VERSION
This document describes MCE::Examples version 1.608DESCRIPTION
MCE comes with various examples showing real-world scenarios on parallelizing something as small as cat (try with -n) to searching for patterns and word count aggregation. MCE 1.522 adds sampledb to the list demonstrating DBI with MCE. MCE 1.600 adds biofasta (folder), mutex.pl, and relay.pl.INCLUDED WITH THE DISTRIBUTION
A wrapper script for parallelizing the grep binary. Hence, processing is done by the binary, not Perl. This wrapper resides under the bin directory.
mce_grep
A wrapper script with support for the following C binaries.
agrep, grep, egrep, fgrep, and tre-agrep
Chunking may be applied either at the [file] level, for large
file(s), or at the [list] level when parsing many files
recursively.
The gain in performance is noticeable for expensive patterns,
especially with agrep and tre-agrep.
The following scripts are located under the examples directory.
cat.pl, egrep.pl, wc.pl
Concatenation, egrep, and word count scripts similar to the
cat, egrep, and wc binaries respectively.
files_flow.pl, files_mce.pl, files_thr.pl
Demonstrates MCE::Flow, MCE::Queue, and Thread::Queue.
See MCE::Queue synopsis for another variation.
findnull.pl
A parallel script for reporting lines containing null fields.
It is many times faster than the egrep binary. Try this against
a large file containing very long lines.
flow_demo.pl, flow_model.pl
Demonstrates MCE::Flow, MCE::Queue, and MCE->gather.
foreach.pl, forseq.pl, forchunk.pl
These examples demonstrate the sqrt example from Parallel::Loops
(Parallel::Loops v0.07 utilizing Parallel::ForkManager v1.07).
Testing was on a Linux VM; Perl v5.20.1; Haswell i7 at 2.6 GHz.
The number indicates the size of input displayed in 1 second.
Output was directed to >/dev/null.
Parallel::Loops: 1,600 Forking each @input is expensive
MCE->foreach...: 23,000 Workers persist between each @input
MCE->forseq....: 200,000 Uses sequence of numbers as input
MCE->forchunk..: 800,000 IPC overhead is greatly reduced
interval.pl, mutex.pl, relay.pl
Demonstration of the interval option appearing in MCE 1.5.
Mutex locking and relaying data among workers.
iterator.pl
Similar to forseq.pl. Specifies an iterator for input_data.
A factory function is called which returns a closure.
pipe1.pl, pipe2.pl
Process STDIN or FILE in parallel. Processing is via Perl for
pipe1.pl, whereas an external command for pipe2.pl.
seq_demo.pl, step_demo.pl
Demonstration of the new sequence option appearing in MCE 1.3.
Run with seq_demo.pl | sort
Transparent use of MCE::Queue with MCE::Step.
sync.pl, utf8.pl
Barrier synchronization demonstration.
Process input containing unicode data.
The rest are organized into various sub directories.
biofasta/fasta_aidx.pl, fasta_rdr*.pl
Parallel demonstration for Bioinformatics.
matmult/matmult_base*.pl, matmult_mce*.pl, strassen_mce*.pl
Various matrix multiplication demonstrations benchmarking
PDL, PDL + MCE, as well as parallelizing Strassen's
divide-and-conquer algorithm. Included are 2 plain
Perl examples.
sampledb/create.pl, query*.pl, update*.pl
Examples demonstrating DBI (SQLite) with MCE.
tbray/wf_mce1.pl, wf_mce2.pl, wf_mce3.pl
An implementation of wide finder utilizing MCE.
As fast as MMAP IO when file resides in OS FS cache.
2x ~ 3x faster when reading directly from disk.
