10 Parallel Database Systems Chapter 14 Parallel Database

10. Parallel Database Systems Chapter 14 Parallel Database Systems 1

Fact and Database Problem v A large volume of data uses disk and large main memory w Speed(disk) << speed(RAM) << speed(microprocessor) 2

The Solution v Increase the I/O bandwidth w Data partitioning w Parallel data access v Origins (1980's): database machines w Hardware-oriented: bad cost-performance failure w Notable exception: ICL's CAFS Intelligent Search Processor v 1990's: same solution but using standard hardware components integrated in a multiprocessor w Software-oriented w Exploiting continuing technology improvements 3

Multiprocessor Objectives High-performance with better cost-performance than mainframe or vector supercomputer v Use many nodes, each with good costperformance, communicating through network v w Good cost via high-volume components w Good performance via bandwidth v Trends w Microprocessor and memory (DRAM): off-the-shelf w Network (multiprocessor edge): custom v The real challenge is to parallelize applications to run with good load balancing 4

Data Server Architecture 5

Objectives of Data Servers v Avoid the shortcomings of the traditional DBMS approach w Centralization of data and application management w General-purpose OS (not DB-oriented) v By separating the functions between w Application server (or host computer) w Data server (or database computer or back-end computer) 6

Data Server Approach: Assessment v Advantages w Integrated data control by the server (black box) w Increased performance by dedicated system w Can better exploit parallelism w Fits well in distributed environments v Potential problems w Communication overhead between application and data server – High-level interface w High cost with mainframe servers 7

Parallel Data Processing v Three ways of exploiting high-performance multiprocessor systems: w Automatically detect parallelism in sequential programs (e. g. , Fortran, OPS 5) w Augment an existing language with parallel constructs (e. g. , C*, Fortran 90) w Offer a new language in which parallelism can be expressed or automatically inferred 8

Parallel Data Processing (cont. ) v Critique w Hard to develop parallelizing compilers, limited resulting speed-up w Enables the programmer to express parallel computations but too low-level 9

Data-based Parallelism v Inter-operation w operations of the same query in parallel v Intra-operation w the same operation in parallel on different data partitions R R 1 R 2 R 3 R 4 10

Parallel DBMS v Loose definition: a DBMS implemented on a tightly coupled multiprocessor v Naturally extends to distributed databases with one server per site 11

Parallel DBMS -Objectives v Much better cost / performance than mainframe solution v High-performance through parallelism w High throughput with inter-query parallelism w Low response time with intra-operation parallelism v High availability and reliability by exploiting data replication v Extensibility with the ideal goals w Linear speed-up w Linear scale-up 12

Linear Speed-up v Linear increase in performance for a constant DB size and proportional increase of the system components (processor, memory, disk) 13

Linear Scale-up v Sustained performance for a linear increase of database size and proportional increase of the system components. 14

Barriers to Parallelism v Startup w The time needed to start a parallel operation may dominate the actual computation time v Interference w When accessing shared resources, each new process slows down the others (hot spot problem) v Skew w The response time of a set of parallel processes is the time of the slowest one v Parallel data management techniques intend to overcome these barriers 15

Parallel DBMS – Functional Architecture 16

Parallel DBMS Functions v Session w w Host interface Transaction monitoring v Request w w manager Compilation and optimization Data directory management Semantic data control Execution control v Data w w w manager Execution of DB operations Transaction management support Data management 17

Parallel System Architectures v Multiprocessor architecture alternatives w Shared memory (shared everything) w Shared disk w Shared nothing (message-passing) v Hybrid architectures w Hierarchical (cluster) w Non-Uniform Memory Architecture (NUMA) 18

Shared Memory Architectures 19

Shared-Disk Architecture Examples: DEC's VAXcluster, IBM's IMS/VS Data Sharing - network cost, extensibility, migration from uniprocessor - complexity, potential performance problem for copy coherency 20

Shared-Nothing Architecture Examples: Teradata (NCR), Non. Stop. SQL (Tandem-Compaq), Gamma (U. of Wisconsin), Bubba (MCC) - Extensibility, availability - Complexity, difficult load balancing 21

