Do you mean that sleep $sleep_period should be replaced with select( '','','', $sleep_period )?

Yes. Because: sleep uses whole seconds; whereas select "The timeout, if specified, is in seconds, which may be fractional.... You can effect a sleep of 250 milliseconds this way:select( undef, undef, undef, 0.25 );".

That would allow you to avoid setting $pm->set_waitpid_blocking_sleep(0);; by instead setting set_waitpid_blocking_sleep( $seconds ); to a much lower -- but non-zero! -- value than the standard one whole second (say:0.01) which would also achieve the goal of avoiding the pregnant pause between one child process ending and its replacement being started.

The reason why this latter method is preferable to setting the timeout (sleep period) to 0; is documented in the "BLOCKING CALLS" section of the P::FM docs. It basically comes down to the fact that when the timeout is set to 0; instead of using wait to wait for particular child processes; it uses waitpid to wait for any child process. In most circumstances that will be fine because the only child processes will be those that were started by, and thus being managed by, P::FM.

But, if the parent process is also starting other child processes, then waitpid can also be woken by one of those other children that shouldn't be under P::FMs control; thus a) the code that started this 'other' child won't be able to obtain its return code; b) P::FM will start another child thinking that one of its child processes has ended, when it hasn't.

Thank you very much for this in-depth explanation.

Update:

I applied your advice:

karls-mac-mini:monks karl$ ./buks_patch.pl 0.01
processes: 4
fork: 13.8262140750885

karls-mac-mini:monks karl$ ./buks_patch.pl 0.25
processes: 4
fork: 13.9020979404449

karls-mac-mini:monks karl$ ./buks_patch.pl 0.5
processes: 4
fork: 15.0743520259857

karls-mac-mini:monks karl$ ./buks_patch.pl 1.0
processes: 4
fork: 20.0701560974121
[download]

It seems like it works as expected, very nice!

For completeness, the modified code:

#!/usr/bin/env perl 

use strict;
use warnings;
use Math::BigInt;
use Parallel::ForkManager;
use Time::HiRes qw ( time );
use feature qw(say);

my $sleep_period = shift;
my $processes    = qx(sysctl -n hw.ncpu);

chomp $processes;

my $pm = Parallel::ForkManager->new($processes);

$pm->set_waitpid_blocking_sleep($sleep_period);

say qq(processes: $processes);

my $start = time;

for my $number ( 1 .. 50 ) {
    $pm->start($number) and next;
    my $factorial = Math::BigInt->bfac(2000) for 1 .. 10;
    $pm->finish(0);
}
$pm->wait_all_children;

say qq(fork: ), time - $start;
[download]

ForkManager.pm:

sub _waitpid_blocking {
    my $self = shift;
    # pseudo-blocking  
    if( my $sleep_period = $self->{waitpid_blocking_sleep} ) {
        while() {
            my $pid = $self->_waitpid_non_blocking;
            return $pid if $pid;
            # sleep $sleep_period;
            select( '','','', $sleep_period );
        }
    }

    return waitpid -1, 0;
}
[download]

My best regards, Karl

«The Crux of the Biscuit is the Apostrophe»