CHUNK_SIZE => 1 (in essence, wanting no chunking on input data)
Imagine a long running process and wanting to parallelize an array against a pool of workers. The sequence option may be used if simply wanting to loop through a sequence of numbers instead.Below, a callback function is used for displaying results. The logic shows how one can output results immediately while still preserving output order as if processing serially. The %tmp hash is a temporary cache for out-of-order results.
use MCE;
## Return an iterator for preserving output order.
sub preserve_order {
my (%result_n, %result_d); my $order_id = 1;
return sub {
my ($chunk_id, $n, $data) = @_;
$result_n{ $chunk_id } = $n;
$result_d{ $chunk_id } = $data;
while (1) {
last unless exists $result_d{$order_id};
printf "n: %5d sqrt(n): %7.3f\n",
$result_n{$order_id}, $result_d{$order_id};
delete $result_n{$order_id};
delete $result_d{$order_id};
$order_id++;
}
return;
};
}
## Use $chunk_ref->[0] or $_ to retrieve the element.
my @input_data = (0 .. 18000 - 1);
my $mce = MCE->new(
gather => preserve_order, input_data => \@input_data,
chunk_size => 1, max_workers => 3,
user_func => sub {
my ($mce, $chunk_ref, $chunk_id) = @_;
MCE->gather($chunk_id, $_, sqrt($_));
}
);
$mce->run;
This does the same thing using the foreach ``sugar'' method.
use MCE;
sub preserve_order {
...
}
my $mce = MCE->new(
chunk_size => 1, max_workers => 3,
gather => preserve_order
);
## Use $chunk_ref->[0] or $_ to retrieve the element.
my @input_data = (0 .. 18000 - 1);
$mce->foreach( \@input_data, sub {
my ($mce, $chunk_ref, $chunk_id) = @_;
MCE->gather($chunk_id, $_, sqrt($_));
});
The 2 examples described above were done using the Core API. MCE 1.5 comes with several models. The MCE::Loop model is used below.
use MCE::Loop;
sub preserve_order {
...
}
MCE::Loop::init {
chunk_size => 1, max_workers => 3,
gather => preserve_order
};
## Use $chunk_ref->[0] or $_ to retrieve the element.
my @input_data = (0 .. 18000 - 1);
mce_loop {
my ($mce, $chunk_ref, $chunk_id) = @_;
MCE->gather($chunk_id, $_, sqrt($_));
} @input_data;
MCE::Loop::finish;
CHUNKING INPUT_DATA
Chunking has the effect of reducing IPC overhead by many folds. A chunk containing $chunk_size items is sent to the next available worker.
use MCE;
## Return an iterator for preserving output order.
sub preserve_order {
my (%result_n, %result_d, $size); my $order_id = 1;
return sub {
my ($chunk_id, $n_ref, $data_ref) = @_;
$result_n{ $chunk_id } = $n_ref;
$result_d{ $chunk_id } = $data_ref;
while (1) {
last unless exists $result_d{$order_id};
$size = @{ $result_d{$order_id} };
for (0 .. $size - 1) {
printf "n: %5d sqrt(n): %7.3f\n",
$result_n{$order_id}->[$_], $result_d{$order_id}->[$_];
}
delete $result_n{$order_id};
delete $result_d{$order_id};
$order_id++;
}
return;
};
}
## Chunking requires one to loop inside the code block.
my @input_data = (0 .. 18000 - 1);
my $mce = MCE->new(
gather => preserve_order, input_data => \@input_data,
chunk_size => 500, max_workers => 3,
user_func => sub {
my ($mce, $chunk_ref, $chunk_id) = @_;
my (@n, @result);
foreach ( @{ $chunk_ref } ) {
push @n, $_;
push @result, sqrt($_);
}
MCE->gather($chunk_id, \@n, \@result);
}
);
$mce->run;
This does the same thing using the forchunk ``sugar'' method.
use MCE;
sub preserve_order {
...
}
my $mce = MCE->new(
chunk_size => 500, max_workers => 3,
gather => preserve_order
);
## Chunking requires one to loop inside the code block.
my @input_data = (0 .. 18000 - 1);
$mce->forchunk( \@input_data, sub {
my ($mce, $chunk_ref, $chunk_id) = @_;
my (@n, @result);
foreach ( @{ $chunk_ref } ) {
push @n, $_;
push @result, sqrt($_);
}
MCE->gather($chunk_id, \@n, \@result);
});
Finally, chunking with the MCE::Loop model.
use MCE::Loop;
sub preserve_order {
...
