ~S?,.~lcm( A T I S ) applirntioii 131 in which a
3 The Design and Implementation of groiili of t i i i i i ~ i s t swould likr t o rrtrirvc placrs of attrar-
Mobile Caching tioii ani1 ; ~ r ~ ~ ~ ~ i i i i i i ~information
~ i l a t i o i i froni a ilat,;ihase
hlobilr, rlirnts suffer from n r t ~ x i r kdisrnnnrrtioii s r r r c ~via ;I 1 ~ r t a h l ccompnter ~ rqriippcd witli \vir<,-
quit.(. easily and freqiicntly. For instaiicr, a recrnt ex- Ira.; riiiiiiiiiiiiii~iitio~i iiiterfact!. Fignrr I n il11istr;rtrs a
jierienci~i n iisiiig ii rornmrrrial nirolrss rqi~ipnii~nt to sinil>liticdO O D n f i r this travrlcr inforriiat,ioii systcrii
rss onr scrvc'r data t l i r ~ ~ n g h o nour t cIri~artinrnt,al 5vIiili~Figiirr I b iiidir;itrs ttic [~iartial) str~~ctiirti
[if thr,
Iriing iiidiratr~lthat a clicmt ciirild hr discc~ri~~rctrd I w i i I diitiiIms(. niainttiinr~lby ii n i i h i l ~rlimt,. ~
from tlir airtlrss ~ictworksimply l ~ isi~liiting
y thr rlirn~ .I(>riicli<, ; i i t r i h i i t ~of an o l i j i ~ t .I. l~rlongingt o
xith s r v c ~ i i lt1iii.k offirr p;irtitioiis. C d i i n g invcli~i- r l i i s h X f r o i l l tlir strvrr, a11 objrrt is rrc;itrtl for X i n

55

Authorized licensed use limited to: Universidad Nacional Autonoma De Mexico (UNAM). Downloaded on February 24,2024 at 04:59:43 UTC from IEEE Xplore. Restrictions apply.
 Assume that the local database schema a t C is as
shown in Figure l b . Further assume that before the
query is initiated. C's local database only contains
surrogate L arid only vacancy of I is cached as in-
dicated by the shaded region (a) in Figure lb. Fi-
nally, assunie that only two uhjccts. r and y, in S
satisfy the query. When evaluating tlie where claiise
of the query, the niethod vacancy of Places t o Stay will
retrieve vacancy(r) from c.vacancy(r). .I remote re-
quest will be sent to S containing the list: (R.oid(r),
vacancy), as vacancy(r) h a s been cached. S will return
- " IL rj a list uf valurs for name(z), city(z), name(y), city(?/),
and vacancy(y) since only objects z and y satisfy the
Figure 1: .I sample ATIS databnse application query. C can then. cache name and city, of surrogate z
under c-name and c-city rcspectively, as indicated by
shaded region (b) in Figure l b . A new surrogate for
the rlicrit's local datahaw. This loral object is called ohject y will also he created in the local database. C
;IIUC;II surrogate for t h e remote (ihjwt L a t the server. can t h m cache name, city, and vacancy of y.
Each local surrogate of X will inhrrit two attributcs
from Remote: R.oid holding t.he ohjrct iilentifirr used 3.1.3 O b j e c t C a c h i n g
hy t.hc s i ~ v e rt o rcferrnce I and R-host holding t h e
;iddrt~ssof tlic servrr wherr I origirdly resides. T h e Each inobile client trnds t o liave its own srt of objects
local surrtigatc I is aililctl t . C ~ X via niultiplt: itietii- t h a t it accesses most frequently. Furthermore. a client
hership rnorlding construct (if OODD rnorlcl. The sur- might access diffrrcnt attributes of an ohject in differ-
rogate. thus, intirrits attributes rlrfincil for C X . pro- ent queries it initiates. When the server receives a
viding stor;igcs as placrholilers to carlie attributes of request. it might be worthwhile for t h e server to push
r. The value of a n attriliiite. u . of an object. r , u ( z ) , all attributes of a qualified object, r , t o the initiated
