MARK-OPT Fat-Btree Fat-Btree Fat-Btree

Transcription

1 DEWS26 7C-o3 Btree Fat-Btree, {yoshihara,daik}@de.cs.titech.ac.jp, taguchi.r-cs@nhk.or.jp,, yokota@cs.titech.ac.jp Fat-Btree MARK-OPT Fat-Btree Fat-Btree MARK-OPT Fat-Btree Fat-Btree DB A Concurrency Control Method to Reduce the Cost of Request Transfers Abstract on the Fat-Btree, Parallel Btree Structure Tomohiro YOSHIHARA, Dai KOBAYASHI, Ryo TAGUCHI, and Haruo YOKOTA, Department of Computer Science, Graduate School of Information Science and Engineering, Tokyo Institute of Technology Ookayama, Meguro-ku, Tokyo , Japan Science and Technical Research Laboratories, Japan Broadcasting Corporation Kinuta, Setagaya-ku, Tokyo , Japan Global Scientic Information and Computing Center, Tokyo Institute of Technology Ookayama, Meguro-ku, Tokyo , Japan {yoshihara,daik}@de.cs.titech.ac.jp, taguchi.r-cs@nhk.or.jp,, yokota@cs.titech.ac.jp As an effective concurrency control method for the Fat-Btree, the MARK-OPT method has been proposed. However, the latch-couplings in the method enlarge the costs of request transfers on the Fat-Btree. In this paper, a new concurrency control method not latch-coupling during optimistic operations latch coupling is proposed to improve the system performance. To detect access path errors in the proposed method, the index pages have boundary values at both ends. In the Fat-Btree, a page split may cause a page deletion. Swap of deleted page and non-deleted page in the method is more effective than normal page deletion. To compare the performance of the proposed method and the MARK-OPT method, we implemented them on an autonomous-disk system adopting the Fat-Btree. The experimental results on varied environments indicate that the proposed method always improves the system throughput, and are especially effective for large scale configuration and high workloads. Key words Access Path, Concurrency Control and Recovery, Parallel Distributed DB

2 1. root page PE Processing Element index page PE PE [1], [2] [3] 1 Fat-Btree I/O Fat-Btree PE 3 Btree Btree MARK-OPT Btree Fat-Btree INC- 1 PE MARK- OPT Btree Fat-Btree [4] [8] Fat-Btree [4] 2. Fat-Btree ncube3 [6] LAN PC [7] Fat-Btree Btree 5. INC-OPT [9] INC-OPT Btree 2. Fat-Btree B-OPT [1] ARIES/IM [11] [9] INC-OPT 2. 1 Fat-Btree Btree SMO: Structure Modication Operation Btree Btree PE Btree INC-OPT MARK-OPT [12] PE Btree MARK-OPT SMO OPT SMO SMO PE Btree Fat-Btree PE PE [13] [15] MARK-OPT PE PE1 PE2 PE3 leaf page Fat-Btree [4] [8] B + -tree PE Btree PE Btree PE [12] PE MARK-OPT [16] MARK-OPT Btree

3 1 Mode IS IX S SIX X IS IX S SIX X [9] INC-OPT SMO 2. 2 Fat-Btree PE PE PE X PE PE 2. 4 MARK-OPT MARK-OPT MARKing OPTimistic [12] Btree INC OPT SMO X Btree MARK-OPT Btree SMO SMO 1 X X 5 IS IX S SIX X 1 [17] MARK-OPT IS IS S 1 IS S IX X X SIX 1 PE 2 IS PE 1 IS IX PE S S SIX X MARK-OPT PE Fat-Btree Fat-Btree Btree 1 : IX [9], [12] X 2.1 Btree 1 2 S SIX SMO X SMO 3 X 2 Btree B-OPT [1] OPT- DLOCK [18] ARIES/IM [11] B-OPT OPT-DLOCK ARIES/IM Fat-Btree INC-OPT [9] Fat-Btree Btree INC-OPT Btree B-OPT [1] ARIES/IM [11] [17]