}
MCE::Loop::init {
chunk_size => 500, max_workers => 3,
gather => preserve_order
};
## Chunking requires one to loop inside the code block.
my @input_data = (0 .. 18000 - 1);
mce_loop {
my ($mce, $chunk_ref, $chunk_id) = @_;
my (@n, @result);
foreach ( @{ $chunk_ref } ) {
push @n, $_;
push @result, sqrt($_);
}
MCE->gather($chunk_id, \@n, \@result);
} @input_data;
MCE::Loop::finish;
DEMO APPLYING SEQUENCES WITH USER_TASKS
The following is an extract from the seq_demo.pl example included with MCE. Think of having several MCEs running in parallel. The sequence and chunk_size options may be specified uniquely per each task.The input scalar $_ (not shown below) contains the same value as $seq_n in user_func.
use MCE;
use Time::HiRes 'sleep';
## Run with seq_demo.pl | sort
sub user_func {
my ($mce, $seq_n, $chunk_id) = @_;
my $wid = MCE->wid;
my $task_id = MCE->task_id;
my $task_wid = MCE->task_wid;
if (ref $seq_n eq 'ARRAY') {
## seq_n or $_ is an array reference when chunk_size > 1
foreach (@{ $seq_n }) {
MCE->printf(
"task_id %d: seq_n %s: chunk_id %d: wid %d: task_wid %d\n",
$task_id, $_, $chunk_id, $wid, $task_wid
);
}
}
else {
MCE->printf(
"task_id %d: seq_n %s: chunk_id %d: wid %d: task_wid %d\n",
$task_id, $seq_n, $chunk_id, $wid, $task_wid
);
}
sleep 0.003;
return;
}
## Each task can be configured uniquely.
my $mce = MCE->new(
user_tasks => [{
max_workers => 2,
chunk_size => 1,
sequence => { begin => 11, end => 19, step => 1 },
user_func => \&user_func
},{
max_workers => 2,
chunk_size => 5,
sequence => { begin => 21, end => 29, step => 1 },
user_func => \&user_func
},{
max_workers => 2,
chunk_size => 3,
sequence => { begin => 31, end => 39, step => 1 },
user_func => \&user_func
}]
);
$mce->run;
-- Output
task_id 0: seq_n 11: chunk_id 1: wid 2: task_wid 2
task_id 0: seq_n 12: chunk_id 2: wid 1: task_wid 1
task_id 0: seq_n 13: chunk_id 3: wid 2: task_wid 2
task_id 0: seq_n 14: chunk_id 4: wid 1: task_wid 1
task_id 0: seq_n 15: chunk_id 5: wid 2: task_wid 2
task_id 0: seq_n 16: chunk_id 6: wid 1: task_wid 1
task_id 0: seq_n 17: chunk_id 7: wid 2: task_wid 2
task_id 0: seq_n 18: chunk_id 8: wid 1: task_wid 1
task_id 0: seq_n 19: chunk_id 9: wid 2: task_wid 2
task_id 1: seq_n 21: chunk_id 1: wid 3: task_wid 1
task_id 1: seq_n 22: chunk_id 1: wid 3: task_wid 1
task_id 1: seq_n 23: chunk_id 1: wid 3: task_wid 1
task_id 1: seq_n 24: chunk_id 1: wid 3: task_wid 1
task_id 1: seq_n 25: chunk_id 1: wid 3: task_wid 1
task_id 1: seq_n 26: chunk_id 2: wid 4: task_wid 2
task_id 1: seq_n 27: chunk_id 2: wid 4: task_wid 2
task_id 1: seq_n 28: chunk_id 2: wid 4: task_wid 2
task_id 1: seq_n 29: chunk_id 2: wid 4: task_wid 2
task_id 2: seq_n 31: chunk_id 1: wid 5: task_wid 1
task_id 2: seq_n 32: chunk_id 1: wid 5: task_wid 1
task_id 2: seq_n 33: chunk_id 1: wid 5: task_wid 1
task_id 2: seq_n 34: chunk_id 2: wid 6: task_wid 2
task_id 2: seq_n 35: chunk_id 2: wid 6: task_wid 2
task_id 2: seq_n 36: chunk_id 2: wid 6: task_wid 2
task_id 2: seq_n 37: chunk_id 3: wid 5: task_wid 1
task_id 2: seq_n 38: chunk_id 3: wid 5: task_wid 1
task_id 2: seq_n 39: chunk_id 3: wid 5: task_wid 1
GLOBALLY SCOPED VARIABLES AND MCE MODELS
It is possible that Perl may create a new code ref on subsequent runs causing MCE models to re-spawn. One solution to this is to declare global variables, referenced by workers, with ``our'' instead of ``my''.Let's take a look. The $i variable is declared with my and being reference in both user_begin and mce_loop blocks. This will cause Perl to create a new code ref for mce_loop on subsequent runs.