will Iir cachrd unrlrr c.n of C X . For instance. in Fig- client, thus eliminating future requests for z from the
ure Ib, one ubjrct, brlonging to class Places t o Stay is client. The client can rachc the returned attributes
caclird in t.he Iucal storage of i i clirnt. under class C X as in attribute caching. For instance.
Each ;it.t,ributc of class X in a client's local data1)ase the shaded region (c) i n Figure l b shows that all at-
is i i rnirthod. This rncthml cnc;ipsul;~tcsthe, tasks ill- tribntes of L are cached when query (2 is evaluated.
v o l v ~ din query processing. First, the initiatrd query
is scnt t o the server. Second. for rarh attrihutr, (1. re-
quired Iiy the query, it retricvcxs c.R(L), for each lural 3.1.4 H y b r i d C a c h i n g
si1rrogat.r I of class X from the local storage if ~ ( r ) In object caching, the database server will prcfctch all
has ;rlrr;idy 1)c~encaclied in c - a ( r ) uf t h e Iwnl data- attributes of i i qualified otiject. L, to the client. It is
liasc. Third. it s r n r l s an r!xisterit list of (R-oid. a ) i o very oft,tm that nut all attributrs of c. will he accrssed.
t,he srrvrr. infcirn~i~ig the srrvrr ;~liuutt h attriliutf~s
~ ~
This not only \nates the transinission b;mdwidth. but
which Iiavc, Iicrn satisfird 1or:iIly. so that they do not also occupies storage for caching other inure frequently
need t . ~lit, t.r;insrnittetl hack. Finally, iipoii r r c p i v accessed attributes. Hylirid ciiching restricts the data-
ing such ii list and evaluating the qucr>-, the s e r w r base server t u prcfrtch only those attrihutc!s uf a quali-
replies with a list of (oid, attribute value) pairs, fied object with a high likelihood t o be accessed in the
for those that satisfy the query but are not cached in future. Only attributes with access probability ahove
the dient's storage. An advantage for encapsulating a prefetching threshold. e , will be prefetched.
the cache rriorlel within a method is its transparency
to a client. By simply having the method returning 3.2 Cache Coherence
a null rrsult during disconnection. the client can con- .I cache coherence strategy usually involves cache
tinue to operate using its locally cached items. without invalidation and update schemes t o invalidate and up-
concerning if it is connected to the xireless network. date an out-dated cached item. We note that most ap-
plications in a mobile environment will generate more
read operations than write operations [a]. Further-
3.1.2 A t t r i b u t e Caching more, a mobile client usually can accept a slight rle-
In attribute caching, after the server S has evaluated grre of o u t d a t e d dat.a i n return for faster d a t a re-
the query subrnitted by a client C. S only returns trieval. We proposr a "lazy pull-based" invalidation
those attrihutes of those qualified objects that are re- approach i n which rnch client is responsible to inval-
quested by C. To illustrate. consider t h e following idate its cached itrin. In addition, we employ a n on-
OQL qiiery, Q, tieing evaluated against t h e ATIS data- dentand update approach in which a stale cached item
base (if Figure l u by C: is ouly updated !!-hen it is next accessed. We illustrate
the idea using ohjrct caching; coherrnce for attribute
select s.name. z.cit caching and h!-brid caching will be similar.
from z i n Places to tay
where r.vacancy > 0 .
l \\'e borrow the idea proposed in the Leases file
caching mechanism [il by estimating a refresh time

