MCE - Many-Core Engine for Perl. Provides parallel processing capabilities.
This document describes MCE version 1.415
Many-core Engine (MCE) for Perl helps enable a new level of performance by maximizing all available cores. MCE spawns a pool of workers and therefore does not fork a new process per each element of data. Instead, MCE follows a bank queuing model. Imagine the line being the data and bank-tellers the parallel workers. MCE enhances that model by adding the ability to chunk the next n elements from the input stream to the next available worker.
Both chunking and input are optional in MCE. One can simply use MCE to have many workers run in parallel.
This is a simplistic use case of MCE running with 4 workers.
use MCE; my $mce = MCE->new( max_workers => 4, user_func => sub { my ($self) = @_; print "Hello from ", $self->wid, "\n"; } ); $mce->run; -- Output Hello from 3 Hello from 1 Hello from 2 Hello from 4
The anonymous user_func above can be written using one less line. There are 3 constants defined in MCE which are exportable. Using the constants in lieu of 0,1,2 makes it more legible when accessing the array elements.
## The following exports SELF => 0, CHUNK => 1, CID => 2 use MCE EXPORT_CONST => 1; use MCE CONST => 1; ## Allowed in 1.415 and later user_func => sub { # my ($self, $chunk_ref, $chunk_id) = @_; print "Hello from ", $_[SELF]->wid, "\n"; }
In addition to EXPORT_CONST, five additional arguments can be passed to MCE. This allows one to override the default configured in MCE.pm.
use Sereal qw(encode_sereal decode_sereal); use MCE MAX_WORKERS => 4, ## Default 1 CHUNK_SIZE => 100, ## Default 1 TMP_DIR => "/path/to/app/tmp", ## $MCE::Signal::tmp_dir FREEZE => \&encode_sereal, ## \&Storable::freeze THAW => \&decode_sereal; ## \&Storable::thaw my $mce = MCE->new( ... );
Below, a new instance is configured with all available options.
use MCE; my $mce = MCE->new( max_workers => 8, ## Default $MCE::MAX_WORKERS ## Number of workers to spawn. This can be set automatically ## with MCE 1.412 and later releases. ## max_workers => 'auto', ## = MCE::Util::get_ncpu() ## max_workers => 'Auto-1', ## = MCE::Util::get_ncpu() - 1 ## max_workers => 'AUTO + 3', ## = MCE::Util::get_ncpu() + 3 ## max_workers => 'AUTO * 1.5', ## max_workers => 'auto / 1.333 + 2', chunk_size => 2000, ## Default $MCE::CHUNK_SIZE ## Less than or equal to 8192 is number of records. ## Greater than 8192 is number of bytes. MCE reads ## till the end of record before calling user_func. ## chunk_size => 1, ## Consists of 1 record ## chunk_size => 1000, ## Consists of 1000 records ## chunk_size => 16384, ## Approximate 16384 bytes ## chunk_size => 50000, ## Approximate 50000 bytes ## Chunk_size is set to 1 for $self->foreach and ## the $self->forseq methods. tmp_dir => $tmp_dir, ## Default $MCE::TMP_DIR ## Default is $MCE::Signal::tmp_dir which points to ## $ENV{TEMP} if defined. Otherwise, tmp_dir points ## to a location under /tmp. freeze => \&encode_sereal, ## Default $MCE::FREEZE thaw => \&decode_sereal, ## Default $MCE::THAW ## Release 1.412 allows freeze and thaw to be overridden. ## Just include a serialization module prior to loading MCE. ## use Sereal qw(encode_sereal decode_sereal); ## use JSON::XS qw(encode_json decode_json); ## use MCE; input_data => $input_file, ## Default undef RS => "\n::\n", ## Default undef ## input_data => '/path/to/file' for input file ## input_data => \@array for input array ## input_data => \*FILE_HNDL for file handle ## input_data => \$scalar to treat like a file ## Use the sequence option if simply wanting to loop ## through a sequence of numbers instead. ## Release 1.4 and later allows one to specify the ## record separator (RS) applicable to input files, ## file handles, and scalar references. use_slurpio => 1, ## Default 0 use_threads => 1, ## Default 0 or 1 ## Whether or not to enable slurpio when reading files ## (passes raw chunk to user function). ## By default MCE does forking (spawns child processes). ## MCE also supports threads via 2 threading libraries. ## ## The use of threads in MCE requires that you include ## threads support prior to loading MCE. The use_threads ## option defaults to 1 when a thread library is present. ## ## use threads; use