A DATA MINING ALGORITH FOR MULTI LEVEL PREFETCHING IN STORAGE SYSTEMS

Transcription

1 Special Iue o ICIT 20 Conerence A DATA MINING ALGORITH FOR MULTI LEVEL PREFETCHING IN STORAGE SYSTEMS Mai Nijim, Young Lee, Nuri Yilmazer, Remzi Seker * Texa A&M Univeriy Kingville, USA Deparmen o Elecrical Engineering an Compuer Science {mai.nijim, nuri.yilmazer, young.lee}@amuk.eu Univeriy o Arkana a Lile Rock, USA rxeker@uarl.eu ABSTRACT Parallel orage yem have been highly calable an wiely ue in uppor o aainenive applicaion. In uure yem wih he naure o maive aa proceing an oring, hybri orage yem op or a oluion o ulill a variey o eman uch a large orage capaciy, high I/O perormance an low co. Hybri orage yem (HSS) conain boh high-en orage componen (e.g. oli-ae ik an har ik rive) o guaranee perormance, an low-en orage componen (e.g. ape) o reuce co. In HSS, ranerring aa back an orh among oli-ae ik (SSD), har ik rive (HDD), an ape play a criical role in achieving high I/O perormance. Preeching i a promiing oluion o reuce he laency o aa ranerring in HSS. However, preeching in he conex o HSS i echnically challenging ue o an inereing ilemma: aggreive preeching i require o eicienly reuce I/O laency, wherea overaggreive preeching may wae I/O banwih by ranerring uele aa rom HDD o SSD or rom ape o HDD. To are hi problem, we propoe a aa-minin muli-layer preeching algorihm ha can juiciouly preech aa rom ape o HDD an rom HDD o SSD. To evaluae our algorihm, we evelop an analyical moel an he experimenal reul reveal ha our preeching algorihm improve he perormance in hybri orage yem. Keywor: hybri orage yem, oli ae ik INTRODUCTION A recen uy how ha new aa i growing annually a he rae o 30% an ive exabye (5 260) o new inormaion were generae in 2002 [6]. Major corporaion, upercompuing cener an rich meia organizaion, incluing Lawrence Livermore Naional Laboraorie, Oak Rige Naional Laboraory, NASA Ame, Google, Boeing, an CNN rely on large-cale orage yem o mee emaning requiremen o large aa capaciy wih high perormance an reliabiliy [7]. Large-cale orage yem have o be evelope o ulill rapily increaing eman on boh large orage capaciy an high I/O perormance [8][9]. Traiionally, he orage capaciy an I/O perormance o a yem are cale up by imply employing more HDD [0][]. However, I/O perormance an capaciy improvemen hrough he increae number o HDD are an expenive oluion ue o huge expene on new orage equipmen an he increae mainenance ee. Hybri orage yem (HSS) conaining SSD, HDD, an ape ubyem can provie an ieal aa orage oluion o igniicanly improve orage capaciy an I/O perormance a low co. Auhor who are no luen in Englih are requee o ak a naive Englih peaking colleague o correc heir manucrip beore ubmiion. Large-cale hybri orage yem will become increaingly popular in he nex ew year or he ollowing wo reaon. Fir, HSS will keep i high perormance by preeching an caching highly accee orage objec in high-pee orage componen uch a SSD or HDD. Secon, hybri orage yem are co-eecive, becaue inexpenive ape help in increaing orage capaciie a very low co. Thereore, i i believe ha hybri orage yem, which have high perormance, long archive lie, an low co, are ieal aa orage plaorm or a wie variey o aa-inenive applicaion rom human genome analyi [] o remoe-ening applicaion [3]; rom long running cieniic imulaion [4] o biological equence analyi [2]. Hybri orage yem are pracically eaible becaue SSD, a one o he newly evelope orage componen, can be eaily connece o any oher ype o orage evice [5]. Highly eicien aa raner beween ape, HDD an SSD i imporan o hybri orage yem. For a wie range o aa-inenive applicaion, criical aa are require o perioically UbiCC Journal Page 0

