VERSION

This document describes Parallel::Iterator version 1.00

SYNOPSIS

use Parallel::Iterator qw( iterate );
# A very expensive way to double 100 numbers...

my @nums = ( 1 .. 100 );

my $iter = iterate( sub {
my ( $id, $job ) = @_;
return $job * 2;
}, \@nums );

my @out = ();
while ( my ( $index, $value ) = $iter->() ) {
$out[$index] = $value;
}

DESCRIPTION

The "map" function applies a user supplied transformation function to each element in a list, returning a new list containing the transformed elements.

This module provides a 'parallel map'. Multiple worker processes are forked so that many instances of the transformation function may be executed simultaneously.

For time consuming operations, particularly operations that spend most of their time waiting for I/O, this is a big performance win. It also provides a simple idiom to make effective use of multi CPU systems.

There is, however, a considerable overhead associated with forking, so the example in the synopsis (doubling a list of numbers) is not a sensible use of this module.

Example

Imagine you have an array of URLs to fetch:

    my @urls = qw(
        http://google.com/
        http://hexten.net/
        http://search.cpan.org/
        ... and lots more ...
    );

Write a function that retrieves a URL and returns its contents or undef if it can't be fetched:

    sub fetch {
        my $url = shift;
        my $resp = $ua->get($url);
        return unless $resp->is_success;
        return $resp->content;
    };

Now write a function to synthesize a special kind of iterator:

    sub list_iter {
        my @ar = @_;
        my $pos = 0;
        return sub {
            return if $pos >= @ar;
            my @r = ( $pos, $ar[$pos] );  # Note: returns ( index, value )
            $pos++;
            return @r;
        };
    }

The returned iterator will return each element of the array in turn and then undef. Actually it returns both the index and the value of each element in the array. Because multiple instances of the transformation function execute in parallel the results won't necessarily come back in order. The array index will later allow us to put completed items in the correct place in an output array.

Get an iterator for the list of URLs:

    my $url_iter = list_iter( @urls );

Then wrap it in another iterator which will return the transformed results:

    my $page_iter = iterate( \&fetch, $url_iter );

Finally loop over the returned iterator storing results:

    my @out = ( );
    while ( my ( $index, $value ) = $page_iter->() ) {
        $out[$index] = $value;
    }

Behind the scenes your program forked into ten (by default) instances of itself and executed the page requests in parallel.

Simpler interfaces

Having to construct an iterator is a pain so "iterate" is smart enough to do that for you. Instead of passing an iterator just pass a reference to the array:

    my $page_iter = iterate( \&fetch, \@urls );

If you pass a hash reference the iterator you get back will return key, value pairs:

    my $some_iter = iterate( \&fetch, \%some_hash );

If the returned iterator is inconvenient you can get back a hash or array instead:

    my @done = iterate_as_array( \&fetch, @urls );
    my %done = iterate_as_hash( \&worker, %jobs );

How It Works

The current process is forked once for each worker. Each forked child is connected to the parent by a pair of pipes. The child's STDIN, STDOUT and STDERR are unaffected.

Input values are serialised (using Storable) and passed to the workers. Completed work items are serialised and returned.

Caveats

Parallel::Iterator is designed to be simple to use - but the underlying forking of the main process can cause mystifying problems unless you have an understanding of what is going on behind the scenes.

Worker execution enviroment

All code apart from the worker subroutine executes in the parent process as normal. The worker executes in a forked instance of the parent process. That means that things like this won't work as expected:

    my %tally = ();
    my @r = iterate_as_array( sub {
        my ($id, $name) = @_;
        $tally{$name}++;       # might not do what you think it does
        return reverse $name;
    }, @names );
    # Now print out the tally...
    while ( my ( $name, $count ) = each %tally ) {
        printf("%5d : %s\n", $count, $name);
    }

Because the worker is a closure it can see the %tally hash from its enclosing scope; but because it's running in a forked clone of the parent process it modifies its own copy of %tally rather than the copy for the parent process.

That means that after the job terminates the %tally in the parent process will be empty.

In general you should avoid side effects in your worker subroutines.

Serialization

Values are serialised using Storable to pass to the worker subroutine and results from the worker are again serialised before being passed back. Be careful what your values refer to: everything has to be serialised. If there's an indirect way to reach a large object graph Storable will find it and performance will suffer.

To find out how large your serialised values are serialise one of them and check its size:

    use Storable qw( freeze );
    my $serialized = freeze $some_obj;
    print length($serialized), " bytes\n";

In your tests you may wish to guard against the possibility of a change to the structure of your values resulting in a sudden increase in serialized size:

    ok length(freeze $some_obj) < 1000, "Object too bulky?";

See the documetation for Storable for other caveats.

