I had a chance to benchmark various PerlIO layers. The result was striking.
Benchmarking code:
use strict;
use warnings;
use Benchmark qw(timethese cmpthese);
our $test_file = 'test.txt';
# 64 chars including LF;
my $test_line = join '', ( '0' .. '9', 'A' .. 'Z', 'a' .. 'z', '_', "\n" );
sub make_test_file {
my $kb = shift; # size in KB
my $block = $test_line x 16; # 1024 bytes;
my $fh;
open $fh, ">:raw", $test_file or die "$test_file:$!";
print {$fh} $block for ( 1 .. $kb );
close $fh;
# slurp once to make sure the file buffer is in effect;
open $fh, "<:raw", $test_file or die "$test_file:$!";
my $content = do { local $/; <$fh> };
close $fh;
}
sub make_tester {
my $layer = shift;
my $openarg = $layer ? "<:$layer" : "<";
sub {
open my $fh, $openarg, $test_file or die "$test_file:$!";
$_ ne $test_line and die while (<$fh>);
close $fh;
}
}
my %these = map { $_ => make_tester($_) } qw( raw perlio mmap );
$these{nolayer} = make_tester(q());
for my $filesize ( map { 4**$_ } ( 0 .. 4 ) ) {
make_test_file($filesize);
print "#### Testing for $filesize KB Read\n";
cmpthese( timethese( 0, \%these ) );
}
delete $these{raw}; # because this guy is too slow on large files!
for my $filesize ( map { 4**$_ } ( 5 .. 8 ) ) {
make_test_file($filesize);
print "#### Testing for $filesize KB Read\n";
cmpthese( timethese( 0, \%these ) );
}
__END__
The result on my MacBook Pro:
#### Testing for 1 KB Read
Benchmark: running mmap, nolayer, perlio, raw for at least 3 CPU seconds...
Rate raw mmap perlio nolayer
raw 213/s -- -95% -98% -98%
mmap 4492/s 2007% -- -55% -64%
perlio 10088/s 4632% 125% -- -19%
nolayer 12488/s 5758% 178% 24% --
#### Testing for 4 KB Read
Benchmark: running mmap, nolayer, perlio, raw for at least 3 CPU seconds...
Rate raw mmap perlio nolayer
raw 55.7/s -- -99% -99% -99%
mmap 3722/s 6578% -- -42% -49%
perlio 6375/s 11338% 71% -- -13%
nolayer 7305/s 13008% 96% 15% --
#### Testing for 16 KB Read
Benchmark: running mmap, nolayer, perlio, raw for at least 3 CPU seconds...
Rate raw mmap perlio nolayer
raw 13.5/s -- -99% -99% -99%
mmap 1849/s 13550% -- -25% -30%
perlio 2469/s 18123% 34% -- -7%
nolayer 2649/s 19455% 43% 7% --
#### Testing for 64 KB Read
Benchmark: running mmap, nolayer, perlio, raw for at least 3 CPU seconds...
Rate raw mmap perlio nolayer
raw 3.53/s -- -99% -100% -100%
mmap 693/s 19567% -- -10% -14%
perlio 767/s 21651% 11% -- -5%
nolayer 810/s 22872% 17% 6% --
#### Testing for 256 KB Read
Benchmark: running mmap, nolayer, perlio, raw for at least 3 CPU seconds...
Rate raw perlio nolayer mmap
raw 0.937/s -- -99% -100% -100%
perlio 187/s 19804% -- -4% -6%
nolayer 194/s 20582% 4% -- -2%
mmap 198/s 21062% 6% 2% --
#### Testing for 1024 KB Read
Benchmark: running mmap, nolayer, perlio for at least 3 CPU seconds...
Rate perlio nolayer mmap
perlio 46.8/s -- -5% -11%
nolayer 49.2/s 5% -- -7%
mmap 52.8/s 13% 7% --
#### Testing for 4096 KB Read
Benchmark: running mmap, nolayer, perlio for at least 3 CPU seconds...
Rate perlio nolayer mmap
perlio 12.4/s -- -4% -4%
nolayer 12.9/s 4% -- -1%
mmap 13.0/s 4% 1% --
#### Testing for 16384 KB Read
Benchmark: running mmap, nolayer, perlio for at least 3 CPU seconds...
Rate perlio nolayer mmap
perlio 2.96/s -- -5% -6%
nolayer 3.11/s 5% -- -1%
mmap 3.14/s 6% 1% --
#### Testing for 65536 KB Read
Benchmark: running mmap, nolayer, perlio for at least 3 CPU seconds...
s/iter perlio mmap nolayer
perlio 1.35 -- -8% -8%
mmap 1.25 8% -- -0%
nolayer 1.24 9% 0% --
no layer being the fastest is no surprise for me. It's synonym for :perlio on most platforms and with no option you have no overhead for setting layer. What's so surprising is that :raw being so slow. I got similar result on my FreeBSD boxes.
Dan the Just Another Perl Benchmarker