2 Special Iue o ICIT 20 Conerence or coninuouly backe up o ape o ha aa reoraion i poible in cae o yem crah or aa lo. I i imperaive o minimize aa reore ime in orer o improve aa I/O perormance, which largely epen on high aa raner rae beween ik an ape. Thu, ranerring aa back an orh among SSD, HDD, an ape play a criical role in achieving high I/O perormance. Preeching i a promiing oluion o reuce he laency o aa ranerring among SSD, HDD, an ape. Preeching i a proce ha aim a reucing he number o reque iue o HDD or ape by caching popular aa in SSD. Preeching can be ue o preven I/O banwih uneruilizaion by ully exploiing ile ime o orage componen o hie I/O laency. Preeching in he conex o hybri orage yem i echnically challenging ue o an inereing ilemma: aggreive preeching i neee o eicienly reuce I/O laency, wherea overaggreive preeching may wae I/O banwih by ranerring uele aa rom HDD o SSD or rom ape o HDD. To overcome hee echnical obacle, we inveigae a mulilayer preeching algorihm (PreHySy) o enhance I/O perormance o hybri orage yem. The re o he paper i organize a ollow. Secion 2 ummarize he relae work. Secion 3 preen he archiecure o he hybri orage yem. In ecion 4, we preen he muli-layer preeching moule. Secion 5 ecribe he analyical moel o he muli-layer preeching. The concluion o he paper i icue in ecion 6. 2 RELATED WORK Inegraing Har Dik wih Tape: Tapebae orage yem are eenial archival orage componen require by high-perormance compuing communiie. In he pa wo wo ecae, magneic ape have been coniere he mo co eecive an reliable way o archive an backup aa or a long perio o ime. Alhough ape orage yem oer high capaciy a low co, ape have a perormance rawback high I/O laencie an lower banwih ue o equenial aa accee. Convenionally, he perormance o ape-bae orage yem can be improve by inegraing har ik wih ape [26]. Har ik (e.g, RAID reunan array o inepenen ik) can be employe in ape orage yem o ore requenly ue backup aa. To reuce aa raner ime beween ik an ape, aa riping iea are applie o increae I/O hroughpu an reuce repone ime [2]. increae I/O hroughpu an reuce repone ime [2]. Soli-Sae Dik: Soli Sae Dik are mae o emiconucor memory evice. Currenly, here are wo ype o SSD: lah memory bae SSD an DRAM bae SSD. Boh lah memory bae an DRAM bae SSD can be inegrae wih HDD an ape o increae I/O perormance o large-cale hybri orage yem. Recenly, lah memory bae SSD are becoming very popular or aainenive compuing becaue o he avanage inherie rom lah memory uch a high eniy an low power properie. There ha been ome work on applying cache o booing perormance o parallel ik yem. Koz an Elli inveigae cache managemen echnique ue in parallel ile yem o cloe he gap beween proceor an ik pee [5]. Karela e al. examine he ue o caching o reuce yem repone ime an o improve aa hroughpu [6]. I i believe ha i he cache are eparaely reae, i i eay o conrol cache mi equence. Each pariion or a ik will be manage eparaely uing he convenional LRU (Lea Recenly Ue) replacemen algorihm. We evelope a cache pariioning mechanim or cluer orage yem or achieving high ecuriy or aainenive applicaion running on cluer. More o he poin, a number o cache pariioning meho have been propoe wih ieren opimizaion objecive inclue perormance [7], airne [8], an qualiy o ervice [9]. On he oher han, SSD can provie more avanage han he one provie by he builin cache becaue he buil-in orage cache can only be ue wih ha peciic orage yem while SSD can be connece o any orage evice [20]. Harware-Bae Preeching. Poenial acor egraing I/O perormance o orage yem inclue heavy loa, low banwih, banwih uneruilizaion, long eek ime, an ik roaional elay. A promiing approach o booing I/O perormance i o increae I/O banwih by preeching aa e a variou place, incluing orage erver, clien, an proxie. Preeching echnique oun in he lieraure ake eiher oware-bae or harware-bae approache [2]. Soware-bae preeching cheme epen on oware o eec regular aa acce paern, wherea harware-bae approache rely on harware o reuce aa acce penaly [22][23]. Harware-bae preeching approache can be claiie ino wo group: Spaial cheme where aa acce o he curren block i he bai or making preeching eciion, an emporal cheme where block o be preeche i bae on value o peculae aa raher han aa localiy. Soware-Bae Preeching. Exiing oware-bae preeching oluion can be generally caegorize ino wo camp: inorme preeching [24] an preicive preeching [30]. Inorme preeching algorihm rely on uericloe inormaion abou uure reque o bring aa ino buer. Preicive preeching echnique uilize aa acce hiory o preic uure aa reque. 2 UbiCC Journal Page