forks; ## use threads::shared; (or) use forks::shared; ## ## use MCE use MCE; spawn_delay => 0.035, ## Default undef submit_delay => 0.002, ## Default undef job_delay => 0.150, ## Default undef ## Time to wait, in fractional seconds, before spawning ## each worker, parameters submission to each worker, ## and job commencement (staggered) for each worker. ## For example, use job_delay when wanting to stagger ## many workers connecting to a database. on_post_exit => \&on_post_exit, ## Default undef on_post_run => \&on_post_run, ## Default undef ## Execute code block immediately after a worker exits ## (exit, $self->exit, or die). Execute code block after ## running a job ($self->process or $self->run). ## One can take action immediately after a worker exits ## or wait until after the job has completed. user_args => { env => 'test' }, ## Default undef ## MCE release 1.4 adds a new parameter to allow one to ## specify arbitrary arguments such as a string, an ARRAY ## or a HASH reference. Workers can access this directly: ## my $args = $self->{user_args}; user_begin => \&user_begin, ## Default undef user_func => \&user_func, ## Default undef user_end => \&user_end, ## Default undef ## Think of user_begin, user_func, user_end like the awk ## scripting language: ## awk 'BEGIN { ... } { ... } END { ... }' ## MCE workers calls user_begin once per job, then ## calls user_func repeatedly until no chunks remain. ## Afterwards, user_end is called. user_error => \&user_error, ## Default undef user_output => \&user_output, ## Default undef ## When workers call the following functions, MCE will ## pass the data to user_error/user_output if defined. ## $self->sendto('stderr', 'Sending to STDERR'); ## $self->sendto('stdout', 'Sending to STDOUT'); stderr_file => 'err_file', ## Default STDERR stdout_file => 'out_file', ## Default STDOUT ## Or to file. User_error/user_output take precedence. flush_file => 1, ## Default 0 flush_stderr => 1, ## Default 0 flush_stdout => 1, ## Default 0 ## Flush sendto file, standard error, or standard output. sequence => { ## Default undef begin => -1, end => 1 [, step => 0.1 [, format => "4.1f" ] ] }, ## For looping through a sequence of numbers in parallel. ## STEP, if omitted, defaults to 1 if BEGIN is smaller than ## END or -1 if BEGIN is greater than END. The FORMAT string ## (without the %) is passed to sprintf behind the scene. ## e.g. $seq_n_formated = sprintf("%4.1f", $seq_n); ## Leave input_data set to undef when specifying the sequence ## option. One cannot specify both options together. ## Release 1.4 allows one to specify an array reference ## instead (3rd & 4th values are optional): ## ## sequence => [ -1, 1, 0.1, "4.1f" ] user_tasks => [ ## Default undef { ... }, ## Options for task 0 { ... }, ## Options for task 1 { ... }, ## Options for task 2 ], ## Takes a list of hash references, each allowing for 8 ## options: max_workers, chunk_size, sequence, user_args, ## user_begin, user_func, user_end, & task_end. The other ## options are ignored. Input_data, if specified from the ## main MCE instance, is only applicable to task 0. ## Options not specified here (excluding task_end) will ## default to same option specified from the main MCE ## instance. ## The task_end, if specified, is called after the task ## has completed processing. );
Run calls spawn, kicks off job, workers call user_begin, user_func, and user_end. Run shuts down workers afterwards. Call the spawn method early whenever the need arises for large data structures within the main process prior to running.
$mce->spawn(); ## This is optional $mce->run(); ## Call run or process below ## Acquire data arrays and/or input_files. The same pool of ## workers are used. $mce->process(\@input_data_1); ## Process arrays $mce->process(\@input_data_2); $mce->process(\@input_data_n); $mce->process('input_file_1'); ## Process files $mce->process('input_file_2'); $mce->process('input_file_n'); ## Shutdown workers afterwards. $mce->shutdown();
The 1.3 release and above allow workers to loop through a sequence of numbers in calculated mathematically without the overhead of an array. Sequence can be specified separately per each user_task entry unlike input_data. Input_data is applicable to the first task only.
