Programa de Verão LNCC 2015 Jornada Ciência de Dados

Transcrição

1 Progrm de Verão LNCC 0 Jornd Ciênci de Ddos Prticionmeto de Ddos Gerencimento de Grndes Volumes de Ddos Prte III Fio Porto (fporto@lncc.r) LNCC CCC - DEXL L Intuição Q Imgem de um região do céu Consults e processos de workflow sore ddos prticiondos presentm um tempo de processmento correspondente o tmnho d mior prtição Q Q Q LIneA - HQOOP //

2 Distriuição Uniforme Distriuição não uniforme Prticionmento x Crg de Trlho Prticionmento de Ddos Definição: f (ei) -> v, v um número inteiro correspondente um item de locção de ddos (por exemplo um máquin n rede) Considerções: não locl Ddo um conjunto de ddos D={ e, e,, en} definir um função de distriuição sensível à crg de trlho W={w, w,, wn} locl 8

3 Ojetivo Funções de distriuição minimizr o volume de ddos cessdo pels plicções, em um contexto de execução prlel; mximiz loclidde de cesso idelmente, cd plicção cessri loclmente (i.e. sem trnsferênci de ddos pel rede) 9 Round-Roin Histogrm chve sed em crg-trlho 0 Round-Roin Por Intervlo de vlores Ddo D={ e, e,, en} e N={ nn,, nn}, onde ni é um nó de locção de ddos f(ei) -> nj, tl que j=i, se i n, ou j=mod(i/n) Cd elemento é ssocido um nó, segundo su ordem de leitur plicção não present restrição qunto o conjunto de ddos processdo em cd nó; A função de distriuição ssoci um vlor do ojeto um intervlo conhecido do domínio Sej A={,,,.., k} P=[,..,000] P=[00,..,000].. Pn=[, k] se ei(v) = 000 então f(ei) = P

4 Intervlo de Vlores (ex idimensionl) Por Intervlo de Vlores Interessnte qundo plicções processm conjunto de ddos segundo intervlos: dt inicio dt fim intervlo de vlor de vend de produto intervlo de nomes É um cso especil de histogrms Intervlo de Vlores (ex) Intervlo de Vlores (ex)

5 Distriuição não uniforme Distriuição por Intervlo não uiformes 8 Histogrm de igul lrgur 9 Histogrms: Lrgur x Altur 0

6 Prticionmento ciente d crg de trlho Uso do Histogrm por ltur R(,,9) B B Bm Overview of the solution Prolem sttement Given Accessing lock ccessing ll its tuples Periodiclly new tuples nd queries rrive No privilege to prticulr ttriute Nodes: ech dt item (e.g. tuple) represent node in the grph Edges: n edge etween two dt items if re ccessed y common query Edge weight : the numer of queries tht ccess oth dt items Gol: prtition the grph into m equl size su-grphs with minimum edge cut Blocks re explined in terms of queries Optiml prtitioning of R s dt in locks Optiml query execution Ech lock is ssigned n explining query Bi = vi(r) Query processing Adpt to the rrivl of new dt nd queries Use min-cut lgorithm Block explntion Minimize the totl lock ccess during the execution of queries y: Gol Dt prtitioning : grph sed lgorithm Assumptions Single reltion dtse R(,,n), n ~000 Initil worklod: set of k queries W0 = {q,,qk} m empty fixed size locks Queries re compred to explining queries Mtching locks re selected (we hven t worked on tht yet)

7 Schism: VLDB00 We crete node for ech row We crete node for ech row For ech verticl frgment We crete node for ech row We crete node for ech row 8

8 For ech verticl frgment For ech verticl frgment q q We increment the rc weight when two rows re ccessed together 9 We increment the rc weight when two rows re ccessed together 0 For ech verticl frgment W = {q,,qn} For ech verticl frgment We increment the rc weight when two rows re ccessed together We execute min-cut lgorithm 8

9 Prtitioned dt with queries Ctlog B {q,q,, q} {q,q,, q} Ech lock is ssocited with the queries tht ccess Some records of the lock B Ech prtition is ssigned lock B B For given query q the numer of ccessed locks is minimized Alloction sed on ffinity to locks Adptive Strtegy New tuple rrivl: [DEXA 0] Select the est lock Chllenges: i.e. lock to which the new tuple is more correlted How to select the est lock with minimum effort? Initil pproch : find it sed on the correltion of queries to locks Define optiml lloction Compute ctul lloction efficiency Compute lock ffinity Wht if the est lock is full? Initil pproch: split the lock 9

