Index Compression

Adapted from Lectures by Prabhakar Raghavan (Yahoo and Stanford) and Christopher Manning (Stanford)

Prasad

L07IndexCompression

Last lecture index construction

n n

Key step in indexing sort This sort was implemented by exploiting diskbased sorting
n n

Fewer disk seeks Hierarchical merge of blocks

n n

Distributed indexing using MapReduce Running example of document collection: RCV1

Prasad

L07IndexCompression

Today
n

Collection statistics in more detail (RCV1) Dictionary compression Postings compression

Prasad

L07IndexCompression

Recall Reuters RCV1

n n n n n

symbol N L M

statistic documents avg. # tokens per doc terms (= word types) avg. # bytes per token
(incl. spaces/punct.)

value 800,000 200 ~400,000 6 4.5

avg. # bytes per token

(without spaces/punct.)

n n
Prasad

avg. # bytes per term 7.5 non-positional postings 100,000,000

L07IndexCompression 4

Why compression?
n

Keep more stuff in memory (increases speed, caching effect)

n

enables deployment on smaller/mobile devices

Increase data transfer from disk to memory

n

[read compressed data and decompress] is faster than [read uncompressed data]

Premise: Decompression algorithms are fast

n

True of the decompression algorithms we use

L07IndexCompression 5

Prasad

Compression in inverted indexes

n

First, we will consider space for dictionary

n

Make it small enough to keep in main memory Reduce disk space needed, decrease time to read from disk Large search engines keep a significant part of postings in memory

Then the postings

n

(Each postings entry is a docID)

Prasad

L07IndexCompression

Index parameters vs. what we index

(details Table 5.1 p80)
size of word types (terms) dictionary Size (K) Unfiltered No numbers Case folding 30 stopwords 150 stopwords stemming 484 474 391 391 -2 -0 -0 -2 -19 -19 -19 -33 392 -17 % cumul % non-positional postings non-positional index Size (K) 109,971 100,680 96,969 -8 -3 -8 -12 -24 -39 -42 % positional postings positional index % -9 0 cumul % -9 -9 -38 -52 -52

cumul Size (K) % 197,879 179,158 179,158

83,390 -14 67,002 -30 63,812 -4

121,858 -31 94,517 -47 94,517 0

322 -17

Exercise: give intuitions for all the 0 entries. Why do some zero entries correspond to big deltas in other columns?

Lossless vs. lossy compression

n

Lossless compression: All information is preserved.

n

What we mostly do in IR. Several of the preprocessing steps can be viewed as lossy compression: case folding, stop words, stemming, number elimination. Chap/Lecture 7: Prune postings entries that are unlikely to turn up in the top k list for any query.
n

Lossy compression: Discard some information

n

Almost no loss of quality for top k list.

L07IndexCompression 8

Prasad

Vocabulary vs. collection size

n

Heaps Law: M = kTb M is the size of the vocabulary, T is the number of tokens in the collection. Typical values: 30 k 100 and b 0.5. In a log-log plot of vocabulary vs. T, Heaps law is a line.
L07IndexCompression 9

n n

Prasad

Heaps Law
For RCV1, the dashed line log10M = 0.49 log10T + 1.64 is the best least squares fit. Thus, M = 101.64T0.49 so k = 101.64 44 and b = 0.49.

Fig 5.1 p81

Prasad

L07IndexCompression

10

Zipfs law
n n

We also study the relative frequencies of terms. In natural language, there are a few very frequent terms and very many rare terms. Zipfs law: The ith most frequent term has frequency proportional to 1/i . cfi 1/i = c/i where c is a normalizing constant cfi is collection frequency: the number of occurrences of the term ti in the collection.
L07IndexCompression 11

n n

Prasad

Zipf consequences
n

If the most frequent term (the) occurs cf1 times

n

then the second most frequent term (of) occurs cf1/2 times the third most frequent term (and) occurs cf1/3 times

Equivalent: cfi = c/i where c is a normalizing factor, so n log cfi = log c - log i n Linear relationship between log cfi and log i
L07IndexCompression 12

Prasad

Compression
n

First, we will consider space for dictionary and postings n Basic Boolean index only n No study of positional indexes, etc. n We will devise compression schemes

Prasad

L07IndexCompression

13

DICTIONARY COMPRESSION
Prasad L07IndexCompression 14

Why compress the dictionary

n

Must keep in memory n Search begins with the dictionary n Memory footprint competition n Embedded/mobile devices

Prasad

L07IndexCompression

15

Dictionary storage - first cut

n

Array of fixed-width entries

n

~400,000 terms; 28 bytes/term = 11.2 MB.

