Meet the Walkers! Accelerating Index Traversals for In-Memory Databases"

Transcription

1 Meet the Walkers! Accelerating Index Traversals for In-Memory Databases Onur Kocberber Boris Grot, Javier Picorel, Babak Falsafi, Kevin Lim, Parthasarathy Ranganathan

2 Our World is Data-Driven! Data resides in huge databases Most frequent task: find data Indexes used for fast data lookup Indexing efficiency is critical Rely on pointer-intensive data structures Many requests, abundant parallelism Power-limited hardware Index! Data! Need high-throughput and energy-efficient index lookups 2

3 Index Lookups on General-Purpose Cores Index Lookups! OoO Cores! Data in memory Pointer-chasing! Low MLP Inherent parallelism Limited OoO inst. window Index Lookups OoO Core Inst. Window One lookup at a time Memory OoO cores are ill-matched to indexing 3

4 Widx: an Indexing Widget Specialized: Custom HW for index lookups Switch fewer transistors per lookup Parallel: Multiple lookups at once Extract parallelism beyond the OoO exec. window Programmable: Simple RISC cores Target a wide range of DBMSs 3x higher throughput, 5.5x energy reduction vs. OoO 4

5 Outline Introduction Indexing in database systems Indexing inefficiencies in modern processors Widx Evaluation highlights Summary 5

6 Modern Databases & Indexing Indexes are essential for all database operators Data structures for fast data lookup Hash index: fundamental index structure Dominant operation: join via hash index K! Hash Index R T C Q E K Walk 6

7 Join via Hash Index Lookup Join: find on the index matching for every values entry in A in and A B A Hash B Index on B Join Walk 7

8 How Much Time is Spent Indexing?! Measurement on Xeon 5670 CPU with 100GB Dataset % of Execution Time Index Scan Sort & Join Other TPC-H TPC-DS Indexing is the biggest contributor to execution time 8

9 Dissecting Index Lookups Hash: Avg. 30% time of each lookup Computationally intensive, high cache locality Walk: Avg. 70% time of each lookup Trivial computation, low cache locality Next lookup: Inherently parallel Beyond the inst. window capacity OoO Core Inst. Window 9

10 Outline Introduction Indexing in Database Systems Indexing inefficiencies in modern processors Widx Evaluation highlights Summary 10

11 Roadmap for! Efficient and High-Throughput Indexing 1. Specialize Customize hardware for hashing and walking 2. Parallelize Perform multiple index lookups at a time 3. Generalize Use a programmable building block 11

12 Step 1: Specialize Design a dedicated unit for hash and walk Hash: compute hash values from a key list Walk: access the hash index and follow pointers Hash Walk General-purpose Specialized OoO hash and walk hardware 12

13 Step 2: Parallelize Serial Decoupled Decoupled & Parallel Time H W H W H H W

14 Widx unit: common building block for hash and walk Two-stage RISC core Custom ISA Widx units are programmable Step 3: Generalize Execute functions written in Widx ISA Support limitless number of data structure layouts hash( ) H walk( ) W Widx unit 14

15 Putting it all together: Widx When Widx runs, core goes idle Widx Walkers W W W OoO Core Result Hash Producer W Widx H P MMU L1 Simple, parallel hardware

16 Development Write code for each unit and compile for Widx ISA Hash DBMS Index Code Programming Model hash (arg1, arg2, ) { } Walk walk (arg1, arg2, ) { } Res. Produce emit (arg1, arg2, ) { } Execution Communicate Load the code query-specific inputs H W P 16

17 Outline Introduction Indexing in Database Systems Indexing inefficiencies in modern processors Widx Evaluation highlights Summary 17

18 Methodology Flexus simulation infrastructure [Wenisch '06] Benchmarks TPC-H on MonetDB TPC-DS on MonetDB Dataset: 100GB uarch Parameters Core Types Area and Power Synopsys Design Compiler Technology node: TSMC 40 nm, std. cell Frequency: 2GHz Widx Area: 0.24mm 2 Widx Power: 0.3W OoO: 4-wide, 128-entry ROB In-order: 2-wide Frequency: 2GHz L1 (I & D): 32KB LLC: 4MB 18

19 Widx Performance 6 OoO Widx Indexing Speedup qry2 qry11 qry17 qry19 qry20 qry22 qry5 qry37 qry40 qry52 qry64 qry82 TPC-H TPC-DS 3x higher indexing throughput 19

20 2.5 Widx Efficiency Runtime Energy Energy-Delay x reduction in indexing energy vs. OoO 20

21 Conclusions Indexing is essential in modern DBMSs Modern CPUs spend significant time in index lookups Not efficient & fall short of extracting parallelism Widx: Specialized widget for index lookups Efficient, parallel & programmable 3x higher throughput, 5.5x energy reduction vs. OoO 21

22 Thanks! 22