56

Authorized licensed use limited to: Universidad Nacional Autonoma De Mexico (UNAM). Downloaded on February 24,2024 at 04:59:43 UTC from IEEE Xplore. Restrictions apply.
 ( R T ) for each cached objert. The estimation of the mated score ~,,,+, can be computed incrementally
refresh time for a n object depends on its update prob-
ability. If an object is updated frequently. its re-
as (nn,., + % n + l ) / ( i l + 1).
The mean scheme probably does not adapt well to
fresh time vill be shorter. Consider an object, z, changes in access patterns since every single trace from
cached in a mobile client, C . The refresh time for the beginning of the access history remains in effect. A
I. RT,, indicates the duration that z could be cached better approach is t o use a window for the statistical
in a client. while remaining valid. In attribute and measures. Each object is associated with a window
hybrid caching, since individual attributes of x are of size t i - , storing the access time of TI. most recent
cached, the refresh time is associated with individual operations. The cached object with the highest mean
attributes. \Vhen C accesses z in response t o a query, arrival duration within the window is replaced. This
it checks the validity of z by determining if RT, has is known as the window scheme, whose effectiveness
expired. If RT, has expired, x becomes stale. The ex- depends on the window size, H.. With a window size
istent list t o the server will then not contain an entry -i 53: 1
for x. When the server returns z, in response to the ti;, the new score hl,,,,, for object z is computed as
request from C , it will estimate the refresh tlme for
z and the new estimate will be sent along with z to
-iW=l
AlZ," +
(Afz.n+i - hlr.n-W~+I)/&V~. A problem for
the window scheme is the amount of storage needed
C. In other words, the refresh time for an object is in maintaining the li; intermediate values.
updated dynamically whenever the object is sent t o a
client from the server. This approach does not require
To avoid the need of a moving windorv and to adapt
a client to be always connected in order to invalidate quickly t o changes i n access patterns, our third scheme
I. Furthermore, if z is never acressed again, 2 will
assigns weights to rach arrival duration. such that re-
never be refrrshed even after its refrrsh time expires. cent durations have higher weight,s and t h r weights
tail off as tlic durations become aged. TIIP replace-
To estimatr the refresh time for each cached object. nient score is th? Ezponentially Weighted Moving A w
I. the. iutrr-arrival duration of consecutivr w i t r oper- ernyr of arriv;d durations and is callrd the E W M A
ations. d , . oii 5 is niaintained. The mean. d,. and t h r schrinr. .A parainrter to EIV11.i is t h r weight. n.
standard deviation, s I ,of thr durat,ions-arr computed. wliidi rangrs from 0 to 1. The current duration re-
The refresh timr of z is estimated as d , &sr. The + ceivrs a \wiglit of 1; the previous duration receives
value of govrrns the frrquency of refreshing 2. It a n-right of (1: thc next prcrious duration receives a
indicates the degrer of deviatiorr a client can tolerate xeight of o1 aiid so on. For object x with weight
on z. The smaller the value of fir, thr smaller is the -(e, 1
refresh timr and the higher the possibility that a client the estiinatctl s m r e for z, Af,,,,,,. when adding a new
needs to rrquest z when it acrcsses z in a query. mrasurr. A,,.,,, I , could be computed incrementally as
-(0.1
We define the notion of an error in accessing an ob- nrhlz,,E+ ( 1 - i t z ] ~ 4 f =,,,+ [ . The EIYll.4 schcme is. in
ject. Assume that a niobilr client refreshes an object, effrct.. sinii1;ir to tlic LRD scheme described in [SI.We
I. at time t l ard t 2 . Betwer~it l and 1 2 , the client ctiar;wti.rix, Ihi, prrfornlancc difference betwreri LRD
might issue read oprrations on its local cached copy and E\V\lA i i t Srction 5.
of z. For each read operation, r r . initiatrd brtwecn ti
and t 2 : if the servrr performs a write oprration. t u r . 4 The Simulation Model
or) 2 . hrforr r,. the value of z used by the d i m t will
be inconsistrnt with t h r actual value maintained at, T h r rrpvrirrler1I.s presented here are organized
thp server and trcncc. the read operat.ion rz rtwlts in around tlirr,. otijrctives. First. we would likr to study
an error. This definition is used ill i:h;ir;irterizing t h r tlir prri~iri~i;inrr differences among attribute caching.
performancr versu'i cotirrmce tradroff in Srrtion 5. rrbjrct cacliing. and hybrid raching. Secoud. we would
likr t o coniparv our rcplaceincnt policies with conven-
3.3 Cache Replacement tioii;il onrs. Tliird. wr would like to study the effec-
tivcanrss of o u r coherence strategy in maintaining the
IVe consider a spectrum of replacemetit poliries that frrshncss of i,;ichc~ldatabasr items.
adopt the a c r w probabilities of database items as an Thc, siniu1;it ion model is irnplenientcd using CSIhl.
indicator for the necessity of replacing a cached item. It consists of onis OODB server and 10 mobile clirnts.
We illustrate the idea in the context of objert caching; T a u channels. rach having a typical wireless band-
replacement policies under attribut,e raching and hy- \vidth of 19.2 libps, are shared among all clients t o
brid caching are similar. comniuiiicatr with t h r server. Onr channel is used
For each object. a replacement score indicating the for u p s t r e a ~ nqi~rrieswhile the other is used for down-
prediction of its acress probability is rstiniated. The strrani rrsnlfs. TIir OODB has one class. Root. a i t l i
higher the score. t.he higher the estiiiiiitrd access p r o b 2000 oljjrcts. ;ill rrsiding i n t h r srrver's disk initially.
ability, and the Iowrr is its rhance of being rrplared. Eacli objrct c i i n ~ a i n s9 attrihntrs of primitivi,-v;ilued
The most straightfurxvard way t o cornputW t h r typrs ; u r d 3 oii<'-to-onrrelationships t.u aiiothcr object
score for each object is by rneiisuring the itic'an inter- of Root. Eacli oI1,jrc:t has a sizr of 1024 bytes.
operation arrival durntioii. T h r ciiclied ob,iect with \Via ~nodi:lnmnory caching a t both servrr and
the highest niran a r r i v d dur;ition is riylacril. IVr call cliriits. in aililitiolial to storage caching a t c ; ~ l client.
l
this thr mrmt srilrini,. . i s s n r n ~that t h r tiwv dura- L R P is cmipli~yulfur buffer iiiaiiagrlnent at. tlir srrver
tioil for objrct x i s ,\, iuid t h r mtini;itrd scow and tlic clirriti sinw iiicrnory buffer rq)I;iwiurnt is ini-
-
for the prcvious II Ineiisurrs is Af,,,,. T h r I I C , ~ rsti- plcriiciitril by the oprrating swtrni. Thc, Iiatida-idth

