Parallel::Loops - Execute loops using parallel forked subprocesses
use Parallel::Loops;
my $maxProcs = 5;
my $pl = Parallel::Loops->new($maxProcs);
my @parameters = ( 0 .. 9 );
# We want to perform some hefty calculation for each @input and
# store each calculation's result in %output. For that reason, we
# "tie" %output, so that changes to %output in any child process
# (see below) are automatically transfered and updated in the
# parent also.
my %returnValues;
$pl->share( \%returnValues );
$pl->foreach( \@parameters, sub {
# This sub "magically" executed in parallel forked child
# processes
# Lets just create a simple example, but this could be a
# massive calculation that will be parallelized, so that
# $maxProcs different processes are calculating sqrt
# simultaneously for different values of $_ on different CPUs
# (Do see 'Performance' / 'Properties of the loop body' below)
$returnValues{$_} = sqrt($_);
});
foreach (@parameters) {
printf "i: %d sqrt(i): %f\n", $_, $returnValues{$_};
}
You can also use @arrays instead of %hashes, and/or while loops instead of foreach:
my @returnValues;
$pl->share(\@returnValues);
my $i = 0;
$pl->while ( sub { $i++ < 10 }, sub {
# This sub "magically" executed in parallel forked
# child processes
push @returnValues, [ $i, sqrt($i) ];
});
And you can have both foreach and while return values so that $pl->share() isn't required at all:
my $maxProcs = 5;
my $pl = Parallel::Loops->new($maxProcs);
my %returnValues = $pl->foreach( [ 0..9 ], sub {
# Again, this is executed in a forked child
$_ => sqrt($_);
});
Often a loop performs calculations where each iteration of the loop does not depend on the previous iteration, and the iterations really could be carried out in any order.
This module allows you to run such loops in parallel using all the CPUs at your disposal.
Return values are automatically transfered from children to parents via %hashes or @arrays, that have explicitly been configured for that sort of sharing via $pl->share(). Hashes will transfer keys that are set in children (but not cleared or unset), and elements that are pushed to @arrays in children are pushed to the parent @array too (but note that the order is not guaranteed to be the same as it would have been if done all in one process, since there is no way of knowing which child would finish first!)
If you can see past the slightly awkward syntax, you're basically getting foreach and while loops that can run in parallel without having to bother with fork, pipes, signals etc. This is all handled for you by this module.
$pl->foreach($arrayRef, $childBodySub)
Runs $childBodySub->() with $_ set foreach element in @$arrayRef, except that $childBodySub is run in a forked child process to obtain parallelism. Essentially, this does something conceptually similar to:
foreach(@$arrayRef) {
$childBodySub->();
}
Any setting of hash keys or pushing to arrays that have been set with $pl->share() will automagically appear in the hash or array in the parent process.
If you like loop variables, you can run it like so:
$pl->foreach( \@input, sub {
my $i = $_;
.. bla, bla, bla ... $output{$i} = sqrt($i);
}
);
$pl->while($conditionSub, $childBodySub [,$finishSub])
Essentially, this does something conceptually similar to:
while($conditionSub->()) {
$childBodySub->();
}
except that $childBodySub->() is executed in a forked child process. Return values are transfered via share() like in "foreach loop" above.
Note that incrementing $i in the $childBodySub like in this example will not work:
$pl->while( sub { $i < 5 },
sub {
$output{$i} = sqrt($i);
# Won't work!
$i++
}
);
Because $childBodySub is executed in a child, and so while $i would be incremented in the child, that change would not make it to the parent, where $conditionSub is evaluated. The changes that make $conditionSub return false eventually must take place outside the $childBodySub so it is executed in the parent. (Adhering to the parallel principle that one iteration may not affect any other iterations - including whether to run them or not)
In order to track progress, an optional $finishSub can be provided. It will
be called whenever a child finishes. The return value from the $conditionSub
is remembered and provided to the $finishSub as a reference:
my $i = 0;
my %returnValues = $pl->while (
sub { $i++ < 10 ? $i : 0 },
sub {
return ($i, sqrt($i));
},
sub {
my ($i) = @_;
printf "Child %d has finished\n", $i;
}
);
$pl->share(\%output, \@output, ...)
Each of the arguments to share() are instrumented, so that when a hash key is set or array element pushed in a child, this is transfered to the parent's hash or array automatically when a child is finished.
Note the limitation Only keys being set like $hash{'key'} = 'value' and
arrays elements being pushed like push @array, 'value' will be transfered to
the parent. Unsetting keys, or setting particluar array elements with
$array[3]='value' will be lost if done in the children. Also, if two different
children set a value for the same key, a random one of them will be seen by the
parent.
In the parent process all the %hashes and @arrays are full-fledged, and you can use all operations. But only these mentioned operations in the child processes make it back to the parent.
Note that when using share() for @returnValue arrays, the sequence of elements in @returnValue is not guaranteed to be the same as you'd see with a normal sequential while or foreach loop, since the calculations are done in parallel and the children may end in an unexpected sequence. But if you don't really care about the order of elements in the @returnValue array then share-ing an array can be useful and fine.
