MCE::Flow(3) User Contributed Perl Documentation MCE::Flow(3)
NAME
MCE::Flow - Parallel flow model for building creative applications
VERSION
This document describes MCE::Flow version 1.846
DESCRIPTION
MCE::Flow is great for writing custom apps to maximize on all available
cores. This module was created to help one harness user_tasks within
MCE.
It is trivial to parallelize with mce_stream shown below.
## Native map function
my @a = map { $_ * 4 } map { $_ * 3 } map { $_ * 2 } 1..10000;
## Same as with MCE::Stream (processing from right to left)
@a = mce_stream
sub { $_ * 4 }, sub { $_ * 3 }, sub { $_ * 2 }, 1..10000;
## Pass an array reference to have writes occur simultaneously
mce_stream \@a,
sub { $_ * 4 }, sub { $_ * 3 }, sub { $_ * 2 }, 1..10000;
However, let's have MCE::Flow compute the same in parallel. MCE::Queue
will be used for data flow among the sub-tasks.
use MCE::Flow;
use MCE::Queue;
This calls for preserving output order.
sub preserve_order {
my %tmp; my $order_id = 1; my $gather_ref = $_[0];
@{ $gather_ref } = (); ## clear the array (optional)
return sub {
my ($data_ref, $chunk_id) = @_;
$tmp{$chunk_id} = $data_ref;
while (1) {
last unless exists $tmp{$order_id};
push @{ $gather_ref }, @{ delete $tmp{$order_id++} };
}
return;
};
}
Two queues are needed for data flow between the 3 sub-tasks. Notice
task_end and how the value from $task_name is used for determining
which task has ended.
my $b = MCE::Queue->new;
my $c = MCE::Queue->new;
sub task_end {
my ($mce, $task_id, $task_name) = @_;
if (defined $mce->{user_tasks}->[$task_id + 1]) {
my $n_workers = $mce->{user_tasks}->[$task_id + 1]->{max_workers};
if ($task_name eq 'a') {
$b->enqueue((undef) x $n_workers);
}
elsif ($task_name eq 'b') {
$c->enqueue((undef) x $n_workers);
}
}
return;
}
Next are the 3 sub-tasks. The first one reads input and begins the
flow. The 2nd task dequeues, performs the calculation, and enqueues
into the next. Finally, the last task calls the gather method.
Although serialization is done for you automatically, it is done here
to save from double serialization. This is the fastest approach for
passing data between sub-tasks. Thus, the least overhead.
sub task_a {
my @ans; my ($mce, $chunk_ref, $chunk_id) = @_;
push @ans, map { $_ * 2 } @{ $chunk_ref };
$b->enqueue(MCE->freeze([ \@ans, $chunk_id ]));
return;
}
sub task_b {
my ($mce) = @_;
while (1) {
my @ans; my $chunk = $b->dequeue;
last unless defined $chunk;
$chunk = MCE->thaw($chunk);
push @ans, map { $_ * 3 } @{ $chunk->[0] };
$c->enqueue(MCE->freeze([ \@ans, $chunk->[1] ]));
}
return;
}
sub task_c {
my ($mce) = @_;
while (1) {
my @ans; my $chunk = $c->dequeue;
last unless defined $chunk;
$chunk = MCE->thaw($chunk);
push @ans, map { $_ * 4 } @{ $chunk->[0] };
MCE->gather(\@ans, $chunk->[1]);
}
return;
}
In summary, MCE::Flow builds out a MCE instance behind the scene and
starts running. The task_name (shown), max_workers, and use_threads
options can take an anonymous array for specifying the values uniquely
per each sub-task.
my @a;
mce_flow {
task_name => [ 'a', 'b', 'c' ], task_end => \&task_end,
gather => preserve_order(\@a)
}, \&task_a, \&task_b, \&task_c, 1..10000;
print "@a\n";
If speed is not a concern and wanting to rid of all the MCE->freeze and
MCE->thaw statements, simply enqueue and dequeue 2 items at a time. Or
better yet, see MCE::Step introduced in MCE 1.506.
