Database Management Systems, 2 nd Edition. Raghu Ramakrishnan and Johannes Gehrke1 Parallel Database Systems Taken/tweaked from the Wisconsin DB book slides by Joe Hellerstein (UCB) with much of the material borrowed from Jim Gray (Microsoft Research). See also: Mike Carey CS 295 Fall 2011

Database Management Systems, 2 nd Edition. Raghu Ramakrishnan and Johannes Gehrke2 Why Parallel Access To Data? 1 Terabyte 10 MB/s At 10 MB/s 1.2 days to scan 1 Terabyte 1,000 x parallel 1.5 minute to scan. Parallelism: Divide a big problem into many smaller ones to be solved in parallel. Bandwidth

Database Management Systems, 2 nd Edition. Raghu Ramakrishnan and Johannes Gehrke3 Parallel DBMS: Intro v Parallelism is natural to DBMS processing – Pipelined parallelism: many machines each doing one step in a multi-step process. – Partitioned parallelism: many machines doing the same thing to different pieces of data. – Both are natural in DBMS! Pipeline Partition Any Sequential Program Any Sequential Program Sequential Any Sequential Program Any Sequential Program outputs split N ways, inputs merge M ways

Database Management Systems, 2 nd Edition. Raghu Ramakrishnan and Johannes Gehrke4 DBMS: The || Success Story v For a long time, DBMSs were the most (only?!) successful/commercial application of parallelism. –Teradata, Tandem vs. Thinking Machines, KSR. –Every major DBMS vendor has some || server. –(Of course we also have Web search engines now. ) v Reasons for success: –Set-oriented processing (= partition ||-ism). –Natural pipelining (relational operators/trees). –Inexpensive hardware can do the trick! –Users/app-programmers don’t need to think in ||

Database Management Systems, 2 nd Edition. Raghu Ramakrishnan and Johannes Gehrke5 Some || Terminology v Speed-Up –Adding more resources results in proportionally less running time for a fixed amount of data. v Scale-Up –If resources are increased in proportion to an increase in data/problem size, the overall time should remain constant. degree of ||-ism Xact/sec. (throughput) Ideal degree of ||-ism sec./Xact (response time) Ideal

Database Management Systems, 2 nd Edition. Raghu Ramakrishnan and Johannes Gehrke6 Architecture Issue: Shared What? Easy to program Expensive to build Difficult to scale Hard to program Cheap to build Easy to scale Sequent, SGI, Sun VMScluster, Sysplex Tandem, Teradata, SP2 (Use affinity routing to approximate SN- like non-contention)

Database Management Systems, 2 nd Edition. Raghu Ramakrishnan and Johannes Gehrke7 What Systems Work This Way Shared Nothing Teradata: 400 nodes Tandem: 110 nodes IBM / SP2 / DB2: 128 nodes Informix/SP2 48 nodes ATT & Sybase ? nodes Shared Disk Oracle170 nodes DEC Rdb 24 nodes Shared Memory Informix 9 nodes RedBrick ? nodes (as of 9/1995)

Database Management Systems, 2 nd Edition. Raghu Ramakrishnan and Johannes Gehrke8 Different Types of DBMS ||-ism v Intra-operator parallelism –get all machines working together to compute a given operation (scan, sort, join) v Inter-operator parallelism –each operator may run concurrently on a different site (exploits pipelining) v Inter-query parallelism –different queries run on different sites v We’ll focus mainly on intra-operator ||-ism

Database Management Systems, 2 nd Edition. Raghu Ramakrishnan and Johannes Gehrke9 Automatic Data Partitioning Partitioning a table: RangeHashRound Robin Shared disk and memory less sensitive to partitioning. Shared nothing benefits from "good" partitioning. A...E F...J K...NO...ST...Z A...E F...JK...NO...S T...Z A...EF...J K...N O...S T...Z Good for equijoins, exact-match queries, and range queries Good for equijoins, exact match queries Good to spread load

Database Management Systems, 2 nd Edition. Raghu Ramakrishnan and Johannes Gehrke10 Parallel Scans/Selects v Scan in parallel and merge ( a.k.a. union all). v Selection may not require all sites for range or hash partitioning, but always does for RR. v Indexes can be constructed on each partition. –Indexes useful for local accesses, as expected. –However, what about unique indexes...? (May not always want primary key partitioning!)

Database Management Systems, 2 nd Edition. Raghu Ramakrishnan and Johannes Gehrke11 Secondary Indexes A..C D..F G...M N...R S.. Base Table A..Z Base Table A..Z v Secondary indexes become a bit troublesome in the face of partitioning... v Can partition them via base table key. –Inserts local (unless unique??). –Lookups go to ALL indexes. v Can partition by secondary key ranges. –Inserts then hit 2 nodes (base, index). –Ditto for index lookups (index, base). –Uniqueness is easy, however. v Teradata’s index partitioning solution: –Partition non-unique by base table key. –Partition unique by secondary key.