See a demo at the end of this documentation applying sequences with the user_tasks option including chunking.
Sequence can be specified using an array or a hash reference.
use MCE; my $mce = MCE->new( max_workers => 3, # sequence => [ 10, 19, 0.7, "4.1f" ], sequence => { begin => 10, end => 19, step => 0.7, format => "4.1f" }, user_func => sub { my ($self, $n, $chunk_id) = @_; print $n, " from ", $self->wid(), " id ", $chunk_id, "\n"; } ); $mce->run(); -- Output (sorted afterwards, notice wid and chunk_id in output) 10.0 from 1 id 1 10.7 from 2 id 2 11.4 from 3 id 3 12.1 from 1 id 4 12.8 from 2 id 5 13.5 from 3 id 6 14.2 from 1 id 7 14.9 from 2 id 8 15.6 from 3 id 9 16.3 from 1 id 10 17.0 from 2 id 11 17.7 from 3 id 12 18.4 from 1 id 13
The 1.2 release introduced the user_tasks option to allow for multiple roles. This option takes an array of tasks. Each task allows for 9 sub options.
max_workers, use_threads, user_begin, user_func, user_end, user_args, chunk_size, sequence, task_end
Sequence and chunk_size were added in 1.3. User_args was introduced in 1.4. One can specify the task_end option for the manager process to run when all workers for that role have completed processing.
Forking and threading can be inter-mixed among tasks unless running Cygwin. Input data (if defined) can be processed by the first task only. The run method will remain running until all workers have completed processing.
Visit this URL for further reading on the new user_tasks option. http://code.google.com/p/many-core-engine-perl/wiki/MCE_Tasks
use MCE; my $mce = MCE->new( input_data => $list_file, user_tasks => [{ max_workers => 2, user_func => \¶llel_task1, use_threads => 0, task_end => sub { print "Task 1 completed processing\n"; } },{ max_workers => 4, user_func => \¶llel_task2, use_threads => 1 }] ); $mce->run();
Often times, one may want to capture the exit status. The on_post_exit option, if defined, is executed immediately after a worker exits via exit (children only), $self->exit (both children and threads), or die.
The format of $e->{pid} is PID_123 for children and THR_123 for threads.
sub on_post_exit { my ($self, $e) = @_; print "$e->{wid}: $e->{pid}: $e->{status}: $e->{msg}: $e->{id}\n"; } sub user_func { my $self = $_[0]; $self->exit(0, 'ok', 'pebbles'); } my $mce = MCE->new( on_post_exit => \&on_post_exit, user_func => \&user_func ); $mce->run(); -- Output 2: PID_7625: 0: ok: pebbles
The on_post_run option, if defined, is executed immediately after running $self->process or $self->run. This option receives an array reference of hashes.
The difference between on_post_exit and on_post_run is that the former is called immediately whereas the latter is called after all workers have completed processing or running.
sub on_post_run { my ($self, $status_ref) = @_; foreach my $e ( @{ $status_ref } ) { print "$e->{wid}: $e->{pid}: $e->{status}: $e->{msg}: $e->{id}\n"; } } sub user_func { my $self = $_[0]; $self->exit(0, 'ok', 'pebbles'); } my $mce = MCE->new( on_post_run => \&on_post_run, user_func => \&user_func ); $mce->run();
The user_begin and user_end options, if specified, behave similarly to awk 'BEGIN { ... } { ... } END { ... }'. These are called once per each worker during a run.
sub user_begin { ## Called once at the beginning my $self = shift; $self->{wk_total_rows} = 0; } sub user_func { ## Called while processing my $self = shift; $self->{wk_total_rows} += 1; } sub user_end { ## Called once at the end my $self = shift; printf "## %d: Processed %d rows\n", $self->wid(), $self->{wk_total_rows}; } my $mce = MCE->new( user_begin => \&user_begin, user_func => \&user_func, user_end => \&user_end ); $mce->run();
When processing input data, MCE can pass an array or rows or a slurped chunk. Below, a reference to an array containing the chunk data is processed.