First, task_end must be updated. The number of undef(s) must match the
number of workers times the dequeue count. Otherwise, the script will
stall.
sub task_end {
...
if ($task_name eq 'a') {
# $b->enqueue((undef) x $n_workers);
$b->enqueue((undef) x ($n_workers * 2));
}
elsif ($task_name eq 'b') {
# $c->enqueue((undef) x $n_workers);
$c->enqueue((undef) x ($n_workers * 2));
}
...
}
Next, the 3 sub-tasks enqueuing and dequeuing 2 elements at a time.
sub task_a {
my @ans; my ($mce, $chunk_ref, $chunk_id) = @_;
push @ans, map { $_ * 2 } @{ $chunk_ref };
$b->enqueue(\@ans, $chunk_id);
return;
}
sub task_b {
my ($mce) = @_;
while (1) {
my @ans; my ($chunk_ref, $chunk_id) = $b->dequeue(2);
last unless defined $chunk_ref;
push @ans, map { $_ * 3 } @{ $chunk_ref };
$c->enqueue(\@ans, $chunk_id);
}
return;
}
sub task_c {
my ($mce) = @_;
while (1) {
my @ans; my ($chunk_ref, $chunk_id) = $c->dequeue(2);
last unless defined $chunk_ref;
push @ans, map { $_ * 4 } @{ $chunk_ref };
MCE->gather(\@ans, $chunk_id);
}
return;
}
Finally, run as usual.
my @a;
mce_flow {
task_name => [ 'a', 'b', 'c' ], task_end => \&task_end,
gather => preserve_order(\@a)
}, \&task_a, \&task_b, \&task_c, 1..10000;
print "@a\n";
SYNOPSIS when CHUNK_SIZE EQUALS 1
Although MCE::Loop may be preferred for running using a single code
block, the text below also applies to this module, particularly for the
first block.
All models in MCE default to 'auto' for chunk_size. The arguments for
the block are the same as writing a user_func block using the Core API.
Beginning with MCE 1.5, the next input item is placed into the input
scalar variable $_ when chunk_size equals 1. Otherwise, $_ points to
$chunk_ref containing many items. Basically, line 2 below may be
omitted from your code when using $_. One can call MCE->chunk_id to
obtain the current chunk id.
line 1: user_func => sub {
line 2: my ($mce, $chunk_ref, $chunk_id) = @_;
line 3:
line 4: $_ points to $chunk_ref->[0]
line 5: in MCE 1.5 when chunk_size == 1
line 6:
line 7: $_ points to $chunk_ref
line 8: in MCE 1.5 when chunk_size > 1
line 9: }
Follow this synopsis when chunk_size equals one. Looping is not
required from inside the first block. Hence, the block is called once
per each item.
## Exports mce_flow, mce_flow_f, and mce_flow_s
use MCE::Flow;
MCE::Flow::init {
chunk_size => 1
};
## Array or array_ref
mce_flow sub { do_work($_) }, 1..10000;
mce_flow sub { do_work($_) }, \@list;
## Important; pass an array_ref for deeply input data
mce_flow sub { do_work($_) }, [ [ 0, 1 ], [ 0, 2 ], ... ];
mce_flow sub { do_work($_) }, \@deeply_list;
## File path, glob ref, IO::All::{ File, Pipe, STDIO } obj, or scalar ref
## Workers read directly and not involve the manager process
mce_flow_f sub { chomp; do_work($_) }, "/path/to/file"; # efficient
## Involves the manager process, therefore slower
mce_flow_f sub { chomp; do_work($_) }, $file_handle;
mce_flow_f sub { chomp; do_work($_) }, $io;
mce_flow_f sub { chomp; do_work($_) }, \$scalar;
## Sequence of numbers (begin, end [, step, format])
mce_flow_s sub { do_work($_) }, 1, 10000, 5;
mce_flow_s sub { do_work($_) }, [ 1, 10000, 5 ];
mce_flow_s sub { do_work($_) }, {
begin => 1, end => 10000, step => 5, format => undef
};
SYNOPSIS when CHUNK_SIZE is GREATER THAN 1
Follow this synopsis when chunk_size equals 'auto' or greater than 1.
