Next Generation Indexes For Big Data Engineering (Waterloo University, May 2018)

Maximizing performance in data engineering is a daunting challenge. We present some of our work on designing faster indexes, with a particular emphasis on compressed indexes. Some of our prior work includes (1) Roaring indexes which are part of multiple big-data systems such as Spark, Hive, Druid, Atlas, Pinot, Kylin, (2) EWAH indexes are part of Git (GitHub) and included in major Linux distributions.

We will present ongoing and future work on how we can process data faster while supporting the diverse systems found in the cloud (with upcoming ARM processors) and under multiple programming languages (e.g., Java, C++, Go, Python). We seek to minimize shared resources (e.g., RAM) while exploiting algorithms designed for the single-instruction-multiple-data (SIMD) instructions available on commodity processors. Our end goal is to process billions of records per second per core.

The talk will be aimed at programmers who want to better understand the performance characteristics of current big-data systems as well as their evolution. The following specific topics will be addressed:

1. The various types of indexes and their performance characteristics and trade-offs: hashing, sorted arrays, bitsets and so forth.

2. Index and table compression techniques: advances in integer compression, dictionary coding.

Transcript

1. Daniel Lemire, Waterloo University, May 10th 2018.
2. Next Generation Indexes For Big Data Engineering Daniel Lemire and collaborators blog: https://lemire.me twitter: @lemire Université du Québec (TÉLUQ) Montreal Daniel Lemire, Waterloo University, May 10th 2018.
3. Knuth on performance Premature optimization is the root of all evil Daniel Lemire, Waterloo University, May 10th 2018.
4. Knuth on performance Premature optimization is the root of all evil (...) After a programmer knows which parts of his routines are really important, a transformation like doubling up of loops will be worthwhile. Daniel Lemire, Waterloo University, May 10th 2018.
5. Constants matter fasta benchmark: elapsed time total time (all processors) single‑threaded 1.36 s 1.36 s https://benchmarksgame‑team.pages.debian.net/benchmarksgame/performance/fasta.html Daniel Lemire, Waterloo University, May 10th 2018.
6. Constants matter fasta benchmark: elapsed time total time (all processors) single‑threaded 1.36 s 1.36 s multicore (4 cores) 1.00 s 2.00 s Daniel Lemire, Waterloo University, May 10th 2018.
7. Constants matter fasta benchmark: elapsed time total time (all processors) single‑threaded 1.36 s 1.36 s multicore (4 cores) 1.00 s 2.00 s vectorized (1 core) 0.31 s 0.31 s https://lemire.me/blog/2018/01/02/multicore‑versus‑simd‑instructions‑the‑fasta‑case‑study/ Daniel Lemire, Waterloo University, May 10th 2018.
8. “One Size Fits All”: An Idea Whose Time Has Come and Gone (Stonebraker, 2005) Daniel Lemire, Waterloo University, May 10th 2018. 8
9. Rediscover Unix In 2018, Big Data Engineering is made of several specialized and re‑usable components: Calcite : SQL + optimization Hadoop etc. Daniel Lemire, Waterloo University, May 10th 2018. 9
10. "Make your own database engine from parts" We are in a Cambrian explosion, with thousands of organizations and companies building their custom high‑speed systems. Specialized used cases Heterogeneous data (not everything is in your Oracle DB) Daniel Lemire, Waterloo University, May 10th 2018. 10
11. For high‑speed in data engineering you need... Front‑end (data frame, SQL, visualisation) High‑level optimizations Indexes (e.g., Pilosa, Elasticsearch) Great compression routines Specialized data structures .... Daniel Lemire, Waterloo University, May 10th 2018. 11
12. Sets A fundamental concept (sets of documents, identifiers, tuples...) → For performance, we often work with sets of integers (identifiers). Daniel Lemire, Waterloo University, May 10th 2018. 12