4 + &!" % $!" % $ $#%! "! a b 3 &' ( *)!" # $!" # $ a b 2 2 : 1 B-link [19], [2] X X ARIES/IM [11] B-link SMO 1 X B-link X ARIES/IM [12] SMO X MARK-OPT LSN Log Sequence Number SMO INC-OPT SMO X H SMO 1 X SMO INC-OPT 2 3. Fat-Btree MARK- PE OPT Fat-Btree MARK-OPT Fat-Btree PE 3 3 a Key i 2 a a a I1 MARK-OPT I1 I2 2 b a I2 I1 I2 3 b a I2 MARK- Key i I3 OPT 2.1 MARK- a I3 OPT Btree I a

5 3 Btree 1 l := H; 2 Parent := null; Child := ROOT; h := 1; m := 1; 3 while h < l do begin 4 IX latch on Child; a Key i I2 5 if Access path error is detected then 6 Unlatch Child; goto 2; 7 if Child is not full then I2 8 m := h; /* Marking */ 9 Determine NewChild; Key i-1 1 Parent := Child; Child := NewChild; h := h + 1; 11 Unlatch Parent; 12 end; 13 X latch on Child and its copies; Key i-1 Key i 14 if Access path error is detected then 15 Release all granted latches; goto 2; 16 if Child is full then 17 Release all granted latches; l := m; 18 else begin 19 Update; Key i 2 Release all granted latches; operation completion; 21 end; 22 Parent := null; Child := ROOT; h := 1; m := 1; 23 while h < l do begin 4 24 IX latch on Child; 25 if Access path error is detected then 26 Unlatch Child; goto 22; 27 if Child is not full then 28 m := h; /* Marking */ 29 Determine NewChild; 3 Parent := Child; Child := NewChild; h := h + 1; 31 Unlatch Parent; 32 end; 33 X latch on Child and its copies; 34 if Access path error is detected then Key i Key i 35 Release all granted latches; goto 22; 36 Determine NewChild; 37 Parent := Child; Child := NewChild; h := h + 1; 38 while h < 3. 2 = H do begin 39 X latch on Child and its copies; 4 Determine NewChild; MARK-OPT 41 Parent := Child; Child := NewChild; h := h + 1; 42 end; IS 43 if X latch are not sufficient for SMOs then 44 Release all granted latches; l := m; goto 22; 45 else begin 46 Update including the SMOs; 47 Release all granted latches; operation completion; 1 48 end; 2 3 IS [22] delete bit S delete bit 3. 3 delete bit 1 MARK-OPT delete bit 1 H h 1 H l X l H Btree m Fat-Btree 5 Fat-Btree PE 3 Fat-Btree 6 a b 3 fanout c [21] The rst phase The second phase 5

6 ! "! #%$ &' # $ %! " OS: Linux b Java VM: Sun J2SE SDK Server VM!#" $ 3 2 No. of Nodes: 4 64 (Storages), 16 (Clients) CPU: AMD Athlon XP-M 18+ (1.53GHz) Memory: PC21 DDR SDRAM 1GB Network 1BASE-T Hard Drive: TOSHIBA MK319GAX (3GB, 54rpm, 2.5inch) Page size: 4KB Tuple size: 35B Max No. of entries in an index node (fanout): 64 Max No. of tuples in a leaf node: 8 ac 6 Fat-Btree $ " " " -. +,! %$ &'$!!!!! ( ( )* # # delete bit c 7 4. Fat- Btree [13] [15] Linux Java 2 PC 3. 4 Fat-Btree PE 1 Btree PE PE PE Fat-Btree Fat-Btree Fat-Btree

7 ? > >=4 < 8 ; : : :2 7 2 = B C D E ABDCFEHGJILKNM FHG JILKM!ONQP < ; : / 52 4 / /. -, A :@ ! "#"$&%(')$*+$!"# $%&')(#*,+.- / 8 1 > ; = < 7 ; : A B > = 9 =< ; FG /H&I- J&KFL-'&M"N O P Q R SUTVFXWZY\[^]L( BAC BAC BAC BAC BAC BAC =< ; 3 96 > 3 7 = 2 7 : = 8 92 ; 7 3 :2 7 E 9 = :!"$#$%&')(+* '-,.'! #"$%'&)(+*,-./ 9 Fat-Btree 11 PC 4. 4 PC 1 8 PE % 1% MARK-OPT Fat-Btree MARK-OPT MARK-OPT 4. 3 MARK-OPT PC 1 8 1% MARK-OPT % MARK-OPT PC 1 9 PE % MARK-OPT MARK-OPT 9 Fat-Btree MARK-OPT MARK-OPT Fat-Btree 1 PC