This means having to loop through the chunk from inside the first
block.
use MCE::Flow;
MCE::Flow::init { ## Chunk_size defaults to 'auto' when
chunk_size => 'auto' ## not specified. Therefore, the init
}; ## function may be omitted.
## Syntax is shown for mce_flow for demonstration purposes.
## Looping inside the block is the same for mce_flow_f and
## mce_flow_s.
## Array or array_ref
mce_flow sub { do_work($_) for (@{ $_ }) }, 1..10000;
mce_flow sub { do_work($_) for (@{ $_ }) }, \@list;
## Important; pass an array_ref for deeply input data
mce_flow sub { do_work($_) for (@{ $_ }) }, [ [ 0, 1 ], [ 0, 2 ], ... ];
mce_flow sub { do_work($_) for (@{ $_ }) }, \@deeply_list;
## Resembles code using the core MCE API
mce_flow sub {
my ($mce, $chunk_ref, $chunk_id) = @_;
for (@{ $chunk_ref }) {
do_work($_);
}
}, 1..10000;
Chunking reduces the number of IPC calls behind the scene. Think in
terms of chunks whenever processing a large amount of data. For
relatively small data, choosing 1 for chunk_size is fine.
OVERRIDING DEFAULTS
The following list options which may be overridden when loading the
module.
use Sereal qw( encode_sereal decode_sereal );
use CBOR::XS qw( encode_cbor decode_cbor );
use JSON::XS qw( encode_json decode_json );
use MCE::Flow
max_workers => 8, # Default 'auto'
chunk_size => 500, # Default 'auto'
tmp_dir => "/path/to/app/tmp", # $MCE::Signal::tmp_dir
freeze => \&encode_sereal, # \&Storable::freeze
thaw => \&decode_sereal # \&Storable::thaw
;
From MCE 1.8 onwards, Sereal 3.015+ is loaded automatically if
available. Specify "Sereal => 0" to use Storable instead.
use MCE::Flow Sereal => 0;
CUSTOMIZING MCE
MCE::Flow->init ( options )
MCE::Flow::init { options }
The init function accepts a hash of MCE options. Unlike with
MCE::Stream, both gather and bounds_only options may be specified when
calling init (not shown below).
use MCE::Flow;
MCE::Flow::init {
chunk_size => 1, max_workers => 4,
user_begin => sub {
print "## ", MCE->wid, " started\n";
},
user_end => sub {
print "## ", MCE->wid, " completed\n";
}
};
my %a = mce_flow sub { MCE->gather($_, $_ * $_) }, 1..100;
print "\n", "@a{1..100}", "\n";
-- Output
## 3 started
## 2 started
## 4 started
## 1 started
## 2 completed
## 4 completed
## 3 completed
## 1 completed
1 4 9 16 25 36 49 64 81 100 121 144 169 196 225 256 289 324 361
400 441 484 529 576 625 676 729 784 841 900 961 1024 1089 1156
1225 1296 1369 1444 1521 1600 1681 1764 1849 1936 2025 2116 2209
2304 2401 2500 2601 2704 2809 2916 3025 3136 3249 3364 3481 3600
3721 3844 3969 4096 4225 4356 4489 4624 4761 4900 5041 5184 5329
5476 5625 5776 5929 6084 6241 6400 6561 6724 6889 7056 7225 7396
7569 7744 7921 8100 8281 8464 8649 8836 9025 9216 9409 9604 9801
10000
Like with MCE::Flow::init above, MCE options may be specified using an
anonymous hash for the first argument. Notice how task_name,
max_workers, and use_threads can take an anonymous array for setting
uniquely per each code block.