Terms a aachen . zulu Freq. 656,265 65 . 221 Postings ptr.

Dictionary search structure

20 bytes

4 bytes each
16

L07IndexCompression

Fixed-width terms are wasteful

n

Most of the bytes in the Term column are wasted we allot 20 bytes for 1 letter terms.
n

And we still cant handle supercalifragilisticexpialidocious.

n n

Written English averages ~4.5 characters/word. Ave. dictionary word in English: ~8 characters
n

How do we use ~8 characters per dictionary term?

Short words dominate token counts but not token type (term) average.
L07IndexCompression 17

Prasad

Compressing the term list: Dictionary-as-a-String

n

Store dictionary as a (long) string of characters:

Pointer to next word shows end of current word nHope to save up to 60% of dictionary space.
n

.systilesyzygeticsyzygialsyzygyszaibelyiteszczecinszomo.
Freq. 33 29 44 126 Postings ptr. Term ptr.

Total string length = 400K x 8B = 3.2MB Pointers resolve 3.2M positions: log23.2M = 22bits = 3bytes

Prasad

L07IndexCompression

18

Space for dictionary as a string

n n n n n

4 bytes per term for Freq. Now avg. 11 4 bytes per term for pointer to Postings. bytes/term, not 20. 3 bytes per term pointer Avg. 8 bytes per term in term string 400K terms x 19 7.6 MB (against 11.2MB for fixed width)

Prasad

L07IndexCompression

19

Blocking
n

Store pointers to every kth term string.

n

Example below: k=4.

.7systile9syzygetic8syzygial6syzygy11szaibelyite8szczecin9szomo.

Need to store term lengths (1 extra byte)

Freq. 33 29 44 126 7

Postings ptr. Term ptr.

Save 9 bytes on 3 pointers.

Lose 4 bytes on term lengths.

20

Net
n

Where we used 3 bytes/pointer without blocking

n

3 x 4 = 12 bytes for k=4 pointers,

now we use 3+4=7 bytes for 4 pointers.

Shaved another ~0.5MB; can save more with larger k.

Why not go with larger k?
Prasad L07IndexCompression 21

Exercise
n

Estimate the space usage (and savings compared to 7.6 MB) with blocking, for block sizes of k = 4, 8 and 16.
n n

For k = 8. For every block of 8, need to store extra 8 bytes for length For every block of 8, can save 7 * 3 bytes for term pointer Saving (+8 21)/8 * 400K = 0.65 MB
L07IndexCompression 22

Prasad

Dictionary search without blocking

Assuming each dictionary term equally likely in query (not really so in practice!), average number of comparisons = (1+22+43+4)/8 ~2.6
Prasad L07IndexCompression 23

Dictionary search with blocking

Binary search down to 4-term block;

n

Then linear search through terms in block.

Blocks of 4 (binary tree), avg. = (1+22+23+24+5)/8 = 3 compares

L07IndexCompression 24

Prasad

Exercise
n

Estimate the impact on search performance (and slowdown compared to k=1) with blocking, for block sizes of k = 4, 8 and 16.
n

logarithmic search time to get to get to (n/k) leaves linear time proportional to k/2 for subsequent search through the leaves closed-form solution not obvious

Prasad

L07IndexCompression

25

Front coding
n

Front-coding:
Sorted words commonly have long common prefix store differences only n (for last k-1 in a block of k) 8automata8automate9automatic10automation
n

8automat*a1e2ic3ion

Encodes automat

Extra length beyond automat.

Begins to resemble general string compression.