Hierarchical Architecture v v Combines good load balancing of SM with extensibility of SN Alternatives w Limited number of large nodes, e. g. , 4 x 16 processor nodes w High number of small nodes, e. g. , 16 x 4 processor nodes, has much better cost-performance (can be a cluster of workstations) 22

Shared-Memory vs. Distributed Memory v Mixes two different aspects : addressing and memory w Addressing – Single address space : Sequent, Encore, KSR – Multiple address spaces : Intel, Ncube w Physical memory – Central : Sequent, Encore – Distributed : Intel, Ncube, KSR v NUMA : single address space on distributed physical memory w Eases application portability w Extensibility 23

NUMA Architectures v Cache Coherent NUMA (CC-NUMA) w statically divide the main memory among the nodes v Cache Only Memory Architecture (COMA) w convert the per-node memory into a large cache of the shared address space 24

COMA Architecture 25

Parallel DBMS Techniques v Data placement w Physical placement of the DB onto multiple nodes w Static vs. Dynamic v Parallel data processing w Select is easy w Join (and all other non-select operations) is more difficult v Parallel query optimization w Choice of the best parallel execution plans w Automatic parallelization of the queries and load balancing v Transaction management w Similar to distributed transaction management 26

Data Partitioning v Each relation is divided in n partitions (subrelations), where n is a function of relation size and access frequency v Implementation w Round-robin – Map the i-th element to node (i mod n) – Simple but only exact-match queries w B-tree index – Support range queries but large index size w Hash function – Only exact-match queries but small index size 27

Replicated Data Partitioning v High-availability requires data replication w simple solution is mirrored disks – hurting load balancing when one node fails w more elaborate solutions achieve load balancing – interleaved or chained partitioning 28

Interleaved Partitioning r 3. 3 29

Chained Partitioning 30

Placement Directory v Perform two functions w F 1 (relname, placement attval) = lognode-id w F 2 (lognode-id) = phynode-id v In either case, the data structure for F 1 and F 2 should be available when needed at each node 31

Join Processing v Three basic algorithms for intra-operator parallelism w Parallel nested loop join: no special assumption w Parallel associative join: one relation is declustered on join attribute and equi-join w Parallel hash join: equi-join v They also apply to other complex operators such as duplicate elimination, union, intersection, etc. with minor adaptation 32

Parallel Nested Loop Join 33

Parallel Associative Join 34

Parallel Hash Join 35

Parallel Query Optimization v The objective is to select the "best" parallel execution plan for a query using the following components v Search space w Model alternative execution plans as operator trees w Left-deep vs. Right-deep vs. Bushy trees v Search strategy w Dynamic programming for small search space w Randomized for large search space v Cost model (abstraction of execution system) w Physical schema info. (partitioning, indexes, etc. ) w Statistics and cost functions 36

Execution Plans as Operators Trees 37

Load Balancing v Problems arise for intra-operator parallelism with skewed data distributions w attribute data skew (AVS) w tuple placement skew (TPS) w selectivity skew (SS) w redistribution skew (RS) w join product skew (JPS) v Solutions w sophisticated parallel algorithms that deal with skew w dynamic processor allocation (at execution time) 38

Some Parallel DBMSs v Prototypes w EDS and DBS 3 (ESPRIT) w Gamma (U. of Wisconsin) w Bubba (MCC, Austin, Texas) w XPRS (U. of Berkeley) w GRACE (U. of Tokyo) v Products w Teradata (NCR) w Non. Stop. SQL (Tandem-Compac) w DB 2 (IBM), Oracle, Informix, Ingres, Navigator (Sybase). . . 39

Research Problems v Hybrid v OS architectures support: using micro-kernels v Benchmarks to stress speedup and scale up under mixed workloads v Data placement to deal with skewed data distributions and data replication v Parallel data languages to specify independent and pipelined parallelism v Parallel query optimization to deal with mix of precompiled queries and complex ad-hoc queries v Support of higher functionality such as rules and objects 40

Question & Answer 41