Unlike MCE::Stream which processes from right-to-left, MCE::Flow begins
with the first code block, thus processing from left-to-right.
use threads;
use MCE::Flow;
my @a = mce_flow {
task_name => [ 'a', 'b', 'c' ],
max_workers => [ 3, 4, 2, ],
use_threads => [ 1, 0, 0, ],
user_end => sub {
my ($mce, $task_id, $task_name) = @_;
MCE->print("$task_id - $task_name completed\n");
},
task_end => sub {
my ($mce, $task_id, $task_name) = @_;
MCE->print("$task_id - $task_name ended\n");
}
},
sub { sleep 1; }, ## 3 workers, named a
sub { sleep 2; }, ## 4 workers, named b
sub { sleep 3; }; ## 2 workers, named c
-- Output
0 - a completed
0 - a completed
0 - a completed
0 - a ended
1 - b completed
1 - b completed
1 - b completed
1 - b completed
1 - b ended
2 - c completed
2 - c completed
2 - c ended
API DOCUMENTATION
Although input data is optional for MCE::Flow, the following assumes
chunk_size equals 1 in order to demonstrate all the possibilities for
providing input data.
MCE::Flow->run ( sub { code }, list )
mce_flow sub { code }, list
Input data may be defined using a list, an array ref, or a hash ref.
Unlike MCE::Loop, Map, and Grep which take a block as "{ ... }", Flow
takes a "sub { ... }" or a code reference. The other difference is that
the comma is needed after the block.
# $_ contains the item when chunk_size => 1
mce_flow sub { do_work($_) }, 1..1000;
mce_flow sub { do_work($_) }, \@list;
# Important; pass an array_ref for deeply input data
mce_flow sub { do_work($_) }, [ [ 0, 1 ], [ 0, 2 ], ... ];
mce_flow sub { do_work($_) }, \@deeply_list;
# Chunking; any chunk_size => 1 or greater
my %res = mce_flow sub {
my ($mce, $chunk_ref, $chunk_id) = @_;
my %ret;
for my $item (@{ $chunk_ref }) {
$ret{$item} = $item * 2;
}
MCE->gather(%ret);
},
\@list;
# Input hash; current API available since 1.828
my %res = mce_flow sub {
my ($mce, $chunk_ref, $chunk_id) = @_;
my %ret;
for my $key (keys %{ $chunk_ref }) {
$ret{$key} = $chunk_ref->{$key} * 2;
}
MCE->gather(%ret);
},
\%hash;
# Unlike MCE::Loop, MCE::Flow doesn't need input to run
mce_flow { max_workers => 4 }, sub {
MCE->say( MCE->wid );
};
# ... and can run multiple tasks
mce_flow {
max_workers => [ 1, 3 ],
task_name => [ 'p', 'c' ]
},
sub {
# 1 producer
MCE->say( "producer: ", MCE->wid );
},
sub {
# 3 consumers
MCE->say( "consumer: ", MCE->wid );
};
# Here, options are specified via init
MCE::Flow::init {
max_workers => [ 1, 3 ],
task_name => [ 'p', 'c' ]
};
mce_flow \&producer, \&consumers;
MCE::Flow->run_file ( sub { code }, file )
mce_flow_f sub { code }, file
The fastest of these is the /path/to/file. Workers communicate the next
offset position among themselves with zero interaction by the manager
process.
"IO::All" { File, Pipe, STDIO } is supported since MCE 1.845.
# $_ contains the line when chunk_size => 1
mce_flow_f sub { $_ }, "/path/to/file"; # faster
mce_flow_f sub { $_ }, $file_handle;
mce_flow_f sub { $_ }, $io; # IO::All
mce_flow_f sub { $_ }, \$scalar;
# chunking, any chunk_size => 1 or greater
my %res = mce_flow_f sub {
my ($mce, $chunk_ref, $chunk_id) = @_;
my $buf = '';
for my $line (@{ $chunk_ref }) {
$buf .= $line;
}
MCE->gather($chunk_id, $buf);
},
"/path/to/file";
MCE::Flow->run_seq ( sub { code }, $beg, $end [, $step, $fmt ] )
mce_flow_s sub { code }, $beg, $end [, $step, $fmt ]
Sequence may be defined as a list, an array reference, or a hash
reference. The functions require both begin and end values to run.