RCV1 dictionary compression

Technique Fixed width String with pointers to every term Blocking k=4 Blocking + front coding Size in MB 11.2 7.6 7.1 5.9

Prasad

L07IndexCompression

27

POSTINGS COMPRESSION
Prasad L07IndexCompression 28

Postings compression
n

n n n

The postings file is much larger than the dictionary, by a factor of at least 10. Key desideratum: store each posting compactly. A posting for our purposes is a docID. For Reuters (800,000 documents), we would use 32 bits per docID when using 4-byte integers. Alternatively, we can use log2 800,000 20 bits per docID. Our goal: use a lot less than 20 bits per docID.
L07IndexCompression 29

Prasad

Postings: two conflicting forces

n

A term like arachnocentric occurs in maybe one doc out of a million we would like to store this posting using log2 1M ~ 20 bits. A term like the occurs in virtually every doc, so 20 bits/posting is too expensive.
n

Prefer 0/1 bitmap vector in this case

Prasad

L07IndexCompression

30

Postings file entry

n

We store the list of docs containing a term in increasing order of docID.

n

computer: 33,47,154,159,202 33,14,107,5,43

Consequence: it suffices to store gaps.

n

Hope: most gaps can be encoded/stored with far fewer than 20 bits.

Prasad

L07IndexCompression

31

Three postings entries

Prasad

L07IndexCompression

32

Variable length encoding

n

Aim:
n

For arachnocentric, we will use ~20 bits/gap entry. For the, we will use ~1 bit/gap entry.

If the average gap for a term is G, we want to use ~log2G bits/gap entry. Key challenge: encode every integer (gap) with ~ as few bits as needed for that integer. Variable length codes achieve this by using short codes for small numbers
L07IndexCompression 33

Prasad

Variable Byte (VB) codes

n

For a gap value G, use close to the fewest bytes needed to hold log2 G bits Begin with one byte to store G and dedicate 1 bit in it to be a continuation bit c If G 127, binary-encode it in the 7 available bits and set c =1 Else encode Gs lower-order 7 bits and then use additional bytes to encode the higher order bits using the same algorithm At the end set the continuation bit of the last byte to 1 (c =1) and of the other bytes to 0 (c =0). 34

Example
docIDs gaps VB code 00000110 10111000 824 829 5 10000101 215406 214577 00001101 00001100 10110001

Postings stored as the byte concatenation

00000110 10111000 10000101 00001101 00001100 10110001

Key property: VB-encoded postings are uniquely prefix-decodable. For a small gap (5), VB uses a whole byte.

Prasad

Other variable codes

n

Instead of bytes, we can also use a different unit of alignment: 32 bits (words), 16 bits, 4 bits (nibbles) etc. Variable byte alignment wastes space if you have many small gaps nibbles do better in such cases.
L07IndexCompression 36

Prasad

Gamma codes
n

Can compress better with bit-level codes

n

The Gamma code is the best known of these.

n n

Represent a gap G as a pair length and offset offset is G in binary, with the leading bit cut off
n

For example 13 1101 101 For 13 (offset 101), this is 3. Encode length in unary code: 1110.

length is the length of offset

n n

Gamma code of 13 is the concatenation of length and offset: 1110101

L07IndexCompression 37

Prasad

Gamma code examples

number 0 1 2 3 4 9 13 24 511 1025
Prasad

length 0 10 10 110 1110 1110 11110 111111110 11111111110

offset

-code none 0 0 1 00 001 101 1000 10,0 10,1 110,00 1110,001 1110,101 11110,1000 111111110,11111111 11111111110,0000000001
38

11111111 0000000001

L07IndexCompression

Exercise
n

Given the following sequence of -coded gaps, reconstruct the postings sequence:

1110001110101011111101101111011

From these -codes -- decode and reconstruct gaps, then full postings.
39

Gamma code properties

n

Uniquely prefix-decodable, like VB All gamma codes have an odd number of bits G is encoded using 2 log G +1 bits
n

Almost within a factor of 2 of best possible

Prasad

L07IndexCompression

40

Gamma seldom used in practice

n

Machines have word boundaries 8, 16, 32 bits Compressing and manipulating at individual bitgranularity will slow down query processing Variable byte alignment is potentially more efficient Regardless of efficiency, variable byte is conceptually simpler at little additional space cost
L07IndexCompression 41