e.g. $chunk_ref = [ record1, record2, record3, ... ]
sub user_func { my ($self, $chunk_ref, $chunk_id) = @_; foreach my $row ( @{ $chunk_ref } ) { $self->{wk_total_rows} += 1; print $row; } } my $mce = MCE->new( chunk_size => 100, input_data => "/path/to/file", user_func => \&user_func, use_slurpio => 0 ); $mce->run();
Here, a reference to a scalar containing the raw chunk data is processed.
sub user_func { my ($self, $chunk_ref, $chunk_id) = @_; my $count = () = $$chunk_ref =~ /abc/; } my $mce = MCE->new( chunk_size => 16000, input_data => "/path/to/file", user_func => \&user_func, use_slurpio => 1 ); $mce->run();
Output coming from $self->sendto('stderr/stdout', ...) can be intercepted by specifying the user_error and user_output options. MCE on receiving output will instead direct to user_error and user_output in a serialized fashion. Handy when wanting to filter, modify, and/or direct the output elsewhere.
sub user_error { ## Redirect STDERR to STDOUT my $error = shift; print STDOUT $error; } sub user_output { ## Redirect STDOUT to STDERR my $output = shift; print STDERR $output; } sub user_func { my ($self, $chunk_ref, $chunk_id) = @_; my $count = 0; foreach my $row ( @{ $chunk_ref } ) { $self->sendto('stdout', $row); $count += 1; } $self->sendto('stderr', "$chunk_id: processed $count rows\n"); } my $mce = MCE->new( chunk_size => 1000, input_data => "/path/to/file", user_error => \&user_error, user_output => \&user_output, user_func => \&user_func ); $mce->run();
The methods listed below are callable by the main process and workers.
The abort method is applicable when processing input_data only. This causes all workers to abort after processing the current chunk.
Workers write the next offset position to the queue socket for the next available worker. In essence, the abort method writes the last offset position. Workers, on requesting the next offset position, will think the end of input_data has been reached and leave the chunking loop.
$self->abort();
Returns the chunk_size used by MCE.
my $chunk_size = $self->chunk_size();
Returns the value for max_workers used by MCE.
my $max_workers = $self->max_workers();
Returns the session directory used by the MCE instance. This is defined during spawning and removed during shutdown.
my $sess_dir = $self->sess_dir();
Returns the task ID. This applies to the user_tasks option (starts at 0).
my $task_id = $self->task_id();
Returns the task worker ID (applies to user_tasks). The value starts at 1 per each task configured within user_tasks. The value is set to 0 for the manager process.
my $task_wid = $self->task_wid();
Returns the temporary directory used by MCE.
my $tmp_dir = $self->tmp_dir();
Returns the MCE worker ID. Starts at 1 per each MCE instance. The value is set to 0 for the manager process.
my $wid = $self->wid();
Methods listed below are callable by the main process only.
Forchunk, foreach, and forseq are sugar methods in MCE. Workers are auto- matically spawned, the code block is executed in parallel, and shutdown is called. Do not use these methods if wanting workers to persist afterwards.
Specifying options is optional. Valid options are the same as for the process method.
## Declare a MCE instance. my $mce = MCE->new( chunk_size => 20, max_workers => $max_workers ); ## Arguments inside the code block are the same as for user_func. $mce->forchunk(\@input_data, sub { my ($self, $chunk_ref, $chunk_id) = @_; foreach ( @{ $chunk_ref } ) { $self->sendto("stdout", "$chunk_id: $_\n"); } }); ## Passing chunk_size as an option. $mce->forchunk(\@input_data, { chunk_size => 30 }, sub { ... });
Foreach implies chunk_size => 1 and cannot be overwritten. Arguments inside the code block are the same as for user_func. This is true even if chunk_size is set to 1. MCE is both a chunking engine plus a parallel engine all in one. Arguments within the block are the same whether calling foreach or forchunk.
my $mce = MCE->new( max_workers => $max_workers ); $mce->foreach(\@input_data, sub { my ($self, $chunk_ref, $chunk_id) = @_; my $row = $chunk_ref->[0]; $self->sendto("stdout", "$chunk_id: $row\n"); });
Below, passing an anonymous array as input data. For example, wanting to parallelize a serial loop with MCE.
