Parallel::ForkManager adds sleeps by default.

Parallel::ForkManager is only sleeping during " $pm->wait_all_children; ", correct? Or are you saying during every fork there is a hard-limit sleep?

Read http://search.cpan.org/~yanick/Parallel-ForkManager-1.17/lib/Parallel/ForkManager.pm#BLOCKING_CALLS and then try this to see if it helps:

#!/usr/local/bin/perl
use strict;
use warnings;
use Parallel::ForkManager;

my $pm = new Parallel::ForkManager(24);
$pm->set_waitpid_blocking_sleep(0);
...
[download]

Please see the many postings of BrowserUk in response to questions of the general form "Why doesn't multithreading/multiprocessing solve all my problems?" My impression is that he handles two or three applications like the one you describe before he finishes his first mug of coffee each morning — or maybe it's tea in his case.

open( my $fh, "-|", '/bin/zcat' => $file ) or die "Error opening $file
+: $!";
[download]

Update: added pattern matching to the demonstration.

Greetings,

At first glance, am not seeing NodeJS performing regex for fields containing spaces. That's not really fair, IMHO :) The regex in Perl is likely the main reason for taking longer. Having 24 workers reading IO simultaneously from the nfsv3 mounted SAN may be another reason even though NodeJS seems not affected by it. Maybe, the regex statement in Perl is the reason.

The following is an alternative solution. IO is faster for the upcoming 1.700 release versus MCE 1.608. Thus, the note on using 1.699_010. Also, inlined bits for the Windows platform.

#!/usr/local/bin/perl

use strict;
use warnings;

# IO performance in 1.699_010 is faster than MCE 1.608.
# https://metacpan.org/release/MARIOROY/MCE-1.699_010

use MCE::Loop;
use MCE::Candy;

my $dir = 'logs/*.log.gz';
my @files = sort(glob "$dir");
my $pattern = "some_string";

MCE::Loop::init {
    gather      => MCE::Candy::out_iter_fh(\*STDOUT),
    chunk_size  => '240k',
    max_workers => 24,
    use_slurpio => 1,
};

open( my $fh, "-|", "zcat", @files ) or die "open error: $!\n";

mce_loop {
    my ( $mce, $slurp_ref, $chunk_id ) = @_;
    my $buf = '';

    # Quickly determine if a match is found...
    # ...and process slurped chunk only if true.

    if ( $$slurp_ref =~ /$pattern/m ) {

        # The following is fast on Unix, but performance degrades
        # drastically on Windows beyond 4 workers.

          open my $MEM_FH, '<', $slurp_ref;
          while ( my $line = <$MEM_FH> ) {
             if ( $line =~ /$pattern/ ) {
                 my @matches = $line =~ /".*?"|\S+/g;
                 $buf .= "$matches[0],$matches[1],$matches[3],$matches
+[4]\n";
             }
          }
          close $MEM_FH;

        # Therefore, use the following construction on Windows.

        # while ( $$slurp_ref =~ /([^\n]+\n)/mg ) {
        #    # my $line = $1; # possibly save $1 to not lose the value
        #                     # not necessary for this demonstration
        #    my @matches = $1 =~ /".*?"|\S+/g;
        #    $buf .= "$matches[0],$matches[1],$matches[3],$matches[4]\
+n";
        # }
    }

    # Send output to the manager process for orderly output to STDOUT

    $mce->gather($chunk_id, $buf);

} $fh;

close $fh;
[download]

There is one reader across NFS by the manager process only, not workers. The effect is sequential IO which is typically faster than random IO.

Regards, Mario

Update: added pattern matching to the demonstration

If parallel reads is desired, the following demonstration does the same thing. Basically, specifying chunk_size => 1 is all one needs to do for getting MCE to run like Parallel::ForkManager.

#!/usr/local/bin/perl

use strict;
use warnings;
use MCE::Loop;

my $dir = 'logs/*.log.gz';
my @files = sort(glob "$dir");
my $pattern = "some_string";

MCE::Loop::init {
    max_workers => 24,
    chunk_size  => 1,
};

mce_loop {
    my ( $mce, $chunk_ref, $chunk_id ) = @_;
    open( my $fh, "-|", "zcat", $chunk_ref->[0] ) or die "open error: 
+$!\n";

    while ( my $line = <$fh> ) {
        if ( $line =~ /$pattern/ ) {
            my @matches = $line =~ /".*?"|\S+/g;
            print "$matches[0],$matches[1],$matches[3],$matches[4]\n";
        }
    }

    close $fh;

} @files;
[download]

Regards, Mario

Wow thanks, I'll give this a shot. I will have to read some more on MCE, it looks very useful. I should have clarified, in the "parser" function in nodejs, I'm applying the same regex as perl to be fair. I've done the tests looking for a specific string (/some_string/) and I've done the regex in the above code (/".*?"|\S+/g), which captures everything in an array, since the lines are in this format: ' 1970-01-01 00:00:00 1.1.1.1 "A multi-word field" 2.2.2.2 '

When timing 70 files, nodejs takes 20 seconds

Can you post the nodejs code?

This was typed up quickly from memory. I haven't tested this code yet, but you get the general idea. I'm also brand new at JavaScript so please don't judge.

'use strict'

var cluster = require('cluster');
var spawn = require('child_process').spawn;
var stream = require('stream');
var glob = require('glob');

var files = glob.sync('/data/logs/*.log.gz');

function Parser() {
    var t = new stream.Transform( { objectMode: true } );
    var transform = t._transform;
    var flush = t._flush;
    var parse = t.parse;
    return t;
}

function transform(chunk, enc, done) {
    var data = chunk.toString();
    //if flush gives us remaining data from last line, prepend to next
+ chunk
    if (this._lastLineData) data = this._lastLineData + data;
    
    var lines = data.split('\n');
    this._lastLineData = lines[lines.length-1];
    
    for (var i = 0; i < lines.length; i++) {
        this.push(parse.lines[i]);
    }
    done();
}

function flush(done) {
    //If remaining data, pass to transform to prepend to next chunk
    if (this._lastLineData) this.push(this._lastLineData);
    this._lastLineData = null
    done()
}

function parse(line) {
    //Split line into fields, or apply regex to filter results
    return line;
}

if (cluster.isMaster) {

var workers = 24; // number of CPUs
var logsPerWorker = workers / files;
var remainder = workers % files;
var logSlice {};
var sliceStart = 0;
var sliceEnd = 0;

    for (var i = 0; i < workers; i++) {
        if (remainder) {
            sliceEnd++;
            remainder--;
        }
        logSlice[i] = files.slice(sliceStart,sliceEnd);
        sliceStart++;
        sliceEnd = sliceStart;
        
        if (logsPerWorker === 0 && remainder === 0) {
            //if no more logs to distribute, break out of loop
            break;
        }
    }

    for (var i = 0; i < workers; i++) {
        //Fork and pass list of logs for child to process
        cluster.fork({logs: logSlice[i]});
    }
} else {
    var liner = Parser();
    var logs = split.env.logs(','); //read logs passed by parent
    for (var i = 0; i < logs.length; i++){
        var child = spawn('zcat', [ logs[i] ]);
        child.stdout.pipe(liner); //pipe this to stdout or to writeStr
+eam to write to a file
    }
}
[download]