use MCE::Loop;
my $i = 0; ## <-- this is the reason, try our instead
MCE::Loop::init {
user_begin => sub {
print "process_id: $$\n" if MCE->wid == 1;
$i++;
},
chunk_size => 1, max_workers => 'auto',
};
for (1..2) {
## Perl creates another code block ref causing workers
## to re-spawn on subsequent runs.
print "\n"; mce_loop { print "$i: $_\n" } 1..4;
}
MCE::Loop::finish;
-- Output
process_id: 51380
1: 1
1: 2
1: 3
1: 4
process_id: 51388
1: 1
1: 2
1: 3
1: 4
By making the one line change, we see that workers persist for the duration of the script.
use MCE::Loop;
our $i = 0; ## <-- changed my to our
MCE::Loop::init {
user_begin => sub {
print "process_id: $$\n" if MCE->wid == 1;
$i++;
},
chunk_size => 1, max_workers => 'auto',
};
for (1..2) {
## Workers persist between runs. No re-spawning.
print "\n"; mce_loop { print "$i: $_\n" } 1..4;
}
-- Output
process_id: 51457
1: 1
1: 2
1: 4
1: 3
process_id: 51457
2: 1
2: 2
2: 3
2: 4
One may alternatively specify a code reference to existing routines for user_begin and mce_loop. Take notice of the comma after \&_func though.
use MCE::Loop;
my $i = 0; ## my (ok)
sub _begin {
print "process_id: $$\n" if MCE->wid == 1;
$i++;
}
sub _func {
print "$i: $_\n";
}
MCE::Loop::init {
user_begin => \&_begin,
chunk_size => 1, max_workers => 'auto',
};
for (1..2) {
print "\n"; mce_loop \&_func, 1..4;
}
MCE::Loop::finish;
-- Output
process_id: 51626
1: 1
1: 2
1: 3
1: 4
process_id: 51626
2: 1
2: 2
2: 3
2: 4
MONTE CARLO SIMULATION
There is an article on the web (search for comp.lang.perl.misc MCE) suggesting that MCE::Examples does not cover a simple simulation scenario. This section demonstrates just that.The serial code is based off the one by ``gamo''. A sleep is added to imitate extra CPU time. The while loop is wrapped within a for loop to run 10 times. The random number generator is seeded as well.
use Time::HiRes qw/sleep time/;
srand 5906;
my ($var, $foo, $bar) = (1, 2, 3);
my ($r, $a, $b);
my $start = time;
for (1..10) {
while (1) {
$r = rand;
$a = $r * ($var + $foo + $bar);
$b = sqrt($var + $foo + $bar);
last if ($a < $b + 0.001 && $a > $b - 0.001);
sleep 0.002;
}
print "$r -> $a\n";
}
my $end = time;
printf {*STDERR} "\n## compute time: %0.03f secs\n\n", $end - $start;
-- Output
0.408246276657106 -> 2.44947765994264
0.408099657137821 -> 2.44859794282693
0.408285842931324 -> 2.44971505758794
0.408342292008765 -> 2.45005375205259
0.408333076522673 -> 2.44999845913604
0.408344266898869 -> 2.45006560139321
0.408084104120526 -> 2.44850462472316
0.408197400014714 -> 2.44918440008828
0.408344783704855 -> 2.45006870222913
0.408248062985479 -> 2.44948837791287
## compute time: 93.049 secs
Next, we'd do the same with MCE. The demonstration requires at least MCE 1.509 to run properly. Folks on prior releases (1.505 - 1.508) will not see output for the 2nd run and beyond.