## Serial loops. for (my $i = 0; $i < $max; $i++) { ... ## Runs serially } for my $i (0 .. $max - 1) { ... ## Runs serially } ## Parallel loop via MCE. $mce->foreach([ (0 .. $max - 1) ], sub { my ($self, $chunk_ref, $chunk_id) = @_; my $i = $chunk_ref->[0]; ... ## Runs in parallel });
Sequence can be specified using a hash or an array reference.
my $mce = MCE->new( max_workers => 3 ); $mce->forseq({ begin => 15, end => 10, step => -1 }, sub { my ($self, $n, $chunk_id) = @_; print $n, " from ", $self->wid(), "\n"; }); $mce->forseq([ 20, 40 ], sub { my ($self, $n, $chunk_id) = @_; my $result = `ping 192.168.1.${n}`; ... });
MCE 1.415 allows chunk_size > 1. $n_seq is a reference pointing to an array of sequences below. Chunking reduces IPC overhead behind the scene. Chunk size is 1 when not specified.
$mce->forseq([ 20, 80 ], { chunk_size => 10 }, sub { my ($self, $n_seq, $chunk_id) = @_; for my $n ( @{ $n_seq } ) { my $result = `ping 192.168.1.${n}`; ... } });
The process method will spawn workers automatically if not already spawned. It will set input_data => $input_data. It calls run(0) to not auto-shutdown workers. Specifying options is optional.
Allowable options { key => value, ... } are:
chunk_size input_data job_delay spawn_delay submit_delay RS flush_file flush_stderr flush_stdout stderr_file stdout_file on_post_exit on_post_run sequence user_args user_begin user_end user_func user_error user_output use_slurpio
Options remain persistent going forward unless changed. Setting user_begin, user_end, or user_func will cause already spawned workers to shutdown and re-spawn automatically. Therefore, define these during instantiation if possible.
The below will cause workers to re-spawn after each process. It doesn't really matter for small jobs. One can create a few MCE instances. Perhaps the code block is different altogether.
$mce->new( max_workers => 'auto' ); $mce->process( { user_begin => sub { ## connect to DB }, user_func => sub { ## process each row }, user_end => sub { ## close handle to DB }, }, \@input_data ); $mce->process( { user_begin => sub { ## connect to DB }, user_func => sub { ## process each file }, user_end => sub { ## close handle to DB }, }, "/list/of/files" );
To have persistent workers, it's better to do the following instead. Workers will stay alive until told to shutdown or exiting of the script. Below is shown with overriding the default for max_workers.
use MCE MAX_WORKERS => 'auto'; $mce->new( user_begin => sub { ## connect to DB }, user_func => sub { ## process each chunk or row or host }, user_end => sub { ## close handle to DB }, ); $mce->spawn(); ## Not necessary, unless you want to spawn early $mce->process("/one/very_big_file/_mce_/will_chunk_in_parallel"); $mce->process(\@array_of_files_to_grep); $mce->process("/path/to/host/list"); $mce->process($array_ref); $mce->process($array_ref, { stdout_file => $output_file }); ## This was not allowed before. Fixed in 1.415. $mce->process({ sequence => { begin => 10, end => 90, step 2 } }); $mce->process({ sequence => [ 10, 90, 2 ] }); $mce->shutdown();
One can restart a worker who has died or exited. The job never ends below due to restarting each time. The same wid from the worker, which has exited, is used (recommended).
my $mce = MCE->new( on_post_exit => sub { my ($self, $e) = @_; print "$e->{wid}: $e->{pid}: status $e->{status}: $e->{msg}"; $self->restart_worker($e->{wid}); }, user_begin => sub { my $self = $_[0]; ## Not interested in die messages going to STDERR. ## The die handler calls $self->exit(255, $_[0]). close STDERR; }, user_tasks => [{ max_workers => 5, user_func => sub { my $self = $_[0]; sleep $self->wid(); $self->exit(3, "exited from " . $self->wid() . "\n"); } },{ max_workers => 4, user_func => sub { my $self = $_[0]; sleep $self->wid(); die("died from " . $self->wid() . "\n"); } }] ); $mce->run(); -- Output 1: PID_85388: status 3: exited from 1 2: PID_85389: status 3: exited from 2 1: PID_85397: status 3: exited from 1 3: PID_85390: status 3: exited from 3 1: PID_85399: status 3: exited from 1 4: PID_85391: status 3: exited from 4 2: PID_85398: status 3: exited from 2 1: PID_85401: status 3: exited from 1 5: PID_85392: status 3: exited from 5 1: PID_85404: status 3: exited from 1 6: PID_85393: status 255: died from 6 3: PID_85400: status 3: exited from 3 2: PID_85403: status 3: exited from 2 1: PID_85406: status 3: exited from 1 7: PID_85394: status 255: died from 7 1: PID_85410: status 3: exited from 1 8: PID_85395: status 255: died from 8 4: PID_85402: status 3: exited from 4 2: PID_85409: status 3: exited from 2 1: PID_85412: status 3: exited from 1 9: PID_85396: status 255: died from 9 3: PID_85408: status 3: exited from 3 1: PID_85416: status 3: exited from 1 ...