57

Authorized licensed use limited to: Universidad Nacional Autonoma De Mexico (UNAM). Downloaded on February 24,2024 at 04:59:43 UTC from IEEE Xplore. Restrictions apply.
 of disk is srt ti) 4 l l h l l ~ p sto i i i o i l t ~fast.
l SCSI disk r v h i k to 1I):OO (with inciiii :irri\d riitc~of 0.027). inodeling
that nf riirinory is sct to 10l)XIlips. Thc size nf nieiii- the biirsty traffic for taking otf fsorn rvnrk. The two
ory butfcr at thr st3rvi.r is 25% o f t h a t of the d;itabase. non-hiirsty periods will atjsorh 211% o f t lie querics: the
i.e., 500 ohjt,cts. Tho size uf stortigr cache ;it each period from 10:00 to 16:00 hiis a ineiin arrival rate of
molilr client is 20% of t h a t of the datal~asc.i.r.. 400 0.005 and t h r one from 1I):iIO to 0 i : O l I hiis ;in
otijects. The incniory huffcr at r w h rlitwt ciiii hold lower mean arriviil rate nf 0.0013. m i d d i n g working
30 ohjects. \Yv mudel a scenario i n whicli tlie server hours arid happy/rest liours rrsprctivrl>-.
has a large hutkrr coniparcd w i t h t t i ~(1at;il)ase. The Tlic sixth iliniensioii studies the e f f m of update
storage iit racli clirnt is uf a nioilr,r;itc size \vliilc! its prnhability of an objcct. U . on thc caching perfor-
niemory liuffc>r is the sni;illcst. mance. For each object accessed hy a query. there will
T h r size of a remote request i i i i i l i i r q ~ l yriiessagc be a probability U that it will he updatrd. All selected
depends on the caching granularity. Ijut huth have a n attributes of an object for update will be modified.
11-byte header including an IP i i i l i l r i w arid ii CRC Filially, we experiment with rhe performance of
for rrror detection. Each qurry has ;I low selectivity a caching schenie during disconnrction. siiice discon-
of I%, Le.. 20 ohji,cts. since the ainoiiiit of iiifnrma- ncctcrl operation is an important aspect in a mobile
tion rtyuested for cach qucry in a niihili, i~nviroiinieiit environment [ l l ] . Here. we mould like to study the
should be rrlativcly small in general. performance from two different perspectives. First,
We oxperinwrit riioliile cacliing ;iIong s r w i dinirn- we vary the duration of disconnrction nf each client.
sioris. ;\long the first dirncnsion, wo irivcstigatv thrr 2). Second. we vary the number of clients t h a t are
rffect uf caching granularity, C (iittriliute ciichiiig or disconnected, .V.. Each experimmt is conducted for a
.IC,uhjcct mching o r O C iind hybrid r;whing o r HC). prriod of 4 sirnulatcd days, i.e.. 96 simulated hours.
on t.hc prrfnrniancv [ i f rriihile mctiiiig iis cotiipiirrd The nvrr;ige metrics are ineasurid. Thcx staridard de-
with no racliirig (NC). viation of our iniwurimeiits is fouiirl t o be very small.
-\lorig the sccoiid (liini~risinri.!vi' vqwsiiiimt x i t h tlius yielding w r y tight cnnfidiwci~intesvals.
two heat distrilnitions. A. ;iinong ttw rlatabasc oh-
jects. First, we adopt ;in 80120 ride i n ildiriiiig the hot 5 Performance Evaluation
objects: 20% of tlie objects are pickivl randimily such We characterize the perforniniire of a caching
that 80% uf the accesses fall withiii this s i i l m t . This sclienir hy three nietrics. Avc.rage cache hit rntio of
is called tlie skewed heat (SH) p;ittr,rii. \Ye rnsure all mobile clients measures the percentage of accesses
that the hot ohjects of rwch clirnt i i w i i o t i<leiitic;il. that c a n he satisfied hy rrtriwing it locally unexpired
Second, we allow the 20%)hot ohjcrts t i i lx, riwlrcted caclicd database itrm (attribute or object). .Iverage
aftt>r every A<- r l u c ~ i i , ~This . is ciiIIwI I Iw c h n n g i n g respotise time of all mohile clients rneasiirrs t.trc aver-
skewed hrnt ( C S H ) ptittrrii. age time spent (in secoiids) from the iiiomrnt a quc:ry
Along the third iliinensiori. we rxpi,riiiiimt X L - i t h t,wo is issued tn the moment the. rrsults tn t h r query arc
different typrs nf query. G.In arsocz~itnrf~ p r r y (.4(J), generated. either hy nsing locally iinexpircd database
Q,, primitive-valued attritnitras nf r x l i si~lwtcrlohjrct itrnis o r rrniot,r riwilts from thc, srrvw. Fin;illy, error
will he acrrssril. The attrilnitr~sof v x l i o h j i x r t sr- rate measures the prrcrntagc (if r f v d ( m o r s the c1icnt.s
lected by a query follow a nriiforrii i k t w d ~ l i s t r ~ l ~ n t i ~ ~ rnroiintcrtd
~i. (sre Srctiuii 3 . 2 ) . Six w t s of rsperiinerits
All ;ittrihiitt>s iiavr i i non-zeru iiccess priitxihility. 111 arr conducted. Thc p;ir;urictcr srttirigs of tlie erprri-
tiavigatiorinl p e r ! / (.Nc)). for eiich sr.lc.i.tid o t i j t ~ t G,, . nients presrntril arr snniniarized i n Tiitk 1.
inter-ohjrct relationship will Ix triivrrsi,<iin ad~litinniil
t o accessing Q , I)riinitivt,-ralned attriliutrs of the o b
ject. \Vhcn travcrsing the relationship. Q,, attributes
of t,he rrlatrd object will ais0 he rrfcriwced.
Along the fourth dimension, wr mpcriment with
different repl;icrniriit policies a t a (.liiwt's lnral stor-
age, 'Rd,,r. \$'e wperirnent with LRL. i.e.. LRU-1.
LRU-k. LRD. mean. window. and EXVX1.A schemes.
We specifically select LRL. LRU-k. and LRD schemes
for comparison with our schemes mainly hecausr LRC
is the most popular approach and LRD is intuitively
similar to our EtVA1.I scheme. LRU-A has also breri
shown in [14] to have a better perforrriaiicc than LRL
when access patterns on objects change periodically.
We experiment the fifth dimension w i t h t x o arrival
patterns of queries. P . \Vith Poisson arrival. the iiiran
arrival rate of queries is set t o 0.01 [lo]. \Vith Bursty
arrival. w itio~lt'l the veliicte traffic pattern [ll)]. with
the same average rate of 0.01 ovcr one day. Of all the
qurries initiated within 24 hours. 80% of the queries
are clustrrrd withiii two hursty periods: from 07:00
to 10:OO (witti mean arrival rate uf 0.037). inodr3ling
the hursty traffic for traveling to a o r k . and from 16:OO

58

Authorized licensed use limited to: Universidad Nacional Autonoma De Mexico (UNAM). Downloaded on February 24,2024 at 04:59:43 UTC from IEEE Xplore. Restrictions apply.