If you need to be able to determine which iteration generated what output, use a hash instead.
This is about blocking calls. When it comes to waiting for child processes to terminate, Parallel::ForkManager (and hence Parallel::Loops) is between a rock and a hard place. The underlying Perl waitpid function that the module relies on can block until either one specific or any child process terminate, but not for a process part of a given group.
This means that the module can do one of two things when it waits for one of its child processes to terminate:
Only wait for its own child processes This is the default and involves sleeping between checking whether a process has exited. This is the reason why the above simple examples needlessly all take at least one second. But it is safe, in that other processes can exit safely.
Block until any process terminate This is faster, but not the default as it is potentialy unsafe.
To get the unsafe behavior:
$pl->set_waitpid_blocking_sleep(0);
All set_waitpid_blocking_sleep does is setup exactly the same behavior in
Parallel::ForkManager. See Parallel::ForkManager#BLOCKING-CALLS for a
much more thorough description.
Note that no check is performed for recursive forking: If the main process encouters a loop that it executes in parallel, and the execution of the loop in child processes also encounters a parallel loop, these will also be forked, and you'll essentially have $maxProcs^2 running processes. It wouldn't be too hard to implement such a check (either inside or outside this package).
If you want some measure of exception handling you can use eval in the child like this:
my %errors;
$pl->share( \%errors );
my %returnValues = $pl->foreach( [ 0..9 ], sub {
# Again, this is executed in a forked child
eval {
die "Bogus error"
if $_ == 3;
$_ => sqrt($_);
};
if ($@) {
$errors{$_} = $@;
}
});
# Now test %errors. $errors{3} should exist as the only element
Also, be sure not to call exit() in the child. That will just exit the child and that doesn't work. Right now, exit just makes the parent fail no-so-nicely. Patches to this that handle exit somehow are welcome.
Keep in mind that a child process is forked every time while or foreach calls the provided sub. For use of Parallel::Loops to make sense, each invocation needs to actually do some serious work for the performance gain of parallel execution to outweigh the overhead of forking and communicating between the processes. So while sqrt() in the example above is simple, it will actually be slower than just running it in a standard foreach loop because of the overhead.
Also, if each loop sub returns a massive amount of data, this needs to be communicated back to the parent process, and again that could outweigh parallel performance gains unless the loop body does some heavy work too.
On the same VMware host, I ran this script in Debian Linux and Windows XP virtual machines respectively. The script runs a "no-op" sub in 1000 child processes two in parallel at a time
my $pl = Parallel::Loops->new(2);
$pl->foreach( [1..1000], sub {} );
For comparison, that took:
7.3 seconds on Linux
43 seconds on Strawberry Perl for Windows
240 seconds on Cygwin for Windows
On some platforms the fork() is emulated. Be sure to read perlfork.
E.g. on Windows, select is only supported for sockets, and not for pipes. So we use temporary files to store the information sent from the child to the parent. This adds a little extra overhead. See perlport for other platforms where there are problems with select. Parallel::Loops tests for a working select() and uses temporary files otherwise.
This module uses fork(). ithreads could have been possible too, but was not chosen. You may want to check out:
When to use forks, when to use threads ...? http://www.perlmonks.org/index.pl?node_id=709061
The forks module (not used here) http://search.cpan.org/search?query=forks
threads in perlthrtut http://perldoc.perl.org/perlthrtut.html
I believe this is the only dependency that isn't part of core perl:
use Parallel::ForkManager;
These are in perl's core:
use Storable; # Since perl v5.7.3
use IO::Handle; # Since perl 5.00307
use Tie::Array; # Since perl 5.005
use Tie::Hash; # Since perl 5.002
No bugs are known at the moment. Send any reports to [email protected].
Enhancements:
Optionally prevent recursive forking: If a forked child encounters a Parallel::Loop it should be possible to prevent that Parallel::Loop instance to also create forks.
Determine the number of CPUs so that new()'s $maxProcs parameter can be optional. Could use e.g. Sys::Sysconf, UNIX::Processors or Sys::CPU.
Maybe use function prototypes (see Prototypes under perldoc perlsub).
Then we could do something like
pl_foreach @input {
yada($_);
};
or
pl_foreach $pl @input {
yada($_);
};
instead of
$pl->foreach(\@input, sub {
yada($_);
});
and so on, where the first suggestion above means global variables (yikes!). Unfortunately, methods aren't supported by prototypes, so this will never be posssible:
$pl->foreach @input {
yada($_);
};
An alternative pointed out by the perlmonks chatterbox could be to use Devel::Declare "if I can stand pain".
See the git source on github https://github.com/pmorch/perl-Parallel-Loops
Copyright (c) 2008 Peter Valdemar Mørch [email protected]
All right reserved. This program is free software; you can redistribute it and/or modify it under the same terms as Perl itself.
Peter Valdemar Mørch <[email protected]>