Performance

Process forking is expensive. Only use Parallel::Iterator in cases where:

the worker waits for I/O

The case of fetching web pages is a good example of this. Fetching a page with LWP::UserAgent may take as long as a few seconds but probably consumes only a few milliseconds of processor time. Running many requests in parallel is a huge win - but be kind to the server you're talking to: don't launch a lot of parallel requests unless it's your server or you know it can handle the load.

the worker is CPU intensive and you have multiple cores / CPUs

If the worker is doing an expensive calculation you can parallelise that across multiple CPU cores. Benchmark first though. There's a considerable overhead associated with Parallel::Iterator; unless your calculations are time consuming that overhead will dwarf whatever time they take.

INTERFACE

iterate( [ $options ], $worker, $iterator )

Get an iterator that applies the supplied transformation function to each value returned by the input iterator.

Instead of an iterator you may pass an array or hash reference and "iterate" will convert it internally into a suitable iterator.

If you are doing this you may wish to investigate "iterate_as_hash" and "iterate_as_array".

Options

A reference to a hash of options may be supplied as the first argument. The following options are supported:

"workers"

The number of concurrent processes to launch. Set this to 0 to disable forking. Defaults to 10 on systems that support fork and 0 (disable forking) on those that do not.

"nowarn"

Normally "iterate" will issue a warning and fall back to single process mode on systems on which fork is not available. This option supresses that warning.

"batch"

Ordinarily items are passed to the worker one at a time. If you are processing a large number of items it may be more efficient to process them in batches. Specify the batch size using this option.

Batching is transparent from the caller's perspective. Internally it modifies the iterators and worker (by wrapping them in additional closures) so that they pack, process and unpack chunks of work.

"adaptive"

Extending the idea of batching a number of work items to amortize the overhead of passing work to and from parallel workers you may also ask "iterate" to heuristically determine the batch size by setting the "adaptive" option to a numeric value.

The batch size will be computed as

    <number of items seen> / <number of workers> / <adaptive>

A larger value for "adaptive" will reduce the rate at which the batch size increases. Good values tend to be in the range 1 to 2.

You can also specify lower and, optionally, upper bounds on the batch size by passing an reference to an array containing ( lower bound, growth ratio, upper bound ). The upper bound may be omitted.

    my $iter = iterate(
        { adaptive => [ 5, 2, 100 ] },
        $worker, \@stuff );

"onerror"

The action to take when an error is thrown in the iterator. Possible values are 'die', 'warn' or a reference to a subroutine that will be called with the index of the job that threw the exception and the value of $@ thrown.

    iterate( {
        onerror => sub {
            my ($id, $err) = @_;
            $self->log( "Error for index $id: $err" );
        },
        $worker,
        \@jobs
    );

The default is 'die'.

iterate_as_array

As "iterate" but instead of returning an iterator returns an array containing the collected output from the iterator. In a scalar context returns a reference to the same array.

For this to work properly the input iterator must return (index, value) pairs. This allows the results to be placed in the correct slots in the output array. The simplest way to do this is to pass an array reference as the input iterator:

    my @output = iterate_as_array( \&some_handler, \@input );

iterate_as_hash

As "iterate" but instead of returning an iterator returns a hash containing the collected output from the iterator. In a scalar context returns a reference to the same hash.

For this to work properly the input iterator must return (key, value) pairs. This allows the results to be placed in the correct slots in the output hash. The simplest way to do this is to pass a hash reference as the input iterator:

    my %output = iterate_as_hash( \&some_handler, \%input );

CONFIGURATION AND ENVIRONMENT

Parallel::Iterator requires no configuration files or environment variables.

DEPENDENCIES

None.

INCOMPATIBILITIES

None reported.

AUTHOR

Andy Armstrong "<[email protected]>"

THANKS

Aristotle Pagaltzis for the END handling suggestion and patch.

LICENCE AND COPYRIGHT

Copyright (c) 2007, Andy Armstrong "<[email protected]>". All rights reserved.

This module is free software; you can redistribute it and/or modify it under the same terms as Perl itself. See perlartistic.

DISCLAIMER OF WARRANTY

BECAUSE THIS SOFTWARE IS LICENSED FREE OF CHARGE, THERE IS NO WARRANTY FOR THE SOFTWARE, TO THE EXTENT PERMITTED BY APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT HOLDERS AND/OR OTHER PARTIES PROVIDE THE SOFTWARE ``AS IS'' WITHOUT WARRANTY OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE SOFTWARE IS WITH YOU. SHOULD THE SOFTWARE PROVE DEFECTIVE, YOU ASSUME THE COST OF ALL NECESSARY SERVICING, REPAIR, OR CORRECTION.

IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MAY MODIFY AND/OR REDISTRIBUTE THE SOFTWARE AS PERMITTED BY THE ABOVE LICENCE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY GENERAL, SPECIAL, INCIDENTAL, OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE USE OR INABILITY TO USE THE SOFTWARE (INCLUDING BUT NOT LIMITED TO LOSS OF DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD PARTIES OR A FAILURE OF THE SOFTWARE TO OPERATE WITH ANY OTHER SOFTWARE), EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH DAMAGES.