TurboPFor: Fastest Integer Compression

• 100% C (C++ compatible headers), without inline assembly

- Fastest **"Variable Byte"** implementation

- Novel **"Variable Simple"** faster than simple16 and more compact than simple8-b

- Scalar **"Bit Packing"** with bulk decoding as fast as SIMD FastPFor in realistic and practical (No "pure cache") scenarios - Bit Packing with **Direct/Random Access** without decompressing entire blocks - Access any single bit packed entry with **zero decompression** - **New:** **Direct Update** of individual bit packed entries - Reducing **Cache Pollution**

- Novel **"TurboPFor"** (Patched Frame-of-Reference) scheme with direct access or bulk decoding. Outstanding compression

- Several times faster than other libraries - Usage in C/C++ as easy as memcpy - Most functions optimized for speed and others for high compression ratio - **New:** Include more functions

- Instant access to compressed *frequency* and *position* data in inverted index with zero decompression - **New:** Inverted Index Demo + Benchmarks: Intersection of lists of sorted integers. - more than **1000 queries per second** on gov2 (25 millions documents) on a **SINGLE** core. - Decompress only the minimum necessary blocks (Ex. 10-15%).

Benchmark:

CPU: Sandy bridge i7-2600k at 4.2GHz, gcc 4.9, ubuntu 14.10, Single thread.

• Realistic and practical benchmark with large integer arrays.
• No PURE cache benchmark

Synthetic data:

• Generate and test skewed distribution.

  ./icbench -a1.5 -m0 -M8 -n100000000

Size	Ratio in %	Bits/Integer	C Time MB/s	D Time MB/s	Function
63392801	15.85	5.07	316.96	893.67	TurboPFor
63392801	15.85	5.07	315.59	227.15	TurboPForDA
65359916	16.34	5.23	7.09	638.96	OptPFD
72364024	18.09	5.79	85.31	762.00	Simple16
78514276	19.63	6.28	229.21	748.32	SimpleV
95915096	23.98	7.67	221.46	1049.70	Simple-8b
99910930	24.98	7.99	1553.92	1904.21	SIMDPackFPF
99910930	24.98	7.99	953.29	1872.02	TurboPack
99910930	24.98	7.99	953.13	869.84	TurboPackDA
102074663	25.52	8.17	1131.47	1184.68	TurboVbyte
102074663	25.52	8.17	1110.75	897.86	VbyteFPF
112500000	28.12	9.00	305.85	1899.15	VarintG8IU
400000000	100.00	32.00	1470.87	1477.93	Copy

####----------------------------------------------------

Data files

• gov2.sorted (from http://lemire.me/data/integercompression2014.html) Blocksize=128
  (+ SimpleV 64k). Benchmark repeated several times.

  ./icbench -c1 gov2.sorted

Size	Ratio in %	Bits/Integer	C Time MB/s	D Time MB/s	Function
3214763689	13.44	4.30	279.93	665.41	SimpleV 64k
3337758854	13.95	4.47	5.06	513.00	OptPFD
3357673495	14.04	4.49	270.57	813.83	TurboPFor
3501671314	14.64	4.68	258.56	720.76	SimpleV
3820190182	15.97	5.11	118.81	650.21	Simple16
4521326518	18.90	6.05	209.17	824.26	Simple-8b
4953768342	20.71	6.63	647.75	1501.24	TurboPack
5203353057	21.75	6.96	1560.34	1806.60	SIMDPackFPFD1
6699519000	28.01	8.96	502.86	624.12	TurboVbyte
6699519000	28.01	8.96	472.01	495.12	VbyteFPF
7622896878	31.87	10.20	208.73	1197.74	VarintG8IU
23918861764	100.00	32.00	1391.82	1420.03	Copy

####-------------------------------------------------------

Compressed Inverted Index Intersections with GOV2

GOV2: 426GB, 25 Millions documents, average doc. size=18k.

• Aol: 1100 queries per second
  18000 queries in 16.31s [1103.9 q/s] [0.906 ms/q]
  Ratio = 14.37% Decoded/Total Integers.

• TREC Million Query Track (1MQT): 950 queries per second
  20000 queries in 21.03s, [951.0 q/s] [1.052 ms/q]
  Ratio = 11.59% Decoded/Total Integers.

Compile:

make

Testing

Synthetic data:

1. test all functions
   

*./icbench -a1.0 -m0 -M8 -n100000000*

- zipfian distribution alpha = 1.0 (Ex. -a1.0=uniform -a1.5=skewed distribution)
- number of integers = 100000000
- integer range from 0 to 255 (integer size = 0 to 8 bits)

2. individual function test (ex. copy TurboPack TurboPack Direct access)
   

*./icbench -a1.5 -m0 -M8 -ecopy/turbopack/turbopackda -n100000000*

Data files:

• Data file Benchmark (file format as in FastPFOR)

  ./icbench -c1 gov2.sorted

Benchmarking intersections

• Download gov2 (or clueweb09) + query file (Ex. "1mq.txt")
  from "http://lemire.me/data/integercompression2014.html"

• Create index file

  ./idxcr gov2.sorted .

  create inverted index file "gov2.sorted.i" in the current directory

• Benchmarking intersections

  ./idxqry gov2.sorted.i 1mq.txt

  run queries in file "1mq.txt" over the index of gov2 file

8GB RAM required (16GB recommended for benchmarking "clueweb09" files).

Function usage:

In general compression/decompression functions are of the form:

char *endptr = compress( unsigned *in, int n, [int b,] char *out)
endptr : set by compress to the next character in "out" after the compressed buffer
in : input integer array
n : number of elements
out : pointer to output buffer
b : number of bits. Only for bit packing functions

char *endptr = decompress( char *in, int n, [int b,] unsigned *out)
endptr : set by decompress to the next character in "in" after the decompressed buffer
in : pointer to input buffer
n : number of elements
out : output integer array
b : number of bits. Only for bit unpacking functions

header files to use with documentation :

vint.h	Variable byte
vsimple.h	Variable simple
vp4dc.h, vp4dd.h	TurboPFor
bitpack.h bitunpack.h	Bit Packing

Reference:

• "SIMD-BitPack FPF" from FastPFor https://github.com/lemire/simdcomp
• Sorted integer datasets from http://lemire.me/data/integercompression2014.html
• OptPFD (OptP4) and Simple-16 from http://jinruhe.com/

#------------------------------------------------