Step and format are optional. The format is passed to sprintf (% may be
omitted below).
my ($beg, $end, $step, $fmt) = (10, 20, 0.1, "%4.1f");
# $_ contains the sequence number when chunk_size => 1
mce_flow_s sub { $_ }, $beg, $end, $step, $fmt;
mce_flow_s sub { $_ }, [ $beg, $end, $step, $fmt ];
mce_flow_s sub { $_ }, {
begin => $beg, end => $end,
step => $step, format => $fmt
};
# chunking, any chunk_size => 1 or greater
my %res = mce_flow_s sub {
my ($mce, $chunk_ref, $chunk_id) = @_;
my $buf = '';
for my $seq (@{ $chunk_ref }) {
$buf .= "$seq\n";
}
MCE->gather($chunk_id, $buf);
},
[ $beg, $end ];
The sequence engine can compute 'begin' and 'end' items only, for the
chunk, and not the items in between (hence boundaries only). This
option applies to sequence only and has no effect when chunk_size
equals 1.
The time to run is 0.006s below. This becomes 0.827s without the
bounds_only option due to computing all items in between, thus creating
a very large array. Basically, specify bounds_only => 1 when boundaries
is all you need for looping inside the block; e.g. Monte Carlo
simulations.
Time was measured using 1 worker to emphasize the difference.
use MCE::Flow;
MCE::Flow::init {
max_workers => 1, chunk_size => 1_250_000,
bounds_only => 1
};
# Typically, the input scalar $_ contains the sequence number
# when chunk_size => 1, unless the bounds_only option is set
# which is the case here. Thus, $_ points to $chunk_ref.
mce_flow_s sub {
my ($mce, $chunk_ref, $chunk_id) = @_;
# $chunk_ref contains 2 items, not 1_250_000
# my ( $begin, $end ) = ( $_->[0], $_->[1] );
my $begin = $chunk_ref->[0];
my $end = $chunk_ref->[1];
# for my $seq ( $begin .. $end ) {
# ...
# }
MCE->printf("%7d .. %8d\n", $begin, $end);
},
[ 1, 10_000_000 ];
-- Output
1 .. 1250000
1250001 .. 2500000
2500001 .. 3750000
3750001 .. 5000000
5000001 .. 6250000
6250001 .. 7500000
7500001 .. 8750000
8750001 .. 10000000
MCE::Flow->run ( { input_data => iterator }, sub { code } )
mce_flow { input_data => iterator }, sub { code }
An iterator reference may be specified for input_data. The only other
way is to specify input_data via MCE::Flow::init. This prevents
MCE::Flow from configuring the iterator reference as another user task
which will not work.
Iterators are described under section "SYNTAX for INPUT_DATA" at
MCE::Core.
MCE::Flow::init {
input_data => iterator
};
mce_flow sub { $_ };
GATHERING DATA
Unlike MCE::Map where gather and output order are done for you
automatically, the gather method is used to have results sent back to
the manager process.
use MCE::Flow chunk_size => 1;
## Output order is not guaranteed.
my @a1 = mce_flow sub { MCE->gather($_ * 2) }, 1..100;
print "@a1\n\n";
## Outputs to a hash instead (key, value).
my %h1 = mce_flow sub { MCE->gather($_, $_ * 2) }, 1..100;
print "@h1{1..100}\n\n";
## This does the same thing due to chunk_id starting at one.
my %h2 = mce_flow sub { MCE->gather(MCE->chunk_id, $_ * 2) }, 1..100;
print "@h2{1..100}\n\n";
The gather method may be called multiple times within the block unlike
return which would leave the block. Therefore, think of gather as
yielding results immediately to the manager process without actually
leaving the block.