c
 ~~~
IUI

~~
--
-u*

~~
-9 "4

~~
.
9

~~
-9
9-
~~~
-4 WUI .i3 -y.l
9-
wk.. "4

~~

Figure 2: Performance of storage caching schemes uersus no

5.1 Experiment #1 of i i i i ubjert. tlir overhead for transmitting cold at-

tribiitt3s uf hiit objrcts will be wasted. The ovt!rlicad
Our first set of nprriments compares the perfor- savrd clur to t h r high cache hit is oHsrt Iiy the F ~ P I I
mancr of AC, OC. and HC with the base case, i.e., Iiiglirr transniission delay d u r to the low bandwidth of
NC. In HC. tlir, prrfrtrhing tlirrsliold, e , is set t o two xirrlrss clia1i1ic~l.HC performs vrry well. ;ichiering re-
standard deviations. a, below the mean access rates spoiisr timrs as low as thosr of AC and carlie hit ratios
of all attributes, 1'. T h r basr c a r is achirvcd by dis- closr t o t l i o s i ' of OC. Prcfetrhing in HC apprars to he.
ahling st,oragr raching at each client; only memory "intvlligrnt" , in that most of thc attributes prrf~tched
caching employing LRC at the server arid each rlient will. most 1ik1,ly. be accrssed i n ttic futurc.
is enabled. For storagr caching. E\VhlA with CI = 0.5
(EWMA-0.5) is euiploye~lfor Rdlst. T h e number of CSH acrrss pattern results in slightly 1ow.t-r hit ra-
clients is fixed a t 10. and update probability U is 0.1. tius aiirl slightly higher respoiisr tinics than thosc of
SH iircrss pat trrn. This is I,rcausr wlirii tlicrc is a
T h e results are drpicted in Figure 2 , arranged a a c h a i i g ~(if I i u t spot r w r y 500 qiieries, tlir rlient nrrds
two-dimensional array of graphs. The first row (Fig- tu riqurst n r w liot itrms from t h c sprvcr. Tlic lwhar-
ures 2a to 2 d ) illiistratrs t h r performanre of AQ while ior is olisimml in lioth Poisson and Bursty cl~icryar-
the srrond row (Figures 2~ t o 211) illustrates that of riwl jiat,tcrfis. The only cxccption is the, high rrsponsr
K Q . Thc. first two columns show tho perforrnanw of timrs from 1IC for N Q arriving ill Bursty (Figure 2 h ) .
Poisson arrival pattern while tho last twu columns \Vr \vi11 explain this ano~nalyi n Exlicrinirnt # 3 w11e1i
show that of Bursty arrival pattcwi. For the sake, of w r iiiwstigat(, HC in detail.
clarity, only caclie~liit ratios and response times are
depicted herc. \Vr will leave tlir error rates measure- 5.2 Experiment # 2
ment to Experiment #S. \Vr \vould like t o idrntify t h r best ~iossihlr,pcr-
In Figure 2, it is clear that the base case performs forrnancr of rnrious replacement policirs fur storiigtx
a lot worsr than any storagc caching scheme. It has carhing. This occurs in a read-only environnient, i.e..
a much lower hit ratio and a niucli higher response 24 = 0, with only one client. \Ve experinlent with six
time than those of any storagr caching scheme. This replarrrnrnt pdiries: LRU, LRL-k with I; = 3 (LRU-
is becausr a storagr caching scheme trades network 3). LRD. Mean. window with 1%'= 10 (Win-10). and
transmission for local disk access which has a rnucli E\V\IA witli (1 = 0.5 (EWMA-0.5). For LRD. tlic ref-
higher bandwidtli than that of a wireless channel. crriirc' ruunt of riicli databasr item is divided b y 2
\Ve ohsrrvr that OC. in genrral, yields liiglirr hit ra- cvrr? 1000 stconds. Choosing 0.5 f o r (I ill EWMA has
tios than AC. This is mainly Iic~c;iuseO C \vi11 caclir all tllr rlosrst iiitiiitirc mi:anisig as dividing tlir rdrreliw
attrihutes of tlie liot uhjtscts. tliiis rlirninating extra re- ciiiiiit I)? 2 . 01iIy rrsults of HC arc cliyirtid siiirr I T
quests wlirii (IiHcrrnt attriliutes uf the s ~ i n olijrrt e arr yidclh uptima1 prrforr1i;mre iri Ex~icririir~it #1
r ~ ! q u r s t t d i l l the futurc. It is isiterrsting that clrspitt, Tliv results art, illustrated i i i Figiirv 3. Tiit, first
the higlirr r;iclic hit ratios from O C alien r o m ~ ~ a r ~ d roiv drpirt s tlir pidormasire of AQ \vIiik tlir w i . u i i i 1
with tliosr from ..IC. OC rcsults i n liiglirr rrspusista ro\v ilqiirts that of N Q Ttw first tr\o ruliiuius sIiu\v
times as well. This anomaly is m i i i i l y duc. tu "blind t i l c l ~ ~ ~ r r ~ ~ ljf
r rSII
i l acrc3ss
~ i i i ~l ~~ ~~ ~ t t all(^
l ~ r till,
l i inst trv0
prcfetctiing" i n OC. Sinw OC prefrtclics it11 attributes c o l ~ i m msliuw t hiit of CSH iircrss j,;itt,rrii. Oiily rarlw

59

Authorized licensed use limited to: Universidad Nacional Autonoma De Mexico (UNAM). Downloaded on February 24,2024 at 04:59:43 UTC from IEEE Xplore. Restrictions apply.
 -U
-