3 Special Iue o ICIT 20 Conerence LAN SAN Daa mi Daa mi Upper level preeching SAN o FTP Figure : The Hybri Sorage Syem Archiecure Lower level preeching ape o SAN There exi wo ype o preicive preeching approache: epenency-graph-bae preicive preeching (DGP) an parial maching-bae preicive preeching (PMP). DGP algorihm ue epenency graph o moel acce paern, wherea PMP algorihm mainain Markov preicor o calculae probabiliie o uure reque. Thee exiing preeching oluion have everal eec: Fir, inorme preeching i inapplicable in cenario ha aa-inenive applicaion are unable o provie a priori inormaion abou which aa block are likely o be accee. Secon, DGP algorihm preic orhcoming reque uing previouly accee aa wih only coniering ir orer epenencie. Thir, PMP algorihm o no are he iue o how o chooe a conan value or maximum orer. La, none o hee exiing preeching algorihm conier bringing popular aa e rom har ik ino oli-ae ik. 3 THE HYBRID ARCHITECTURE Figure reveal he hybri orage yem archiecure wih preeching. I coni o FTP erver wih SSD, orage area nework (SAN) wih HDD, an a ape ubyem. The oli-ae ik are eigne o keep he mo highly accee ile. File wih le prioriy or le acce requency are ore in HDD o he SAN ubyem an all oher ile are ore in he ape. The SSD an HDD are coniere a high-en orage componen while he ape are claiie a low-en orage componen. A preeching cheme i eigne o bring aa ino HDD o SAN an opimally, ino he SSD beore i i requee. Our muli-layer preeching algorihm coni o wo par, he upper level preeching an he lower level preeching. The upper level preeching raner he aa rom he HDDS o he SSD an he lower level preeching ranpor he aa rom he ape o he HDD. A mi comman will be iue when requee ile canno be oun in SSD an he miing ile will be eche o he SSD rom he HDD. I he requee ile are no locae in he HDD, he algorihm will iue nex-level mi comman an place he miing ile rom he ape o he HDD an inally o he SSD. 4 PREFETCHING ALGORITHMS 4. Parallel Daa Traner rom Tape o Dik In hi ecion, we propoe an approach o quickly move aa rom parallel ape orage yem o har ik o achieve high I/O perormance.parallel ape librarie can inuiively increae he aggregae banwih beween he ik orage an he ape orage an reuce he ape wich ime by inroucing parallel loa/unloa 3 UbiCC Journal Page 2

4 Special Iue o ICIT 20 Conerence operaion. Thi preeching mechanim nee o cheule rea reque in a way ha he highe prioriy aa block will be eche ir. To uppor he aa raner parallelim, a riping an aa placemen echnique or he hybri orage yem are ue. Sriping i a well-known echnique or improving he eecive I/O banwih or orage yem. We conier aa riping o acce apereien aa e in parallel. Daa e will be ivie ino uniorm chunk o be preeche an ore o ik imulaneouly. The aa riping cheme i compleely ranparen o he PreHySy preeching mechanim. To eermine an opimal riping wih, we conier boh aa ize an I/O workloa. The riping echnique can uppor aa raner parallelim, hereby horening I/O repone ime. However, riping caue a large number o mall I/O call, which in urn increae wich ime o ape rive [27]. Coniering he big penaly aociae wih ape wiche an he long raner ime aociae wih he huge objec ize, we propoe a aa placemen cheme o leverage objec acce probabiliie an he relaionhip beween objec. Thi cheme can improve ape wich parallelim an ynchronize aa eek wih high probabiliy, an hu increae he parallelim o he objec raner. The main goal o our aa placemen algorihm i o reuce he ape wiche wihin he ape library, increae ape wich parallelim among ape librarie, an increae he aa raner parallelim among ape rive. To achieve hee goal, we propoe a aa cluering mechanim ha cluer objec wih high probabiliy o be requee ogeher. Our aa cluering iea i bae on an aumpion ha relae aa reque are highly o be requee ogeher. Daa block wih a high probabiliy o be requee ogeher will be groupe