use MCE::Flow chunk_size => 1, max_workers => 3;
my @hosts = qw(
hosta hostb hostc hostd hoste
);
my %h3 = mce_flow sub {
my ($output, $error, $status); my $host = $_;
## Do something with $host;
$output = "Worker ". MCE->wid .": Hello from $host";
if (MCE->chunk_id % 3 == 0) {
## Simulating an error condition
local $? = 1; $status = $?;
$error = "Error from $host"
}
else {
$status = 0;
}
## Ensure unique keys (key, value) when gathering to
## a hash.
MCE->gather("$host.out", $output);
MCE->gather("$host.err", $error) if (defined $error);
MCE->gather("$host.sta", $status);
}, @hosts;
foreach my $host (@hosts) {
print $h3{"$host.out"}, "\n";
print $h3{"$host.err"}, "\n" if (exists $h3{"$host.err"});
print "Exit status: ", $h3{"$host.sta"}, "\n\n";
}
-- Output
Worker 3: Hello from hosta
Exit status: 0
Worker 2: Hello from hostb
Exit status: 0
Worker 1: Hello from hostc
Error from hostc
Exit status: 1
Worker 3: Hello from hostd
Exit status: 0
Worker 2: Hello from hoste
Exit status: 0
The following uses an anonymous array containing 3 elements when
gathering data. Serialization is automatic behind the scene.
my %h3 = mce_flow sub {
...
MCE->gather($host, [$output, $error, $status]);
}, @hosts;
foreach my $host (@hosts) {
print $h3{$host}->[0], "\n";
print $h3{$host}->[1], "\n" if (defined $h3{$host}->[1]);
print "Exit status: ", $h3{$host}->[2], "\n\n";
}
Although MCE::Map comes to mind, one may want additional control when
gathering data such as retaining output order.
use MCE::Flow;
sub preserve_order {
my %tmp; my $order_id = 1; my $gather_ref = $_[0];
return sub {
$tmp{ (shift) } = \@_;
while (1) {
last unless exists $tmp{$order_id};
push @{ $gather_ref }, @{ delete $tmp{$order_id++} };
}
return;
};
}
## Workers persist for the most part after running. Though, not always
## the case and depends on Perl. Pass a reference to a subroutine if
## workers must persist; e.g. mce_flow { ... }, \&foo, 1..100000.
MCE::Flow::init {
chunk_size => 'auto', max_workers => 'auto'
};
for (1..2) {
my @m2;
mce_flow {
gather => preserve_order(\@m2)
},
sub {
my @a; my ($mce, $chunk_ref, $chunk_id) = @_;
## Compute the entire chunk data at once.
push @a, map { $_ * 2 } @{ $chunk_ref };
## Afterwards, invoke the gather feature, which
## will direct the data to the callback function.
MCE->gather(MCE->chunk_id, @a);
}, 1..100000;
print scalar @m2, "\n";
}
MCE::Flow::finish;
All 6 models support 'auto' for chunk_size unlike the Core API. Think
of the models as the basis for providing JIT for MCE. They create the
instance, tune max_workers, and tune chunk_size automatically
regardless of the hardware.
The following does the same thing using the Core API. Workers persist
after running.
use MCE;
sub preserve_order {
...
}
my $mce = MCE->new(
max_workers => 'auto', chunk_size => 8000,
user_func => sub {
my @a; my ($mce, $chunk_ref, $chunk_id) = @_;
## Compute the entire chunk data at once.
push @a, map { $_ * 2 } @{ $chunk_ref };
## Afterwards, invoke the gather feature, which
## will direct the data to the callback function.
MCE->gather(MCE->chunk_id, @a);
}
);
for (1..2) {
my @m2;
$mce->process({ gather => preserve_order(\@m2) }, [1..100000]);
print scalar @m2, "\n";
}
$mce->shutdown;
MANUAL SHUTDOWN
MCE::Flow->finish
MCE::Flow::finish
Workers remain persistent as much as possible after running. Shutdown
occurs automatically when the script terminates. Call finish when
workers are no longer needed.
use MCE::Flow;
MCE::Flow::init {
chunk_size => 20, max_workers => 'auto'
};
mce_flow sub { ... }, 1..100;
MCE::Flow::finish;
INDEX
MCE(3), MCE::Core(3)
AUTHOR
Mario E. Roy, <marioeroyA ATA gmailA DOTA com> perl v5.28.2 2019-08-27 MCE::Flow(3)
mce 1.846.0 - Generated Wed Aug 28 16:36:56 CDT 2019