Speed up calculation by running in parallel

原文链接：https://perlmaven.com/speed-up-calculation-by-running-in-parallel

In this example we have a bunch of numbers. We need to make some heavy calculation on each number (using our calc function), and then collect the results in a hash where the keys are the original numbers and the values are the results.

For the sake of our example the “heavy computation” is generating and summing up lots of random numbers.

The linear solution

This is quite simple to do in linear code. We just need to iterate over the elements of the original array that holds the input numbers and call the calc function on each one of them. When the function is done it returns the result that we can then assign to the appropriate element in the hash.

examples/calc_no_fork.pl

use strict;
use warnings;
use Data::Dumper qw(Dumper);
 
my @numbers = map { $_ * 2000000 } reverse 1 .. 10;
my %results;
 
foreach my $q (@numbers) {
    $results{$q} = calc($q);
}
 
print Dumper \%results;
 
sub calc {
    my ($n) = @_;
    my $sum = 0;
    for (1 .. $n) {
        $sum += 1 + rand()/100;
    }
    return $sum;
}

(If it is unclear what is in the @numbers array and how it is generated then use the Dumper function to print the content of @numbers right after it was created and read about map in Perl.

We can run this code using the time program of the Unix/Linux shell:

$ time perl calc_no_fork.pl

The result on my computer was:

real    0m21.032s

It took 21 seconds to do all the calculations.

While the program was running I’ve also used the htop program in another console and saw that none of the 4 cores of my computer is fully used.

I think the reason that none of them is fully used is that the operating system moves the process around form CPU to CPU, but I am not 100% sure in this.

Using fork

In order to allow the computer to better utilize all its resources we could use either threads or fork. As threads are not a recommended technique in Perl we opt to use fork as described in the article on Using fork to spread load to multiple cores. The problem, as it is also mentioned at the end of that article is that forked processes don’t have shared memory, so the forked process cannot simple write back to the common %results hash. That’s where the module Parallel::ForkManager comes into play.

It is a wrapper around the regular fork function of Perl that provides us with various nice extra services. Including the possibility to return data from the child processes to the parent process that launched them.

examples/calc_fork_manager.pl

use strict;
use warnings;
use Parallel::ForkManager;
use Data::Dumper qw(Dumper);
 
my $forks = shift or die "Usage: $0 N\n";
 
my @numbers = map { $_ * 2000000 } reverse 1 .. 10;
my %results;
 
print "Forking up to $forks at a time\n";
my $pm = Parallel::ForkManager->new($forks);
 
$pm->run_on_finish( sub {
    my ($pid, $exit_code, $ident, $exit_signal, $core_dump, $data_structure_reference) = @_;
    my $q = $data_structure_reference->{input};
    $results{$q} = $data_structure_reference->{result};
});
 
foreach my $q (@numbers) {
    my $pid = $pm->start and next;
    my $res = calc($q);
    $pm->finish(0, { result => $res, input => $q });
}
$pm->wait_all_children;
 
print Dumper \%results;
 
sub calc {
    my ($n) = @_;
    my $sum = 0;
    for (1 .. $n) {
        $sum += 1 + rand()/100;
    }
    return $sum;
}

The code starts with the creation of the Parallel::ForkManager object and setting the maximum number of parallel child processes.

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

Then we create an anonymous function (a sub without a name) and pass it to the run_on_finish method of Parallel::ForkManager. This function will be called once for each child process immediately as the child process terminates. It receives a number of parameters, but the one that is interesting to us now is the 6th, the last parameter which we assigned to the $data_structure_reference variable.

This variable will hold everything we sent back from the child process. In our case that will be a hash reference with two keys. “input” will contain the value from the original @numbers array the specific child process is dealing with. The “result” will contain the value returned by the calc() function.

$pm->run_on_finish( sub {
    my ($pid, $exit_code, $ident, $exit_signal, $core_dump, $data_structure_reference) = @_;
    my $q = $data_structure_reference->{input};
    $results{$q} = $data_structure_reference->{result};
});

Then comes the main part of the code.

We have a simple foreach loop iterating over the @numbers array. For each iteration we call my $pid = $pm->start and next; This will try to create a new fork. If it is successful then at this point two processes will continue to run almost exactly the same way: the value returned by the start method is assigned to $pid. There is however a small difference in the two processes.

In the parent process, this value is going to be the process ID of the child process, a non-zero number, and therefore the right-hand side of the and boolean operator will be evaluated and the main process will go to the next iteration of the foreach loop.

In the child process the value returned by start will be 0. Which is false. Which means the right-hand side of the and operator will not be executed. In the child process the next evaluated statement will be the calc($q);. While the child process is calculating using one of the CPUs of the computer, the main process can run using the other CPU and it can create more child-processes.

The Parallel::Forkmanager will also count how many child processes have been forked and if we reach the value passed to the new constructor then the start command will wait till one of the earlier child-processes finishes and will only fork a new child-process after that.

In the meantime all the child processes are running on one of the CPUs of the computer. When one of them finishes the calc function it will call the finish method of Parallel::Forkmanager and it will pass to it two values. The first one is the exit code it wishes to have. 0 means success. The second one is a reference to a data structure it wishes to send back to the main process. This is the data structure we have used in the anonymous subroutine in $data_structure_reference.

foreach my $q (@numbers) {
    my $pid = $pm->start and next;
    my $res = calc($q);
    $pm->finish(0, { result => $res, input => $q });
}
$pm->wait_all_children;

The call to wait_all_children makes sure that the parent process will indeed wait till all the processes it created have finished and will only continue running once that happened and once it had a chance to run the respective run_on_finish function.

We can run this script as

time perl ~/work/perlmaven.com/examples/calc_fork_manager.pl 8

The result is about twice as fast as the linear version:

real    0m11.138s

At the same time htop shows that all the CPUs are saturated.

In some other measurements I’ve seen a 3-time speedup, but I think you can’t expect anything better with a 4-core machine. After all there are many other tasks running in the system, so we don’t really have 4 times more free CPU power than earlier, and the whole forking and managing the communication has some overhead.

Combined example

Just in case you’d like to tweak the calc() function and would like to further experiment with this code, I’ve included a version of the two scripts combined together. If you run it without any parameter it will run the linear version. If you run it with any positive number, it will use that many parallel child processes.

examples/calc_fork_manager_full.pl

use strict;
use warnings;
use Parallel::ForkManager;
use Data::Dumper qw(Dumper);
 
 
my $forks = shift;
 
my @numbers = map { $_ * 2000000 } reverse 1 .. 10;
my %results;
 
if ($forks) {
    print "Forking up to $forks at a time\n";
    my $pm = Parallel::ForkManager->new($forks);
    $pm->run_on_finish( sub {
        my ($pid, $exit_code, $ident, $exit_signal, $core_dump, $data_structure_reference) = @_;
#die Dumper \@_;
        my $q = $data_structure_reference->{input};
        $results{$q} = $data_structure_reference->{result};
    });
 
    DATA_LOOP:
    foreach my $q (@numbers) {
        my $pid = $pm->start and next DATA_LOOP;
        my $res = calc($q);
        $pm->finish(0, { result => $res, input => $q });
    }
    $pm->wait_all_children;
} else {
    print "Non-forking\n";
    foreach my $q (@numbers) {
        $results{$q} = calc($q);
    }
}
 
print Dumper \%results;
 
sub calc {
    my ($n) = @_;
    my $sum = 0;
    for (1 .. $n) {
        $sum += 1 + rand()/100;
    }
    return $sum;
}

BTW, the “DATA_LOOP” in this example is not really needed, it only tries to make the code a bit more readable.