use Time::HiRes qw/sleep time/;
use MCE::Loop;
srand 5906;
## Configure MCE. Move common variables inside the user_begin
## block when not needed by the manager process.
MCE::Loop::init {
user_begin => sub {
use vars qw($var $foo $bar);
our ($var, $foo, $bar) = (1, 2, 3);
},
chunk_size => 1, max_workers => 'auto',
input_data => \&_input, gather => \&_gather
};
## Callback functions.
my ($done, $r, $a);
sub _input {
return if $done;
return rand;
}
sub _gather {
my ($_r, $_a, $_b) = @_;
return if $done;
if ($_a < $_b + 0.001 && $_a > $_b - 0.001) {
($done, $r, $a) = (1, $_r, $_a);
}
return;
}
## Compute in parallel.
my $start = time;
for (1..10) {
$done = 0; ## Reset $done before running
mce_loop {
# my ($mce, $chunk_ref, $chunk_id) = @_;
# my $r = $chunk_ref->[0];
my $r = $_; ## Valid due to chunk_size => 1
my $a = $r * ($var + $foo + $bar);
my $b = sqrt($var + $foo + $bar);
MCE->gather($r, $a, $b);
sleep 0.002;
};
print "$r -> $a\n";
}
printf "\n## compute time: %0.03f secs\n\n", time - $start;
-- Output
0.408246276657106 -> 2.44947765994264
0.408099657137821 -> 2.44859794282693
0.408285842931324 -> 2.44971505758794
0.408342292008765 -> 2.45005375205259
0.408333076522673 -> 2.44999845913604
0.408344266898869 -> 2.45006560139321
0.408084104120526 -> 2.44850462472316
0.408197400014714 -> 2.44918440008828
0.408344783704855 -> 2.45006870222913
0.408248062985479 -> 2.44948837791287
## compute time: 12.990 secs
Well, there you have it. MCE is able to complete the same simulation many times faster.
MANY WORKERS RUNNING IN PARALLEL
There are occasions when one wants several workers to run in parallel without having to specify input_data or seqeunce. These two options are optional in MCE. The ``do'' and ``sendto'' methods, for sending data to the manager process, are demonstrated below. Both process serially by the manager process on a first come, first serve basis.
use MCE::Flow max_workers => 4;
sub report_stats {
my ($wid, $msg, $h_ref) = @_;
print "Worker $wid says $msg: ", $h_ref->{"counter"}, "\n";
}
mce_flow sub {
my ($mce) = @_;
my $wid = MCE->wid;
if ($wid == 1) {
my %h = ("counter" => 0);
while (1) {
$h{"counter"} += 1;
MCE->do("report_stats", $wid, "Hey there", \%h);
last if ($h{"counter"} == 4);
sleep 2;
}
}
else {
my %h = ("counter" => 0);
while (1) {
$h{"counter"} += 1;
MCE->do("report_stats", $wid, "Welcome..", \%h);
last if ($h{"counter"} == 2);
sleep 4;
}
}
MCE->print(\*STDERR, "Worker $wid is exiting\n");
};
-- Output
Note how worker 2 comes first in the 2nd run below.
$ ./demo.pl
Worker 1 says Hey there: 1
Worker 2 says Welcome..: 1
Worker 3 says Welcome..: 1
Worker 4 says Welcome..: 1
Worker 1 says Hey there: 2
Worker 2 says Welcome..: 2
Worker 3 says Welcome..: 2
Worker 1 says Hey there: 3
Worker 2 is exiting
Worker 3 is exiting
Worker 4 says Welcome..: 2
Worker 4 is exiting
Worker 1 says Hey there: 4
Worker 1 is exiting
$ ./demo.pl
Worker 2 says Welcome..: 1
Worker 1 says Hey there: 1
Worker 4 says Welcome..: 1
Worker 3 says Welcome..: 1
Worker 1 says Hey there: 2
Worker 2 says Welcome..: 2
Worker 4 says Welcome..: 2
Worker 3 says Welcome..: 2
Worker 2 is exiting
Worker 4 is exiting
Worker 1 says Hey there: 3
Worker 3 is exiting
Worker 1 says Hey there: 4
Worker 1 is exiting
INDEX
MCEAUTHOR
Mario E. Roy, <marioeroy AT gmail DOT com>