in one cluer. The ape ubyem will be compoe o n ape librarie. Each ape library ha ienical ape rive, wich rive, an ape. The baic iea i o ivie each ape library ino wo group. The ir group conain ape ha are moune all he ime an conain high emergency aa. The number o ape ha are moune all he ime i -. The econ an laer cluer conain ape or each ape library, which i moune uring arup ime an will wappe ou i he requee ripe ca n no be oun wihin he moune ape. Fig.2 i an illurae example o objec cluering mechanim in he ape orage. Aume here are hree ape librarie where each ape library conain even ape, ive ape rive, an wo wich rive. The ir bach or each library conain hree ape, an he econ an laer bache conain wo ape each. We evelope a preeching algorihm calle PRE_TD or ape reien aa. When he requee aa e are no locae in he SAN ubyem, PRE- 4 TD will ech he aa e rom he ape o he HDD in SAN. PRE-TD oe no preech aa block i a higher prioriy aa block mu be evice rom he ik. We aume ha he uer acce paern i known in avance. Each I/O reque in he acce paern i aigne a prioriy level epening on aa acce hiory. I i worh noing ha he prioriie o riping uni are equal o he prioriy o he correponing I/O reque. Moreover, he prioriy o a aa block in a ik i erive rom he prioriy o he nex reerence o he aa block. I here i no reerence o he aa block in he acce paern, he block prioriy will be i-max, where i i he inex o he mo recen reerence in he aa block, an Max i a large poiive ineger ha i greaer han he lengh o he acce paern. I aa block reiing in he ik ubyem are no appeare in acce reerence li, hen he block will be aigne he lowe prioriy. Daa block will be migrae o ape bae on he lea recen ue policy. When PRE-TD i invoke, he riping uni o an I/O reque will be mappe o he ik ubyem uing he roun-robin mapping raegy. PRE-TD will conruc a li o preeche reque in accorance wih heir prioriie. PRE-TD will have all he I/O reque in he li procee by ik rive aer he reque are aigne o he ik. The PRE-TD algorihm i ecribe in ig.3a. Figure 2: The aa rucure o he PRE-TD Algorihm. Fig. 2(a) i an acce paern known in avance. Fig. 2(b) how he riping uni o I/O reque iribue among ape librarie. Fig. 2(c) how a preeching cheule an upae prioriie UbiCC Journal Page 3

5 Special Iue o ICIT 20 Conerence 4.2 Parallel Daa Traner rom har ik rive o oli ae ik The perormance o large-cale orage yem will be ubanially improve by employing a number o oli-ae ik ha can help in reucing he number o har ik accee. We evelop a parallel aa raner algorihm calle PRE-DS ha preeche aa e rom he ik yem o he oli ae ik. The ir componen in PRE-DS i a SSD pariioning algorihm (or PaSSD or hor), which ynamically pariion an array o SSD among HDD in uch a way o maximize I/O perormance. The baic iea o PaSSD i moivae by he obervaion ha I/O workloa i no uniormly iribue among parallel har ik. The oli-ae ik will be allocae ynamically epening on he populariy, ize o conen, an acce paern. Taking hee acor ino conieraion, wo approache are implemene in PaSSD o aign weigh or conen an aociae a oli-ae rive o each conen. The ir approach i calle Conen- Populariy-Bae Weigh Aignmen, in which he weigh o a new conen i aigne bae on i populariy o oher relevan conen ha alreay have meaure populariie. I a large number o uer have requeing conen regularly, he conen houl be eche an cache on a oli-ae rive bae on he meaure weigh. Weigh will be ajue ynamically an regularly o relec any change on he ly. The econ approach i calle Collaboraive-Populariy-Bae Weigh Aignmen. In hi approach, we peciy populariy weigh by coniering correlaion among requee conen. For example, i uer u requee conen c an c2, an uer u2 requee conen c2, hen c2 will receive a higher weigh becaue here i a rong likelihoo ha boh u an u2 will reque c2. PaSSD will keep aeing conen an eimaing he weigh on a regular bai in a ynamic environmen. Aer populariy weigh are aigne uing he above wo approache, PRE-DS will ecie o orwar reque o eiher SSD or HDD. Furhermore, PRE-DS will be ue o raner aa rom HDD o SSD i requee popular aa block are no available in he oli-ae rive, an i ranerring ha aa block oe no caue a aa block wih higher prioriy o be evice rom he oli-ae ik. To improve I/O perormance, PRE- DS rive o ech a many popular aa block a poible rom all har ik an ore hee popular aa in he high-pee oli-ae ik. Fig.3b ecribe he PRE-DS algorihm. 