8 512 [1] George Copeland, William Alexander, Ellen Boughter, and Tom Keller. Data Placement in Bubba. In Proc. of ACM SIGMOD'88, MARK-OPT pp , PC [2] Shahram Ghandeharizadeh and David J. DeWitt. Hybrid-Range Partitioning Strategy: A New Declustering Strategy for Multiprocessor Database Machines. In Proc. of VLDB'9, pp , [3] David J. DeWitt and Jim Gray. Parallel Database Systems: The Future of High Performance Database Systems. Communications of the ACM, Vol. 35, No. 6, pp , [4],,.. AI97-44, DE97-77, [5] Haruo Yokota, Yasuhiko Kanemasa, and Jun Miyazaki. Fat-Btree: An Update-Conscious Parallel Directory Structure. In Proc. of Btree ICDE'99, pp , Fat-Btree [6],. Fat-Btree. DBS , [7],. Fat-Btree MARK-OPT. DEWS21 7A-7, 21. [8] Jun Miyazaki, Yohei Abe, and Haruo Yokota. Availabilities and Costs of Reliable Fat-Btrees. In Proc. of PRDC24, pp , 24. [9] Jun Miyazaki and Haruo Yokota. Concurrency Control and Performance Evaluation of Parallel B-tree Structures. IEICE Trans. Inf. Syst., Vol. E85-D, No. 8, pp , 22. [1] Rudolf Bayer and Mario Schkolnick. Concurrency of Operations on B-trees. Acta Inf., Vol. 9, No. 1, pp. 1 21, [11] C. Mohan and Frank Levine. ARIES/IM: An Efficient and High Concurrency Index Management Method Using Write-Ahead Logging. delete bit In Proc. of ACM SIGMOD'92, pp , [12],,,,. Fat-Btree. DEWS25 2A-o4, 25. [13] Haruo Yokota. Autonomous Disks for Advanced Database Applications. In Proc. of DANTE'99, pp , [14] Daisuke Ito and Haruo Yokota. Automatic Reconguration of an Autonomous Disk Cluster. In Proc. of PRDC21, pp , 21. [15] Haruo Yokota and Ryota Abe. Secondary Storage Conguration for Advanced Data Engineering. In Nontraditional Database Syst., chapter 14, pp Taylor & Francis, 22. B-link [19], [2] Fat-Btree B-link [16],,,,,. Btree. DBS , 25. [17] Jim Gray and Andreas Reuter. Transaction Processing: Concepts and Techniques. Morgan Kaufmann, B-link Fat-Btree [18] V. Srinivasan and Michael J. Carey. Performance of B-Tree Concurrency Control Algorithms. In Proc. of ACM SIGMOD'91, pp , [19] Philip L. Lehman and S. Bing Yao. Efficient Locking for Concurrent MARK-OPT INC-OPT Operations on B-trees. ACM Trans. Database Syst., Vol. 6, No. 4, pp , PO-P [23] PO-P [2] Vladimir Lanin and Dennis Shasha. A Symmetric Concurrent B-tree INC-OPT PO [24] Algorithm. In Proc. of FJCC'86, pp , [21] Theodore Johnson and Dennis Shasha. Utilization of B-trees with Btree Inserts, Deletes and Modies. In Proc. of PODS'89, pp , [22] H. T. Kung and Philip L. Lehman. Concurrent Manipulation of Binary Search Trees. ACM Trans. Database Syst., Vol. 5, No. 3, pp. Fat-Btree , 198. [23] Damdinsuren Amarmend, Masayoshi Aritsugi, and Yoshinari Kanamori. PO-P: A Concurrency Control Protocol for Parallel B- trees. IPSJ Trans. Databases, Vol. 45, No. SIG14, pp. 3 38, 24. CREST SRC [24] Y. Mond and Yoav Raz. Concurrency Control in B+-Trees Databases Using Preparatory Operations. In Proc. of VLDB'85, pp , COE