Mark Nunberg
and 1 contributors

NAME

Text::Prefix::XS - Fast prefix searching

SYNOPSIS

    use Text::Prefix::XS;
    
    my @haystacks = qw(
        garbage
        blarrgh
        FOO-stuff
        meh
        AA-ggrr
        AB-hi!
    );
    
    my @prefixes = qw(AAA AB FOO FOO-BAR);
    
    my $search = prefix_search_create( map uc($_), @prefixes );
    
    my %seen_hash;
    
    foreach my $haystack (@haystacks) {
        if(my $prefix = prefix_search($search, $haystack)) {
            $seen_hash{$prefix}++;
        }
    }
    
    $seen_hash{'FOO'} == 1;
    
    {
        %seen_hash = ();
        my $re = join('|', map quotemeta $_, @prefixes);
        $re = qr/^($re)/;
        
        foreach my $haystack (@haystacks) {
            my ($match) = ($haystack =~ $re);
            if($match) {
                $seen_hash{$match}++;
            }
        }
        $seen_hash{'FOO'} == 1;
    }
    
    #Super fast:
    
    my $match_results = prefix_search_multi($search, @haystacks);
    grep $_ eq 'FOO-stuff', @{ $match_results->{FOO} };

DESCRIPTION

This module implements a variety of algorithms to ensure fast prefix matching.

It is particularly high performant with pessimistic matching (when a match is unlikely), but is still faster than other methods with optimistic matching (when a match is likely).

A common application I had was to pre-filter lots and lots of text for a small amount of preset prefixes.

Interestingly enough, the quickest solution until I wrote this module was to use a large regular expression (as in the synopsis)

FUNCTIONS

The interface is relatively simple. This is alpha software and the API is subject to change

prefix_search_create(@prefixes)

Create an opaque prefix search handle. It returns a thingy, which you should keep around.

Each prefix must be no longer than 256 bytes. For normal ASCII strings, this should be the number of characters - but does not hold true for encodings like UTF8.

prefix_search($thingy, $haystack)

Will check $haystack for any of the prefixes in @prefixes passed to "prefix_search_create". If $haystack has a prefix, it will be returned by this function; otherwise, the return value is undef.

Input strings can be text or random byte sequences (any is acceptable)

prefix_search_multi($thingy, @haystacks)

EXTREMELY FAST!!!

Will check each item in @haystacks for any of the @prefixes passed to "prefix_search_create". The return value is a hash reference. Its keys are matched prefix strings, and its values are array references containing items from @haystacks which matched.

This function is extremely fast. It's four times quicker than the normal "prefix_search" function (which is itself about twice as fast as any other method).

However, it will not gain a lot of performance benefit with optimistic searching (meaning that a match has a good chance of being found), and will just consume more memory (since it needs to store the results in a hash).

PERFORMANCE

In most normal use cases, Text::Prefix::XS will outperform any other module or search algorithm.

Specifically, this module is intended for a pessimistic search mechanism, where most of the input is assumed not to match (which is usually the case anyway).

The ideal position of Text::Prefix::XS would reside between raw but delimited user input, and more complex searching and processing algorithms. This module acts as a layer between those.

In addition to a trie, this module also uses a very fast sparse array to check characters in the input against an index of known characters at the given position. This is much quicker than a hash lookup.

See the trie.pl script included with this distribution for detailed benchmark comparison methods

Here are a bunch of numbers. The entries are in the format of

    [capture (Y/N)] NAME DURATION MATCHES
    

Where capture means whether the test was also able to return the prefix which matched. MATCHES is the amount of matches returned.

Additionally, each test has a few parameters defining the input. These are:

TERMS

The amount of search terms

TERM_MIN

The minimum length of a term

TERM_MAX

The maximum length of a term

INPUT

The count of input strings which will be checked to see if they are prefixed with any of the TERMS. The strings are each exactly one character longer than TERM_MAX

Sample input is taken by making a sha1_hex string of each number from 0 until TERMS, and then encoding that output into Base64, ensuring that both the terms and the input get a diversity of the ASCII charset.

A few methods were benchmarked, and are listed as keys:

TMFA

Text::Match::FastAlternatives match_at function

perl-re

Generic perl regex. The capturing version is qr/^(term1|term2)/, and the non-capturing version is qr/^(?:term1|term2)/, where the terms are joined together in a list2re fashion.