5 MATHEMATICAL MODEL FOR HYBRID STORAGE SYSTEMS In hi ecion, we ecribe he analyical moel o acce he propoe PreHySy algorihm. The mahemaical moel coni o hree par: he erver acce moel, he acce ime wihou preeching, an he acce ime wih preeching. The ulimae goal o hi moel i o provie crieria ha can mahemaically evaluae he perormance o our preeching algorihmn. More peciically, we wan o maximize he average acce ime improvemen S, where S i eine a ollow: S () Where an are he average acce ime when preeching i no carrie ou an when preeching i carrie ou repecively. In he erver acce moel, we conier muliple uer acceing he nework hrough he p erver. In our analyi, we conier M/G/ roun robin queuing yem. In hi yem, he average ime o inih a job, neceiae a ervice ime x, i calculae a ollow r x /( ) (2) Where i he yem uilizaion. Le he average ize o an objec be, le be he ize o an objec locae in he ik yem, an le be he ize o an objec locae in he ape orage where. Thi mean ha par o he objec coul be locae in he ik yem, an he oher par i locae in he ape orage. In hi cae, he upper level an he lower level preeching algorihm nee o be calle. I 0 hen he aa objec i locae in he ape orage an he lower level preeching algorihm will be calle. Oherwie, i 0 hen he aa objec i locae in he ik yem, an he upper level preeching algorihm will be calle. Le b be he banwih in he upper level hybri orage yem (i.e. he banwih rom he ik yem o he oli-ae-ik). Le b be he banwih in he lower level (i.e. he banwih rom he ape orage o he ik yem). We aume ha b b. The average ervice ime x i eine a x ( ) /( b b) (3) Reque a arrival rae o are iue by he uer. Bae on he aumpion ha he arrival rae i no aece by preeching which mean ha preeching i oally ranparen o he uer. Wihou preeching, a proporion h o he uer reque reul in a hi in he oli-ae ik. Thi mean 5 UbiCC Journal Page 4

6 Special Iue o ICIT 20 Conerence ha hi proporion i erve by he oli-ae ik. The ailure raio i which mean ha he h reque are locae in he ik yem an/or he ape. The proporion h reul in a hi in he ik yem which mean ha h proporion o he uer reque are erve by he ik yem. The complemenary value mean ha h h reque are erve by he ape yem. We conier he cae where preeching i no perorme. Uing (2) an (3) he average rerieval ime o he reque rom he ik yem i r (4) b ( ) The reque are iue by he uer a rae λ Where h (5) b Thu, r (6) b h The average rerieval ime o he reque rom he ape orage i r (7) b ) Where ( h b b ( bh b ) bb Thu, b r bb ( bh b (8) (9) We aume ha he acce ime i 0 when he reque are locae in he oli-ae ik. The acce ime or he reque ha are locae in he ik yem an in he ape orage are r an r repecively. Hence, he average acce ime i h 0 h r h r r h b b h bb ( bh b r ) (0) We calculae he acce ime or reque coniering preeching. For each uer reque, an average o n( iem will be preeche where n( n( n( () Where n( i he average number o iem o be preeche rom he ape o he ik yem, an n( i he average number o iem o be preeche rom he ik o he oli-ae ik. We aume ha all iem ha will be preeche have probabiliy p where p p p 2 Where p i he probabiliy o iem o be preeche rom he ape orage o he ik yem, an p 2 i he probabiliy o iem o be preeche rom he ik yem o he oli-ae ik. For impliciy, we aume ha he iem o be preeche rom he ape o he ik have he ame probabiliy, an he iem o be preeche rom he ik o he oli-a ik have he ame probabiliy. When preeching i perorme on he lower level (i.e. rom he ape orage o he ik yem), he hi raio in he ik yem will be increae by he number o he preeche iem. Thu, he hi raio in he ik yem i expece o rie o h, where h h n( p (2) 2 When he aa objec are o preeche rom he ik yem o he oli-ae ik, he hi raio in he oli ae-ik i expece o rie o h where (3) h h n( p By uing