13. tests : x ∈ S? intersections : S ∩ S , unions : S ∪ S , differences : S ∖ S Similarity (Jaccard/Tanimoto): ∣S ∩ S ∣/∣S ∪ S ∣ Iteration f o r x i n S d o p r i n t ( x ) 2 1 2 1 2 1 1 1 1 2 Daniel Lemire, Waterloo University, May 10th 2018. 13
14. How to implement sets? sorted arrays ( s t d : : v e c t o r < u i n t 3 2 _ t > ) hash tables ( j a v a . u t i l . H a s h S e t < I n t e g e r > , s t d : : u n o r d e r e d _ s e t < u i n t 3 2 _ t > ) … bitmap ( j a v a . u t i l . B i t S e t ) compressed bitmaps Daniel Lemire, Waterloo University, May 10th 2018. 14
15. Arrays are your friends w h i l e ( l o w < = h i g h ) { i n t m I = ( l o w + h i g h ) > > > 1 ; i n t m = a r r a y . g e t ( m I ) ; i f ( m < k e y ) { l o w = m I + 1 ; } e l s e i f ( m > k e y ) { h i g h = m I - 1 ; } e l s e { r e t u r n m I ; } } r e t u r n - ( l o w + 1 ) ; Daniel Lemire, Waterloo University, May 10th 2018. 15
16. Hash tables value x at index h(x) random access to a value in expected constant‑time much faster than arrays Daniel Lemire, Waterloo University, May 10th 2018. 16
17. in‑order access is kind of terrible [15, 3, 0, 6, 11, 4, 5, 9, 12, 13, 8, 2, 1, 14, 10, 7] [15, 3, 0, 6, 11, 4, 5, 9, 12, 13, 8, 2, 1, 14, 10, 7] [15, 3, 0, 6, 11, 4, 5, 9, 12, 13, 8, 2, 1, 14, 10, 7] [15, 3, 0, 6, 11, 4, 5, 9, 12, 13, 8, 2, 1, 14, 10, 7] [15, 3, 0, 6, 11, 4, 5, 9, 12, 13, 8, 2, 1, 14, 10, 7] [15, 3, 0, 6, 11, 4, 5, 9, 12, 13, 8, 2, 1, 14, 10, 7] (Robin Hood, linear probing, MurmurHash3 hash function) Daniel Lemire, Waterloo University, May 10th 2018. 17
18. Set operations on hash tables h 1 < - h a s h s e t h 2 < - h a s h s e t . . . f o r ( x i n h 1 ) { i n s e r t x i n h 2 / / c a c h e m i s s ? } Daniel Lemire, Waterloo University, May 10th 2018. 18
19. "Crash" Swift v a r S 1 = S e t < I n t > ( 1 . . . s i z e ) v a r S 2 = S e t < I n t > ( ) f o r i i n d { S 2 . i n s e r t ( i ) } Daniel Lemire, Waterloo University, May 10th 2018. 19
20. Some numbers: half an hour for 64M keys size time (s) 1M 0.8 8M 22 64M 1400 Maps and sets can have quadratic‑time performance https://lemire.me/blog/2017/01/30/maps‑and‑sets‑can‑have‑quadratic‑time‑performance/ Rust hash iteration+reinsertion https://accidentallyquadratic.tumblr.com/post/153545455987/rust‑hash‑iteration‑reinsertion Daniel Lemire, Waterloo University, May 10th 2018. 20
21. Daniel Lemire, Waterloo University, May 10th 2018. 21
22. Bitmaps Efficient way to represent sets of integers. For example, 0, 1, 3, 4 becomes 0 b 1 1 0 1 1 or "27". {0} → 0 b 0 0 0 0 1 {0, 3} → 0 b 0 1 0 0 1 {0, 3, 4} → 0 b 1 1 0 0 1 {0, 1, 3, 4} → 0 b 1 1 0 1 1 Daniel Lemire, Waterloo University, May 10th 2018. 22
23. Manipulate a bitmap 64‑bit processor. Given x , word index is x / 6 4 and bit index x % 6 4 . a d d ( x ) { a r r a y [ x / 6 4 ] | = ( 1 < < ( x % 6 4 ) ) } Daniel Lemire, Waterloo University, May 10th 2018. 23
24. How fast is it? i n d e x = x / 6 4 - > a s h i f t m a s k = 1 < < ( x % 6 4 ) - > a s h i f t a r r a y [ i n d e x ] | - m a s k - > a O R w i t h m e m o r y One bit every ≈ 1.65 cycles because of superscalarity Daniel Lemire, Waterloo University, May 10th 2018. 24
25. Bit parallelism Intersection between {0, 1, 3} and {1, 3} a single AND operation between 0 b 1 0 1 1 and 0 b 1 0 1 0 . Result is 0 b 1 0 1 0 or {1, 3}. No branching! Daniel Lemire, Waterloo University, May 10th 2018. 25
26. Bitmaps love wide registers SIMD: Single Intruction Multiple Data SSE (Pentium 4), ARM NEON 128 bits AVX/AVX2 (256 bits) AVX‑512 (512 bits) AVX‑512 is now available (e.g., from Dell!) with Skylake‑X processors. Daniel Lemire, Waterloo University, May 10th 2018. 26
27. Bitsets can take too much memory {1, 32000, 64000} : 1000 bytes for three values We use compression! Daniel Lemire, Waterloo University, May 10th 2018. 27