Database Management Systems, 2 nd Edition. Raghu Ramakrishnan and Johannes Gehrke12 Grace Hash Join v In Phase 1 in the parallel case, partitions will get distributed to different sites: –A good hash function automatically distributes work evenly! (Diff hash fn for partitioning, BTW.) v Do Phase 2 (the actual joining) at each site. v Almost always the winner for equi-joins. Original Relations (R then S) OUTPUT 2 B main memory buffers Disk INPUT 1 hash function h B-1 Partitions 1 2 B-1... Phase 1

Database Management Systems, 2 nd Edition. Raghu Ramakrishnan and Johannes Gehrke13 Dataflow Network for || Joins v Use of split/merge makes it easier to build parallel versions of sequential join code.

Database Management Systems, 2 nd Edition. Raghu Ramakrishnan and Johannes Gehrke14 Parallel Sorting v Basic idea: –Scan in parallel, range-partition as you go. –As tuples arrive, perform “local” sorting. –Resulting data is sorted and range-partitioned (i.e., spread across system in known way). –Problem: skew! –Solution: “sample” the data at the outset to determine good range partition points.

Database Management Systems, 2 nd Edition. Raghu Ramakrishnan and Johannes Gehrke15 Parallel Aggregation v For each aggregate function, need a decomposition: – count (S) =  count (s(i)), ditto for sum () – avg (S) = (  sum (s(i))) /  count (s(i)) –and so on... v For groups: –Sub-aggregate groups close to the source. –Pass each sub-aggregate to its group’s partition site.

Database Management Systems, 2 nd Edition. Raghu Ramakrishnan and Johannes Gehrke16 Complex Parallel Query Plans v Complex Queries: Inter-Operator parallelism –Pipelining between operators: u note that sort or phase 1 of hash-join block the pipeline! –Bushy Trees A BRS Sites 1-4Sites 5-8 Sites 1-8

Database Management Systems, 2 nd Edition. Raghu Ramakrishnan and Johannes Gehrke17 Observations v It is relatively easy to build a fast parallel query executor. –S.M.O.P., well understood today. v It is hard to write a robust and world-class parallel query optimizer. –There are many tricks. –One quickly hits the complexity barrier. –Many resources to consider simultaneously (CPU, disk, memory, network).

Database Management Systems, 2 nd Edition. Raghu Ramakrishnan and Johannes Gehrke18 Parallel Query Optimization v Common approach: 2 phases –Pick best sequential plan (System R algorithm) –Pick degree of parallelism based on current system parameters. v “Bind” operators to processors –Take query tree, “decorate” it with site assignments as in previous picture.

Database Management Systems, 2 nd Edition. Raghu Ramakrishnan and Johannes Gehrke19 v Best serial plan != Best || plan! Why? v Trivial counter-example: –Table partitioned with local secondary index at two nodes –Range query: all of node 1 and 1% of node 2. –Node 1 should do a scan of its partition. –Node 2 should use secondary index. v SELECT * FROM telephone_book WHERE name < “NoGood”; What’s Wrong With That? N..Z Table Scan A..M Index Scan

Database Management Systems, 2 nd Edition. Raghu Ramakrishnan and Johannes Gehrke20 Parallel DBMS Summary v ||-ism natural to query processing: –Both pipeline and partition ||-ism! v Shared-Nothing vs. Shared-Memory –Shared-disk too, but less “standard” (~older...) –Shared-memory easy, costly. Doesn’t scaleup. –Shared-nothing cheap, scales well, harder to implement. v Intra-op, Inter-op, & Inter-query ||-ism all possible.

Database Management Systems, 2 nd Edition. Raghu Ramakrishnan and Johannes Gehrke21 || DBMS Summary, cont. v Data layout choices important! –In practice, will not N-way partition every table. v Most DB operations can be done partition-|| –Select, sort-merge join, hash-join. –Sorting, aggregation,... v Complex plans. –Allow for pipeline-||ism, but sorts and hashes block the pipeline. –Partition ||-ism achieved via bushy trees.

Database Management Systems, 2 nd Edition. Raghu Ramakrishnan and Johannes Gehrke22 || DBMS Summary, cont. v Hardest part of the equation: optimization. –2-phase optimization simplest, but can be ineffective. –More complex schemes still at the research stage. v We haven’t said anything about xacts, logging, etc. –Easy in shared-memory architecture. –Takes a bit more care in shared-nothing architecture

Database Management Systems, 2 nd Edition. Raghu Ramakrishnan and Johannes Gehrke23 || DBMS Challenges (mid-1990’s) v Parallel query optimization. v Physical database design. v Mixing batch & OLTP activities. –Resource management and concurrency challenges for DSS queries versus OLTP queries/updates. –Also online, incremental, parallel, and recoverable utilities for load, dump, and various DB reorg ops. v Application program parallelism. –MapReduce, anyone...? –(Some new-ish companies looking at this, e.g., GreenPlum, AsterData, …)