preeching, no only he value o he hi raio will be aece, bu alo he yem uilizaion, he rerieval ime, an he acce ime will be aece. One imporan acor o preeching i which iem in he ik yem or in he oli ae ik will be evice in cae hey are ull. To olve hi problem, i i neceary o moel he ineracion beween preeching an he replacemen policy. The rae o uer reque ha require eman eche or he hybri orage yem i ( h) ( h). Now he p erver mu erve boh on eman eche a well a preeching. Thu, he eecive rae o he reque i ( h) ( h) n( n( (4) Bae on (4), he yem uilizaion i preeching i one on he upper level yem (i.e. rom he ik yem o he oli-ae ik) i ( h n( (5) b 6 UbiCC Journal Page 5

7 Special Iue o ICIT 20 Conerence Uing (2) an (3), he average rerieval ime o he aa objec rom he ik yem o he oli-ae ik i r (6) b ) r ( Uing (5), we ge b ( h n ( F )) (7) The yem uilizaion when he aa objec are preeche rom he ape orage i calculae a ollow ( h n( ) ( h n( ) b b (8) The average rerieval ime o aa objec o be rerieve rom he ape orage i r (9) b ) r ( Uing (8) we ge b bb ( b ( h n( )) b ( h n ( ) (20) Thu, he average acce ime when preeching i perorme: h 0 ( h) r ( h h) r h n ( p b ( h ( p )) n( P n F p 2 ( ) b bb b( h n( ( p )) b ( h n ( ( p )) 2 ( 2) Subiuing (0) an (2) in () which eine he acce improvemen when uing preeching, we ge h S { b ( h bb b( h n ( p ( p )) n( P2 n ( pb } n( ( p )) b ( h n ( ( p )) h b { b h bb ( bh b 2 )} ) (22) 6 TESTBED To e he perormance o he muli-layer preeching orage, we will buil a ebe. The ebe plaorm i a peronal compuer running Linux equippe wih Inel Penium Dual Core 3.2 GHZ proceor, IGB o main memory, he har rive i Samung SATA ik, he oli ae rive i 32GB Trancen SSD. Depening on he acce paern known in avance, we meaure he perormance o he yem wih an wihou preeching. Our primary meric or meauring he yem perormance i he overall compleion ime, he average ime o rea a block, an he average ik acce ime. However, our ulimae goal i o e he ucce o our muli-layer preeching algorihm. Figure 3 plo he reul or ive value o he average number o preeche iem again he acce improvemen where he aa ize i 5MB; he value o he probabiliy i 0.4, he arrival rae i 30MB.ec. Figure 3 reveal ha when he average number o preeche iem increae, he average acce ime improvemen S (pleae ee equaion.) o he yem will increae. Figure 3: The Average Acce Improvemen 7 SIMULATION AND EXPERIMENTAL RESULTS To imulae our algorihm, we ue Microo Acce which provie a goo ounaion or he aabae echnique calle aa mining. I ue objec o help he uer li an organize inormaion, a well a prepare pecially eigne repor. Ju a a preahee ue a able wih row an column, Acce ha an array o aa conainer (view, iler, ec.) ha elevae one-imenional li ino he powerul realm o aa managemen. Thoe objec are he aabae able, orm, query, an repor. The able i he warehoue o he inormaion you ore in your aabae. The row an column o he 7 UbiCC Journal Page 6

8 Special Iue o ICIT 20 Conerence preahee become your recor an iel. A recor i one iniviual iem o inormaion ha coni o iel. The aabae uer mu be able o eaily perorm quick iler in a orm or a able. (A iler i a look a ju one peciie caegory o inormaion, e.g., all cuomer who purchae le han a peciie amoun). Clicking on he avance iler quickly eiher in he orm or able view allow he uer o rill ino he aa o in peciie inormaion in our projec. A iler i really a or o "query on he ly" or a quick look a limie par o he aabae. The Acce Form are imply a unique way o viewing he inormaion alreay in your able. Table wih hunre o recor an many iel can be wie an eep. Form help u iplay, view, an ue all or ome o he inormaion in he able. Form alo are a more convenien way o ener new aa, an acce exiing recor. Remember: whaever eleion, change or recor aiion you make on a orm i e ino or remove rom he able. Query i anoher mo imporan par o our work wihou he ruy query, our aabae woul be a unwiely a an overue preahee able. We woul no be able o ee a group o ree, becaue he ore o inormaion in boh urgen an nonurgen aa woul prea wie an