Prasad

RCV1 compression
Data structure dictionary, fixed-width dictionary, term pointers into string with blocking, k = 4 with blocking & front coding collection (text, xml markup etc) collection (text) Term-doc incidence matrix postings, uncompressed (32-bit words) postings, uncompressed (20 bits) postings, variable byte encoded postings, -encoded
Prasad L07IndexCompression

Size in MB 11.2 7.6 7.1 5.9 3,600.0 960.0 40,000.0 400.0 250.0 116.0 101.0
42

Index compression summary

n

n n

n n

We can now create an index for highly efficient Boolean retrieval that is very space efficient Only 4% of the total size of the collection Only 10-15% of the total size of the text in the collection However, weve ignored positional information Hence, space savings are less for indexes used in practice
n

But techniques substantially the same.

L07IndexCompression 43

Prasad

Models of Natural Language

Necessary for analysis : Size estimation

44

Text properties/model
n

How are different words distributed inside each document?

n

Zipfs Law: The frequency of ith most frequent word is 1/i times that of the most frequent word. 50% of the words are stopwords.

How are words distributed across the documents in the collection?

n

Fraction of documents containing a word k times follows binomial distribution.

45

Probability of occurrence of a symbol depends on previous symbol. (Finite-Context or Markovian Model) The number of distinct words in a document (vocabulary) grows as the square root of the size of the document. (Heaps Law) The average length of non-stop words is 6 to 7 letters. 46

Similarity
n

Hamming Distance between a pair of strings of same length is the number of positions that have different characters. Levenshtein (Edit) Distance is the minimum number of character insertions, deletions, and substituitions needed to make two strings the same. (Extensions include transposition, weighted operations, etc) UNIX diff utility uses Longest Common Subsequence, obtained by deletion, to align strings/words/lines.

47

Index Size Estimation Using Zipfs Law

SKIP : from old slides

48

Corpus size for estimates

n

Consider N = 1M documents, each with about L=1K terms. Avg 6 bytes/term incl. spaces/punctuation
n

6GB of data.

Say there are m = 500K distinct terms among these.

49

Recall: Dont build the matrix

n

A 500K x 1M matrix has half-a-trillion 0s and 1s (500 billion). But it has no more than one billion 1s.
n

matrix is extremely sparse. Devised query processing for it

So we devised the inverted index

n

Now let us estimate the size of index

50

Rough analysis based on Zipf

n

n n n n

The i th most frequent term has frequency proportional to 1/i Let this frequency be c/i. 500, 000 Then i =1 c / i = 1.

The k th Harmonic number is H k = i =11 / i. Thus c = 1/Hm , which is ~ 1/ln m = 1/ln(500k) ~ 1/13. So the i th most frequent term has frequency roughly 1/13i.

51

Postings analysis (contd.)

n

Expected number of occurrences of the i th most frequent term in a doc of length L is: Lc/i L/13i 76/i for L=1000.

Let J = Lc ~ 76. Then the J most frequent terms are likely to occur in every document. Now imagine the term-document incidence matrix with rows sorted in decreasing order of term frequency:
52

Informal Observations
n

Most frequent term appears approx 76 times in each document. 2nd most frequent term appears approx 38 times in each document. 76th most frequent term appears approx once in each document. First 76 terms appear at least once in each document. Next 76 terms appear at least once in every two documents.
L07IndexCompression 53

Prasad

Rows by decreasing frequency

N docs
J most frequent terms. J next most frequent terms. J next most frequent terms.

N gaps of 1 each. N/2 gaps of 2 each. N/3 gaps of 3 each. m terms

etc.

54

J-row blocks
n

n n

In the i th of these J-row blocks, we have J rows each with N/i gaps of i each. Encoding a gap of i using Gamma codes takes 2log2 i +1 bits. So such a row uses space ~ (2N log2 i )/i bits. For the entire block, (2N J log2 i )/i bits, which in our case is ~ 1.5 x 108 (log2 i )/i bits. Sum this over i from 1 up to m/J = 500K/76 6500. (Since there are m/J blocks.)
55

Exercise
n

Work out the above sum and show it adds up to about 55 x 150 Mbits, which is about 1GByte. So weve taken 6GB of text and produced from it a 1GB index that can handle Boolean queries!
n

Neat! (16.7%)

Make sure you understand all the approximations in our probabilistic calculation.

56