--
Figure 3: Performance of various replacement policies

hit ratios and response times are measured There will response times are depicted. We will leave the error
be no error there is no write operation. rates measurement to Experiment #5. The results of
For .\Q, Figures 3a and 36 indicate respectively this experirnerit are depicted in Figure 4, arranged in
the cache hit ratios and response timi's for different the same manner as Figure 3.
replacement policies under SH accws Iiattcrn. Here, In t h e presence of write operations, the cache hit
LRU. LRU-3, LRD. and Win-10 perf(irni siniil..irIv.,. re- ratios drop considerably. u p to 10% decrease which
sulting i n i~cache hit ratio in the range of 60%. Both accounts to the increase in responsr times. This is
Mean aiid EWMA-0.5 seem to tie capable of capturing because a query accessing a n expired cached item will
a larger portion of hot objects. yieldiug hit ratios iu have to request the itcm from t,he server again. We fnr-
the range of 70%, as well ;IS a lowvr response time. ther note that the response times from Burst
Both Poisson and Bursty arrival pattr-rns mhit,it sim- arrival pattern are higher than those from Poi
ilar txhav-ior. rival pat,tcrn This is I)ecause for Burstv arrival pat-
For S Q ,Figures 3r a i d 3f clcpict tlic average c x h e tern. queries arrive i l l i i burst a t the server. The re-
hit ratios ;mi1 response times for diffrri'nt replacrnrent sults will h e qneuril up a t the downstream channrl
poiicirs iini1t.r SH iiccess pattern. Tlrc intrrpretation during tiurst?; pcriocl since the low haudaidth wirelms
;mil i!xplanation of the figures are siinilar. \VI= further channrl is not siitticient for delivi!ring the results fast
observe that the response times are double those of enough. This effect is cspecially serious for N Q since
AQ. This is tiecause in K Q , for each object acccssed the selcctivity is double t h a t of .iQ .
One approach
i n a query. (2, = I additional ot).jt,ct would iilso he to address this prolileni is b y a tilnrout heuristic. If
refermced. This doubles the selectirity of each query. the results of a query has been queued a t the top of
For CSH access pattern. LRU, LRU-3, and Win-10 a qucue for more than a duration threshold, the de-
perform similarly. They are still ahle to maintain a livery of prcfctrhcd itenis for current query will be
cache hit ratio in the range of 50%. Mean performs terminated. IVe will report more on t h e effect of this
a lot worse. Since Mean takes every access into ac- heuristic in the future.
count. it cannot adapt t o the rontinuous changes in From Experiments #2 and #3: we observe that
access pattern. LRD performs better clue t o it.s adap- mean is not stable. Its performance is very sensi-
tive natnre, ahout 5% higher hit ratios than those of tive to changes i n access patterns. LRU, LRU-3, LRD,
LRU, LRU-3, or Win-10. EWMA-0.5 adapts even bet- and EWMA-0.5 perform better than Win-10. LRU per-
ter, achieving another 5% higher hit ratios than LRD. forms slightly better than LRU-3 while LRD performs
slightly t)ett.er than LRU. EWMA-0.5, in turn, per-
5.3 Experiment #3 forms slightly better than LRD.
LVe now compare the performance of various re-
placement policies for storage caching under a more 5.4 Experiment #4
realistic environment with write operations and mul- We next compare the performance of LRU. LRU-
tiple clients. We repeat Experiment # S with identical 3, LRD, and EWMA schemes since our previous two
settings except that U = 0.1 and there are 10 mobile experiments have shown t h a t these schemes perform
clients. Again. only the performance of HC is reported similarly, within 10% difference, and relatively invari-
here. For the sake of clarity, only cache hit ratios and ant t o access patterns.

60

Authorized licensed use limited to: Universidad Nacional Autonoma De Mexico (UNAM). Downloaded on February 24,2024 at 04:59:43 UTC from IEEE Xplore. Restrictions apply.
 I

\Vr exprrinirnted thew four schemes on CSH ar- n t s shown that LRU prr-
Si> far. our r ~ x p ~ ~ i m ehave
cess patterii with changing ratrs of 300. 500. and 700 foriris slightly ltc,ttc,r than LRU-3. 111 [ l l ] .it was shown
querirs. Dur to spare limitation. only prrfornianrr of that LRU-3 will Iwrfornl Iicttrr than LRU if t h r access
AQ arriving in a Poisson pattrrn arr drpicted. Thr pattcrii of dat;il,asr qurtirs exhibits a ryrlic behavior,
riumbrr of rlicmts is tired at 10 and updatc probability i.c., the S~IIIC w t uf database itrins arr trfi~crrcrdby
is fixrd at 0.1. Again. we look a t HC only. T h e results datalxisr qiirriiih ;iftrr a rrrtairi duration. 111 ordrr to
are depicted in Figure 5 . cxprrinirnt this cyclic accrss pattern in our environ-
ment, wr folluw [11] in generating the set of database
items to bt. arrrssrd by each database query. T h e
prrfurrnanrr o f LRU, LRU-3. LRD. and EWMA-0.5 is
showii i n Figurr G . Again, only performance of AQ
arriving in I'iiisson pattern is depicted.

Figure 5 : Comparison among LRU, LRU-3, LRD. and

EWMA-0.5 schemes on CSH access pattctn
,% ,%, L/u ma, ,I m,, La -#I

Figurr 5 , illustrates cache hit ratios while Fig- P- -b)-

ure 5 b illustrates the corresporiding rcsponsr times.