eep. We nee he power o querie o bring orer an ocu o our invenory o har-earne aa. I i imporan o remember ha a query i nohing more (or le) han a rerice view o he big aa able. Any change mae o he big able are alo relece in he query able an objec (orm an repor) ha epen on he query. Querie, in oher wor, are rerice, or runcae, view o he overall aabae. A ave query in an MS Acce aabae i like he web bookmark. In a ave query, we have inruce Acce o reerve a peciic view o he inormaion eiher in a able or a orm eigne epecially or he query. Daabae manager requenly ue querie or pecial orm view o ha aa enry people can only ee, ue or ei par o he aabae. The power o he MS Acce aabae i no alway in i robu abiliy o hol on o aa. I i hrough i abiliy o rill own which make Daa Mining more eicien. In our imulaion, we ue wo aa e which are urgen an non urgen. We ue he Marke Bake Analyi Daa Mining Technique which i a moeling echnique bae upon he heory ha i you buy a cerain group o iem, you are more or le likely o buy anoher group o iem. For example, i you are earching o buy a mobile evice hen all he relae iem o mobile will be iplaye uch a creen guar, pouch, ec. Thereore, i erve a an eicien an eecive echnique o preech he aa. We make ue o he combo box ool provie by MS acce an in he back en we wrie he coe o rerieve he aa in viual baic. Figure 4: The Repone ime wih an wihou preeching Figure 5: The rerieval ime wih preeching Figure Figure 4 how 6: The he rerieval rae o aa ime raner wihou rae preeching wih 8 UbiCC Journal Page 7

9 Special Iue o ICIT 20 Conerence Figure 7: The acce improvemen ime when uing PreHyy an wihou preeching. We can noice here ha relaive preech o 5 conecuive aa raner achieve nearly a 2x improvemen over he normal aa raner wihou preeching wih imilar growh curve. Figure 5 an igure 6 epic he aa rerieval rae wih an wihou preeching. In igure 5, he yem ile ime i equal o i repone ime where a in igure 6 he ile ime ecreae an he repone ime ecreae which reul in improvemen in he ae relaive rae. Figure 7 plo reul or ive value o he average number o preeche iem again he acce improvemen where he aa ize i 0 MB; he value o he probabiliy i 0.4. Fig. 4(a) an igure 4(b) how he acce improvemen when he arrival rae i 30MB/ec an 80MB/ec repecively. Figure 4(a) (b) alo reveal ha when he average number o he preeche iem increae, he acce improvemen will increae a well 8 CONCLUSION The ue o large cale parallel ik yem coninue o rie a he eman or aa-inenive applicaion wih large capaciie grow. Traiional orage yem cale up orage capaciy by employing more har ik rive, which en o be an expenive oluion ue o ever increaing co or HDD. Wih he new evoluionary echnology on orage componen like oli-ae ik, hybri orage yem, which conain boh high-en orage componen (SSD an HDD) an low-co orage componen (ape), will be an ieal oluion or nex generaion o aa-inenive applicaion a peacale or exracale level. In hybri orage yem, juiciouly ranerring aa back an orh among SSD, HDDS, an ape i criical or I/O perormance. We propoe a mulilayer preeching algorihm (PreHySy) ha can reuce miing rae o high-en orage componen hereby reucing he average repone ime or aa reque in hybri orage yem. To valiae PreHySy, we alo buil an analyical moel ha can mahemaically evaluae he perormance improvemen (i.e. he average acce ime improvemen) when preeching i carrie ou. The conribuion o hi reearch are wo ol. Fir, o he be o our knowlege, hi i he ir ime muli-layer preeching echnique are propoe in he conex o hybri orage yem. Secon, we preen he ir mahemaical moel o evaluae he preeching algorihm or hybri orage yem. For uure work, we will urher improve our PreHySy algorihm an analyical moel by applying hem o real worl hybri orage yem wih aa-inenive applicaion. ACKNOWLEDGMENT The work repore in hi paper wa uppore by he US Naional Science Founaion uner Gran CNS an Texa A&M Univeriy-Kingville uner ar up un. REFERENCES [] M. Shenalon, R Heller, A Chai, P O Brown, an R W Davi, Parallel human genome analyi: microarray-bae expreion monioring o 000 gene, Proc. Na Aca Sci U S A., vol. 93, no. 20, pp , Oc. 996 [2] J.Hawkin an M. Boen, M., The applicabiliy o recurren neural nework or biological equence analyi, IEEE/ACM Tran. Compuaional Biology an Bioinormaic, vol. 2, no. 3, pp , July-Sep [3] D.B. Trizna, Microwave an HF mulirequency raar or ual-ue coaal remoe ening applicaion, Proc. MTS/IEEE OCEANS, pp , Sep [4] H. Eom an J.K. Hollingworh, Spee v. accuracy in imulaion or I/O-inenive applicaion, Proc. In l Symp. Parallel an Diribue Proceing Sympoium, pp , May UbiCC Journal Page 8