RE2

Same as perl-re, except using re::engine::RE2

TXS

Using a loop of "prefix_search" over the input items

TXS-Multi

Using a single function call to "prefix_search_multi"

    Generated INPUT=2000000 TERMS=20 TERM_MIN=3 TERM_MAX=6
    CAP   NAME       DUR        MATCH
    [N] TMFA            1.12s   M=23578
    [N] perl-re         1.24s   M=23578
    [N] RE2             0.91s   M=23578
    [Y] perl-re         1.44s   M=23578
    [Y] RE2             2.91s   M=23578
    [Y] TXS             0.53s   M=23578
    [Y] TXS-Multi       0.18s   M=23578
    
    Generated INPUT=2000000 TERMS=50 TERM_MIN=10 TERM_MAX=16
    CAP   NAME       DUR        MATCH
    [N] TMFA            1.14s   M=50
    [N] perl-re         1.20s   M=50
    [N] RE2             0.90s   M=50
    [Y] perl-re         1.43s   M=50
    [Y] RE2             1.10s   M=50
    [Y] TXS             0.53s   M=50
    [Y] TXS-Multi       0.17s   M=50
    
    Generated INPUT=2000000 TERMS=49 TERM_MIN=2 TERM_MAX=16
    CAP   NAME       DUR        MATCH
    [N] TMFA            1.18s   M=241799
    [N] perl-re         1.44s   M=241799
    [N] RE2             1.01s   M=241799
    [Y] perl-re         1.77s   M=241799
    [Y] RE2             4.94s   M=241799
    [Y] TXS             1.47s   M=241799
    [Y] TXS-Multi       1.15s   M=241799


    Generated INPUT=2000000 TERMS=10 TERM_MIN=5 TERM_MAX=10
    CAP   NAME       DUR        MATCH
    [N] TMFA            1.12s   M=131
    [N] perl-re         1.27s   M=131
    [N] RE2             0.97s   M=131
    [Y] perl-re         1.50s   M=131
    [Y] RE2             2.76s   M=131
    [Y] TXS             0.46s   M=131
    [Y] TXS-Multi       0.09s   M=131

    Generated INPUT=2000000 TERMS=100 TERM_MIN=3 TERM_MAX=25
    CAP   NAME       DUR        MATCH
    [N] TMFA            1.15s   M=15734
    [N] perl-re         1.26s   M=15734
    [N] RE2             0.94s   M=15734
    [Y] perl-re         1.49s   M=15734
    [Y] RE2             1.69s   M=15734
    [Y] TXS             1.06s   M=15734
    [Y] TXS-Multi       0.68s   M=15734
    
    Generated INPUT=2000000 TERMS=200 TERM_MIN=5 TERM_MAX=25
    CAP   NAME       DUR        MATCH
    [N] TMFA            1.15s   M=1300
    [N] perl-re         1.22s   M=1300
    [N] RE2             1.00s   M=1300
    [Y] perl-re         1.43s   M=1300
    [Y] RE2             1.27s   M=1300
    [Y] TXS             0.73s   M=1300
    [Y] TXS-Multi       0.24s   M=1300

    Generated INPUT=2000000 TERMS=8 TERM_MIN=2 TERM_MAX=5
    CAP   NAME       DUR        MATCH
    [N] TMFA            1.12s   M=88025
    [N] perl-re         1.22s   M=88025
    [N] RE2             0.82s   M=88025
    [Y] perl-re         1.64s   M=88025
    [Y] RE2             1.13s   M=88025
    [Y] TXS             0.63s   M=88025
    [Y] TXS-Multi       0.27s   M=88025

    

I've mainly tested this on Debian's 5.10 - for newer perls, this module performs better, and for el5 5.8, The differences are a bit lower. TBC

SEE ALSO

There are quite a few modules out there which aim for a Trie-like search, but they are all either not written in C, or would not be performant enough for this application.

These two modules are implemented in pure perl, and are not part of the comparison.

Text::Trie

Regexp::Trie

Regexp::Optimizer

Text::Match::FastAlternatives

re::engine::RE2

CAVEATS

Threads

As of version 0.15, threads are fully supported.

Optimistic Matching

The performance gains from this algorithm are less when matches are more likely (optimistic). Nevertheless, when matches are likely, chances are you want to figure out what it was that matched - in which case the performance benefits are still reaped - as this is the fastest performing capturing method.

Do not use "prefix_search_multi" with optimistic matching, as it will provide minimal speed boosts and increase your memory usage.

In the future, an interface to provide a function callback would be handy - but in the case of optimistic matching, would be bad for performance as well

Prefix Lengths

Prefixes may not exceed 256 bytes. You can increase this limit (at the cost of more memory) by changing the #define of CHARTABLE_MAX in the XS code and recompiling.

UTF-8 Support

Doing some basic tests with use utf8; and non-ascii input, it seems to work as expected. The character tables and prefix tries work at the byte level, so no conversion is done for you. This is usually not an issue, but I am no unicode expert, so nag me if you find something wrong

Many Prefixes

In the case where the prefix count becomes insanely high (i.e. over 5,000), the performance of this module will begin to drop. This could probably be solved in the future by a bunch of different methods. Even at a 10,000 prefix count, it still remains on-par with perl regular expressions.

AUTHOR AND COPYRIGHT

Copyright (C) 2011 M. Nunberg

You may use and distribute this software under the same terms, conditions, and licensing as Perl itself.