LRU and LRU-3 perform slightly better than the other Figiirr G : Corriparison among LRU, LRU-3, LRD, and
two schemes when the changing rate is fast, Le., 300 EWMA-0.5 srlirmrs on cyclic access pattern
queries. As the changing rate slows down, EWMA-0.5
and LRD start t o outperforni t h e other two schemes. LRU suffrrs f r u m this rvrlir access pattern, inca-
EWMA-0.5 performs the best arnong the four schemes pablr of rrtaiuirig t h r uscful database itrms. It is
when chiingiiig ratr is slower thaii 500 querirs; is much o r i t ~ i c ~ f o r m cI,!~ l LRU-3 with 21%. EWMA-0.5 also
as i%highrr cachr hit ratios is rcror(ied prrfrrms ii lot I i r t f v r thiul LRD. by as m i l c h as 11%,.
III CSH iirrcss piittrrii. a prcvi~iuslyrold i1;itaiiasr .-\lttiougli n o t slr[.itirally drsign for tlir ryclic pattrrn.
item m a y brromr liot . Sinw LRL:.iiiisril S C I I ~ I I I ~ Swill EWMA-0.5 is iildo t o acliicvr rc;tsonably n ~ lpcrfur- l
always c a r l i ~tlir rerrritly a('rr 11 i t r n i . t l i r y 1if~for111 i i i i i ~ ~ ~ ' < ' ,rlow t o t l i i i t of LRU-3. which is rlrsigncd for
bettvr for fast ch;iiiging ratv siiiw tlw w r r e n t l y rerlird tliis ryrliv 1)iittrrii.
item will h v c ;I iiigli prdxiIiihty of lwiiig lint. IVIirn
the vhiinaiiip rat(, sknvs h v u . LRU-lws~!d srlwmrs 5.5 Experiment #5
could iint t ; i k adv;iiit,iq!,v uf tliv riirrrutly ~ ~ ~ I < ~ I iI t( (YwI O u r f i f t h s('t [ i f rqirsim(~ntsiiiwst.ig;itcs thv rffcart
and EWMA illid LRD i ~ ~ f o r r Iwttcr. ii of i i l x l i i t , . 1,roIiiihility on t l w p w f m " i c v . T h error

61

Authorized licensed use limited to: Universidad Nacional Autonoma De Mexico (UNAM). Downloaded on February 24,2024 at 04:59:43 UTC from IEEE Xplore. Restrictions apply.
 I*". I
,om,

Figure 7: Error rates uersiis update probabilities

rates, hit ratios. and response tirnes for .IC. OC. and eration folloa.ri1 I)?a read operation within any two
HC are measured. Due t,o spare limitation. ~ v eonly consecntive refreshes. the error rate is low. \Vhcn U
present the results for .IQ w i t h Poisson arrival pat,- increases. the prohaljility that thrre is a read opera-
tern, operating on SH access partern. with EWMA-0.5 tion fullowing ii write operation within a refresh dura-
replacement policy. The 0 valuc ranges from -1. 0, t o tion incrrasps m d the error rates i n c r ~ a s eaccordingly.
1 while update probahility. U . ranges from 0.1. 0.3.to One might also observe that thc 1:rror rates iiicreae
0.5. \Ve do not present the resnlts for highrr values of iis ,) incrrascs. This is beranse when .j increases. the
U since we hrlieae that the ntiin1ii~rof write operations refresli tinic of a database item iricrcases accordingly.
shotrld he less than the nurnber uf read operations, es- This increasi's t h e prubahility uf re;rding a stale i t e m
pecially ill a mobile environnirnt. Figuse 7 depicts thc hefore the i t m i is rcfreslicil.
results. T h e first row illustrates the nieasurements for The liit ratios increase as 3 incrwses. since tlic re-
.1 = -1. the second row for , I = 0. and the third fresh time uf ii database itern increases with 3 . thus
row for ,d = 1. The first coltuiin presents error rates. increasing the probahility that a locally cached item
the second column liit ratios. and the third column could Ije accessed. In contrast, the response times de-
response times. crease as J increases due t o iucreased hit ratios.
\Ve ohserve that OC. in general. has higher error
rates than those of AC and HC. This could be ex- 5.6 Experiment # 6
plained as follows. Assume that a client reads at- Our final sPt of experiments is t o study the error
tribute a of a cached object, x, at time t l and t? while rates during disconnection. T h e duration of a discon-
attribute b of x is updated by another client a t time nection period. D.for each client ranges from I to 10
t. t l < t < t p ,a t the server. For OC. since the update hours. \Ye vary the number of clients that are discon-
of attribute b is a n update on L,the object being read nect,ed, .I ..1, 3, 5 , 7, t o 9. The total number Of
from
a t t r , the read operation a t t? is an error. For AC clients is still fixed a t 10. Figure 8 presents the error
and HC, the read operation a t t z is not a n error since rates of AQ with Poisson arrival pattern, operati%
attributes a and b are considered as different items. on SH access pattern, employing EWMA-0.5 replac('-
.IC and HC result in similar number of errors. How- inent policy. The error rate is measured based on a
ever, HC can satisfy more read operations using locally "perfect" knoidedge of all the events in the sirnulatell
cached items due t o the higher hit ratios. Thus, the system. Figures 8a t o 8c depict the error rates for .kc,
error rates of HC are slightly lorver than those of AC. OC, and HC respectively.
The error rates increase with the update probabil- As shoivn i n Figure 8, the error rates increasr' as
ity U . When U is low, there are few write operations. the duration of disconnection increases in all schemes.
Since a n error only occurs \ h e n there is a write op- This is because during disconnecrion. a client will m -