10 Special Iue o ICIT 20 Conerence [5] W.W. Huell, A. Marz, Caching i oli wih ik opion: oli-ae, ik-bae cache can improve he perormance o luggih nework- Sorage, Communicaion New, Dec [6] L. Tian, D. Feng, H. Jiang, K. Zhou, L. Zeng, J. Chen, Z. Wang an Z. Song, PRO: A Populariy-bae Muli-hreae Reconrucion Opimizaion or RAID-Srucure Sorage Syem, Proc. 5h USENIX conerence on File an Sorage Technologie (FAST 07),San Joe, CA, February, [7] S.H. Baek e al., Reliabiliy an perormance o hierarchical RAID wih muliple conroller, Proc. wenieh annual ACM ympoium on principle o iribue compuing, pp , 200. [8] G.R. Ganger, B.L. Worhingon, R.Y. Hou an Y.N. Pa, Dik array: high-perormance,highreliabiliy orage ubyem, IEEE Compuer, Volume 27, Iue 3, pp , March 994. [9] T.Gerel, Sream an anar: elivering mobile vieo, ACM Queue, Vol. 3, Iue 4, pp , May [0] R. Hou, J. Menon, an Y. Pa, Balancing I/O Repone Time an Dik Rebuil Time in a RAID5 Dik Array, Proc. he Hawaii In l Con. on Syem Science, page 70-79, 993. [] G.F. Hughe an J.F. Murray, Reliabiliy an ecuriy o RAID orage yem an D2D archive uing SATA ik rive, ACM Tranacion on Sorage, Vol., Iue, pp , February [2] A.A. Drapeau an R.H. Kaz, Sripe ape array, Proc. 2h IEEE Sympoium on Ma Sorage Syem, 993. [3] C. Georgiai, P. Trianaillou, C. Falouo, Funamenal o Scheuling an Perormance o Vieo Tape Librarie, Mulimeia Tool an Applicaion, pp , [4] J. Li an S. Prabhakar, Daa Placemen or eriary orage, Proc. 9h IEEE Sympoium on Ma Sorage Syem, pp , April [5] D. Koz, D an C.S. Elli, Caching an wrieback policie in parallel ile yem, Proc. he hir IEEE Sympoium on Parallel an Diribue Proceing, 99. [6] R. Karela, J. Spencer Love, an B. Wherry, Caching Sraegie o Improve Dik Syem Perormance, IEEE Compuer, vol. 27, no. 3, pp , March 994. [7] J.Chang an G.S. Sohi, cooperaive cache pariioning or chip muliproceor, In Proc. ICS 07, [8] S.Kim, D. Chanra, an Y.Solihin, Fair Cache Sharing an Pariioning in a Chip Muliproceor Archiecure, In Proc. PACT 04, [9] R.Lyer, QoS: A ramework or enabling qo in hare cache cmp plaorm, In Proc.ICS 04, pp , [20] WOODY Huell, Aaron Marz, Caching i oli wih ik opion: oli-ae, ik-bae cache can improve he perormance o luggih nework-sorage, Communicaion New, Dec [2] R.R.M. Rabbah, H. Sananagobalane, M. Ekpanyapong, an W.-F. Wong, Compiler Preeching via peculaion an preicion, Proc. o he h Inernaional Conerence on Archiecural uppor or programming language an operaing yem, [22] J.-L, Baer an T.-F. Chen, An eecive on-chip preloaing cheme o reuce aa acce penaly, Proc. Supercompuing, pp , 99. [23] Y.Chen, S. Byna, X.-H. Sun, R. Thakur, W. Gropp, Exploring Parallel I/O Concurrency wih Speculaive Preeching, Proc. 37h Inernaional Conerence on Parallel Proceing,2008. [24] K.MK.M. Curewiz, P. Krihnan, an J.S. Vier, Pracical Preeching via Daa Compreion, Proc. ACM Con. Managemen o Daa (SIGMOD 93), pp , June 993. [25] J. Griioen an R. Appleon, Reucing File Syem Laency Uing a Preicive Approach, Proc. USENIX Ann. Technical Con. (USENIX 95), pp , Jan [26] U.U. Hann, W. Dilling, D. Kalea, Improve aapive replacemen algorihm or ik cache in HSM yem, Proc. 6h IEEE Sympoium on Ma Sorage Syem, pp28-40, 999. [27] X..-B. Zhang, D.-S. He, D. Du, an Y.-P. Lu, Objec Placemen in Parallel Tape Sorage, ICPP [28] Z. Jiang an L. Kleinrock, An Aapive Nework Preech cheme, IEEE Journal on Selece Area in Communicaion, pp , 998. [29] L. Kleinrock, Queuing Syem Vol. 2: Compuer Applicaion, Wiley, UbiCC Journal Page 9