10 Elpsed-time of incrementing the DB s the size increses 000 sttic DynPrt, D = 00 k DynPrt, D = M Execution time (s) By Key M M M 8M 0M M DB size M M 8M 0M Experiment: - Slon DR8 0 million tuples - worklod- synthetic 000 queries - PToH hyper-grph prtitioner 8 P-Grid [Aerer0] System model Completely decentrlized Use rndomized lgorithm to crete serch structure Scle grcefully in the totl numer of nodes nd totl numer of dt items; 9 Prtitioning is fine y how to void skewed distriution of dt through nodes? Are there strtegies to prtition dt into hundreds or thousnds of nodes with lnced gurnties? Lessons lerned from PP systems key distriution strtegies Set of peer nodes linked through neighourhood links; Ech peer stores suset of the keys; A virtul inry tree is uilt sed on the key vlue spce Ech peer keeps rnch of n index tree pointing to the nodes storing the keys the peer mnges Additionlly, t ech level of the tree pointers indicte the nodes keeping the complementing key set; A serch my strt t ny peer; 0 0

11 Key mngement Exmple Ech node in the tree keeps pointers to peers holding the complementry key: (p p pk) -> (p p pk) root Ech peer is responsile for n intervl of keys: k= p p pn pi={0,}, vl(k)= i=,n -ipi I(k)=[vl(k),vl(k)n] 0 For ech prefix kl of length l, peer mintins references to other peers tht hve the sme prefix s l ut different vlue t position l for the keys the peer is responsile for A inry tree orders the key spce nd points to peers mnging keys with prefixes ppering t tree nodes; Mnging the tree is independent of mnging keys index keys Strting with ,, 00 - P-Grid Construction P-Grid Construction ) Initilly ech peer points to the whole key spce c) One peer mnges longer pth (p p pk pk.. pn) () root root 0 p p p k p p pkpk.. pn d) The keys peers mnge intercept up to position k (p p pk) p p p k p p p k p p p k 0 p p pkpk.. pn (p p pk pk) ) Meet peers mnging the sme set of keys p p pkpk (0) p p p k p p pkqk.. qm (p p pk 0) Complement pointer p p pkpk.. pn Ref=refs(qk)/{} meet(ref,{}) Ref=refs(pk)/{} meet(ref,{})

12 Distriuted BY Rnge Prllelism & LodBlnce Prllel query Pgpool II Pgpool II Implemented on top of PostgreSQL 9. Centrl node coordintes dt/query distriution/ repliction Requests distriuted through nodes Prllel query Processing Repliction Pgpool II dt prtitioning sed on tle column rnge (e.g. id) For short queries, my reduce the numer of ccessed dt postgresql Lod Blnce PostgreSQL PostgreSQL Concurrent requests directed to different DB copies QServ - LSST Evlution Strength Extends PostgreSQL Lod lnce queries from concurrent workflows Scles up to 8 DB nodes Weknesses postgresql Repliction Pgpool II Lck of support to sptil functions Prtitioning sed on single column Ingestion cn t use COPY Developed y the LSST DM tem Astronomy dt mngement Horizontl prtitioning sed on declintion zones (nodes) nd dt on ech node distriuted into chunks sed on RA-chunk Approx. 000 prtitions Ntive support to sptio-temporl functions Built on top of MySQL 8

13 Evlution Considerções Finis Strong Designed to support stronomy dt surveys Highly sclle: ~000 nodes First performnce results re very promising Alignment with the LSST project Weknesses Gerencir grndes volumes de ddos requer: estrtégis escláveis foco do desenho que fvoreç loclidde de ddos trtmento de incertez trtmento d heterogeneidde prticionmento dos ddos processos de otimizção Current culture sed on PostgreSQL 9 0 Considerções Finis Frmework tuis de processmento nlíticos MpReduce (Hdoop, PivotlHD, Clouder, ) Apche Sprk Sgitri (CEFET-RJ) em reve disponível HQoop em desenvolvimento no DEXL