62

Authorized licensed use limited to: Universidad Nacional Autonoma De Mexico (UNAM). Downloaded on February 24,2024 at 04:59:43 UTC from IEEE Xplore. Restrictions apply.
 References
(I] S. Acharya, R. Alonso, hl. Franklin, and S. Zdonik.
1'1
,, ,

-
Broadcast Disks: Data Manaaement for Awmmetric
Communication Environments. In Proceedings of the
;lb
!/I
.4CM SIGMOD, pages 199-210, 1995.
,//
[2] D. Barbara and T. Imielinski. Sleepers and Worka-
holics: Caching Strategies in Mobile Environments. In
Proceedings of the ACM SIGMOD, pages 1-12, 1994.
[3] \I. Choy. \ I . Kwan, and H.V. Leong. On Real-time
Distributed Geographical Database Systems. In 27'A
Howou International Conference on System Sciences,
pages 33;-346, 1994.
[4] D. DeWitt and D. Maier. A Study of Three Alter-
native 1Vorkstation-Server Architectures for Object-
Oriented Database Systems. In Proceedings of V L D B ,
pages
. .. IOi-121, 1990.
[5] JY. Effelsberg and T. Haerder. Principles of Database
Buffer Management. A C M Rnnsactions on Database
Figure 8: Error rates during disconnection Systems, pages 560-595. December 1984.
i
[6] 11. Franklin, M. Carey, and M. Livny. Global Memory I
!danagement in Client-Server DBMS Architectures.
tinue to use its locally cached, hut expired, items. As In Proceedrngs of V L D B , pages 596-609, 1992.
more clients are disconnected, the total number of er- [7] C . G. Gra? and D. R. Cheriton Leases: A n Efficient
rors will increase. The increase is relatively slow, how-
ever, as illustrated i n Figure ad, which shows the re-
lationship between error rates and number of discon-
Fault-Tulmant Mechanism for Distributed File Cache
Consistency. In Proceedings of SOSP, pages 2 0 2 ~ ~ 2 1 0 , i
I
nected clients when the disconnected duration is fixed 1989. i
at 5 hours. [8] Y.Huang. P. Sistla, and 0 .Wolfson. Data Replication ~

i
for hlobile Computers. In Proceedings of the A C M 'I
ti
- . .
- C;onclusion SIGMOD .mees - 13-24. 1994
[9] T. Imielmski and B. Badrinath. Mobile \!'ireless Com-
We have presented a framework for caching mech- puting: Challcnges in Data Management. Communt-
anism as one way t o improve d a t a access performance cations o/ the ACM, 37(10):18-28, 1994.
in a mobile environment. The caching mechanism is
illustrated and implemented on object-oriented data- [IO] J. Jannink. D. Lam, N. Shivakumar. J. Widom. and
base model. We have shown that page-based caching D.C. Cox Data Management for Uscr Profiles in
is not suitable in this mobile context and proposed JVireless Communications Systems. Technical report,
three different caching granularities, namely, attribute Computer Science & Electrical Enginerring Depart-
cachine. obiect cachinrr. and hvbrid caching. We have uent. Stanford University. 1994.
also sh&n "that conveKtiona1 ;ache coherence and re- [Ill J . J . tiistlrr and M . Satyanarayanan Disconnected
placement schemes are not a effective and modified operation in the Coda file systeni. 111 Proccedings o/
strategies which adapt to object access patterns have SOSP, pages 213-225, 1991.
heen proposed. The behavior of Our caching mecha- 1121 H.\'. Leong and A. Si. Database Caching over the
nism is illustrated through a series of simulated exper- Air-Storapr The Computer Joumol. To appear.
iments. 1131 C. hlin, 51 Chen. and N . Roussopoulos. The Imple-
We intend to extend this study in several direc- mentation and Pcrformance Evaluation of t h e ADMS
tions. First, this study assumes that each mobile client Query Optimizer: Integrating Query Result Caching
only communicates with one server. In real applica- a i d Matching. In Proceedings of Intemational Con-
tions, a mobile client might request items from multi- ference on Eztendtng Database Technology, pages
ple servers, possibly under different cells. This further 323-336. 1991.
complicates the problem because the contact server for
a client might have to request and cache items 1141 E. O'Neil. P. O'Neil, and G. Weikum. The LRU-
from other remote servers on behalf of t h e client which ti Page Replacement Algorithm for Databas Disk
initiates the query. Second, we would like to inves- Buffering In Proceedings ofthe ACM S I G M O D , pages
tigate into the corrertness conditions on concurrent 297-306, 1993.
query processing. Since mobile clieiits arc highly dy- [ l j ] .I.Si and H 1.. Lrong. Query Processing and 01'-
iiarnic and autonomous, the notion of serializability tiimzation for Broadcast Database. In Proccedings
for coiioentiorial transaction proccssing would he too of Inirmntional Conference on Database and Ezpcrt
restrictive for cfficicrit roilcurrency control in this con- Systems Applications, pages 899-914. 1996.
text. Rclaxcd notions should hr explored. Finally. we r161 .\, ill,^^^^^,^^^, H,F, t i o r t h , s, ~ ~ ~ d~ ~ ~t , ,~. ~ h ~ , ~
are incorporating our results irit,o our prototypr to bc h o s e System Concepts. llcGraw-Hill, 1996.
furthcr validated with the simulated results.

63

Authorized licensed use limited to: Universidad Nacional Autonoma De Mexico (UNAM). Downloaded on February 24,2024 at 04:59:43 UTC from IEEE Xplore. Restrictions apply.