28. Git (GitHub) utilise EWAH Run‑length encoding Example: 000000001111111100 est 00000000 − 11111111 − 00 Code long runs of 0s or 1s efficiently. https://github.com/git/git/blob/master/ewah/bitmap.c Daniel Lemire, Waterloo University, May 10th 2018. 28
29. Complexity Intersection : O(∣S ∣ + ∣S ∣) or O(min(∣S ∣, ∣S ∣)) In‑place union (S ← S ∪ S ): O(∣S ∣ + ∣S ∣) or O(∣S ∣) 1 2 1 2 2 1 2 1 2 2 Daniel Lemire, Waterloo University, May 10th 2018. 29
30. Roaring Bitmaps http://roaringbitmap.org/ Apache Lucene, Solr et Elasticsearch, Metamarkets’ Druid, Apache Spark, Apache Hive, Apache Tez, Netflix Atlas, LinkedIn Pinot, InfluxDB, Pilosa, Microsoft Visual Studio Team Services (VSTS), Couchbase's Bleve, Intel’s Optimized Analytics Package (OAP), Apache Hivemall, eBay’s Apache Kylin. Java, C, Go (interoperable) Roaring bitmaps 30
31. Hybrid model Set of containers sorted arrays ({1,20,144}) bitset (0b10000101011) runs ([0,10],[15,20]) Related to: O'Neil's RIDBit + BitMagic Roaring bitmaps 31
32. Roaring bitmaps 32
33. Roaring All containers are small (8 kB), fit in CPU cache We predict the output container type during computations E.g., when array gets too large, we switch to a bitset Union of two large arrays is materialized as a bitset... Dozens of heuristics... sorting networks and so on Roaring bitmaps 33
34. Use Roaring for bitmap compression whenever possible. Do not use other bitmap compression methods (Wang et al., SIGMOD 2017) Roaring bitmaps 34
35. Unions of 200 bitmaps bits per stored value bitset array hash table Roaring census1881 524 32 195 15.1 weather 15.3 32 195 5.38 cycles per input value: bitset array hash table Roaring census1881 9.85 542 1010 2.6 weather 0.35 94 237 0.16 Roaring bitmaps 35
36. Sometimes you do want arrays!!! But you'd like to compress them up. N ot always: compression can be counterproductive. Still, if you must compress, you want to do it fast Integer compression 36
37. Integer compression "Standard" technique: VByte, VarInt, VInt Use 1, 2, 3, 4, ... byte per integer Use one bit per byte to indicate the length of the integers in bytes Lucene, Protocol Buffers, etc. Integer compression 37
38. varint‑GB from Google VByte: one branch per integer varint‑GB: one branch per 4 integers each 4‑integer block is preceded byte a control byte Integer compression 38
39. Vectorisation Stepanov (STL in C++) working for Amazon proposed varint‑G8IU Use vectorization (SIMD) P atented Fastest byte‑oriented compression technique (until recently) SIMD‑Based Decoding of Posting Lists, CIKM 2011 https://stepanovpapers.com/SIMD_Decoding_TR.pdf Integer compression 39
40. Observations from Stepanov et al. We can vectorize Google's varint‑GB, but it is not as fast as varint‑G8IU Integer compression 40
41. Stream VByte Reuse varint‑GB from Google But instead of mixing control bytes and data bytes, ... We store control bytes separately and consecutively... Daniel Lemire, Nathan Kurz, Christoph Rupp Stream VByte: Faster Byte‑Oriented Integer Compression Information Processing Letters 130, 2018 Integer compression 41
42. Integer compression 42
43. Stream VByte is used by... Redis (within RediSearch) https://redislabs.com upscaledb https://upscaledb.com Trinity https://github.com/phaistos‑networks/Trinity Integer compression 43
44. Dictionary coding Use, e.g., by Apache Arrow Given a list of values: "Montreal", "Toronto", "Boston", "Montreal", "Boston"... Map to integers 0, 1, 2, 0, 2 Compress integers: Given 2 distinct values... Can use n‑bit per values (binary packing, patched coding, frame‑of‑reference) n Integer compression 44
45. Dictionary coding + SIMD dict. size bits per value scalar AVX2 (256‑bit) AVX‑512 (512‑bit) 32 5 8 3 1.5 1024 10 8 3.5 2 65536 16 12 5.5 4.5 (cycles per value decoded) https://github.com/lemire/dictionary Integer compression 45
46. To learn more... Blog (twice a week) : https://lemire.me/blog/ GitHub: https://github.com/lemire Home page : https://lemire.me/en/ CRSNG : F aster C ompressed I ndexes O n N ext‑G eneration H ardware (2017‑2022) Twitter @lemire @lemire 46