TMTOWTDI, plus benchmarking

There's a very common Perl idiom for getting "top N elements" from an array: @top = (sort @a) [0 .. $n - 1]. Mostly, it's good enough for anything one would dare to store in RAM.

Then, there is Sort::Key::Top, which allows you to write @top = top $n => @a. Yet another syntax sugar?

Not even close! While the docs don't state it boldly, it is:

  1. a XS based module
  2. partition-based general selection algorithm (also known as quickselect) implementation

So, expect it to be fast. How fast?

Here's the results from my attempt of benchmarking the 10 longest-word selection from system dictionary (total of 235886 words):

               Rate   quicksort    pureperl quickselect
quicksort   0.329/s          --        -75%        -89%
pureperl     1.30/s        295%          --        -57%
quickselect  3.04/s        825%        134%          --

Note that the quicksort is there only to verify this claim from the Wikipedia article:

However, if done properly, a Java implementation is typically a magnitude (10x) faster than the quicksort algorithm.

Yup, seems like so. BTW, the 10 longest words are:


Benchmark code:

#!/usr/bin/env perl
use 5.010000;
use autodie;
use strict;
use warnings qw(all);

use Carp qw(croak);
use Benchmark qw(cmpthese);

use Sort::Key::Top qw(rnkeytopsort);

my %words;
open my $fh, q(<), q(/usr/share/dict/words);
while (<$fh>) {
    $words{$_} = length;
close $fh;

say q(words in hash: ) . scalar keys %words;

my $top_n = 10;
my $code = {
    quickselect => sub { rnkeytopsort { $words{$_} } $top_n => keys %words },
    quicksort => sub {
        use sort qw(_quicksort stable);
            sort { $words{$b} <=> $words{$a} }
            keys %words
        ) [0 .. $top_n - 1];
    pureperl => sub {
            sort { $words{$b} <=> $words{$a} }
            keys %words
        ) [0 .. $top_n - 1];
croak qq(something went VERY wrong)
    unless [$code->{quickselect}->()] ~~ [$code->{pureperl}->()];

cmpthese(100 => $code);


This is a slow laptop, so I only benchmarked the first 5000 words in /usr/share/dict/words, but the following pure Perl solution seems to be faster than all of the above! (Though only 11% faster than quickselect.)

pureperl2 => sub {
	my @top;
	for (keys %words) {
		if (@top < 10 or $words{$_} > $words{$top[0]}) {
			@top = sort { $words{$a} <=> $words{$b} } $_, @top;
			shift @top if @top > 10;
	return reverse @top;

(Granted, it's not quite as pithy.)

True; it doesn't work great for large $top_n. But pureperl2 does have a potential optimisation. When inserting $_ into @top, it should be possible to take advantage of the fact that @top is already sorted, and insert the new word into the right place without needing to sort it again.

(I couldn't get it to work without sacrificing sort stability, but it should be doable.)

I've optimised your pureperl3 slightly more, using the trick of eliminating brace-delimited blocks. Blocks are lexical scopes and thus there is some overhead each time you enter and exit one.

pureperl3 => sub {
	my @top;
	my $cur;
	for my $word (keys %words) {
		next if $words{$word} < $words{$top[0]||'a'};
		(@top = $word) unless @top;
		my ($low, $high) = (0, scalar @top);
			($cur = ($low + $high) >> 1),
			($words{$word} - $words{$top[$cur]} > 0)
				? ($low = $cur + 1)
				: ($high = $cur),
		) while $low < $high;
		splice(@top, $low, 0, $word);
		shift @top if @top > $top_n;
	return reverse @top;

Benchmarking the new pureperl3 on $top_n = 1000 with 483523 words, it's 15% faster than quickselect!

Leave a comment

About stas

user-pic Just another lazy, impatient and arrogant IT guy.