The run method, by default, spawns workers, processes once, and shuts down workers. Set $auto_shutdown to 0 when wanting workers to persist after running (default is 1).
my $mce = MCE->new( ... ); $mce->run(0); ## Disables auto-shutdown
The send method is useful when wanting to spawn workers early to minimize memory consumption and afterwards send data individually to each worker. One cannot send more than the total workers spawned. Workers store the received data as $self->{user_data}.
The data which can be sent is restricted to an ARRAY, HASH, or PDL reference. Workers begin processing immediately after receiving data. Workers set $self->{user_data} to undef after processing. One cannot specify input_data, sequence, or user_tasks for the MCE instance to receive user data via the "send" method.
Passing any options e.g. run(0, { options }) is ignored due to workers running immediately after receiving user data.
my $mce = MCE->new( max_workers => 5, user_func => sub { my ($self) = @_; my $data = $self->{user_data}; my $first_name = $data->{first_name}; print $self->wid, ": Hello from $first_name\n"; } ); $mce->spawn(); ## Optional, send will spawn if necessary. $mce->send( { first_name => "Theresa" } ); $mce->send( { first_name => "Francis" } ); $mce->send( { first_name => "Padre" } ); $mce->send( { first_name => "Anthony" } ); $mce->run(0); ## Wait for workers to complete processing. -- Output 2: Hello from Theresa 5: Hello from Anthony 3: Hello from Francis 4: Hello from Padre
The run method will automatically spawn workers, run once, and shutdown workers automatically. The process method leaves workers waiting for the next job after processing. Call shutdown after processing all jobs.
my $mce = MCE->new( ... ); $mce->spawn(); $mce->process(\@input_data_1); ## Processing multiple arrays $mce->process(\@input_data_2); $mce->process(\@input_data_n); $mce->process('input_file_1'); ## Processing multiple files $mce->process('input_file_2'); $mce->process('input_file_n'); $mce->shutdown();
Workers are normally spawned automatically. The spawn method is beneficial when wanting to spawn workers early.
my $mce = MCE->new( ... ); $mce->spawn();
The greatest exit status is saved among workers exiting during a run. Call this after running or processing to check if a worker exited with a non-zero value. For more granularity, look at the on_post_exit or on_post_run options for callback support.
my $mce = MCE->new( ... ); $mce->run(); my $greatest_exit_status = $mce->status();
Methods listed below are callable by workers only.
MCE can serialized data transfers from worker processes via helper functions do & sendto. The main MCE thread will process these in a serial fashion. This utilizes the Storable Perl module for passing data from a worker process to the main MCE thread. The callback function can optionally return a reply.
[ $reply = ] $self->do('callback' [, $arg1, ... ]);
Passing args to a callback function using references & scalar.
sub callback { my ($array_ref, $hash_ref, $scalar_ref, $scalar) = @_; ... } $self->do('main::callback', \@a, \%h, \$s, 'hello'); $self->do('callback', \@a, \%h, \$s, 'hello');
MCE knows if wanting a void, list, hash, or a scalar return value.
$self->do('callback' [, $arg1, ... ]); my @array = $self->do('callback' [, $arg1, ... ]); my %hash = $self->do('callback' [, $arg1, ... ]); my $scalar = $self->do('callback' [, $arg1, ... ]);
Worker exits entirely from MCE. $id (optional) can be used to pass the primary key or a string along with the message. Please look at the on_post_exit and on_post_run options for callback support.
$self->exit(); ## default is 0 $self->exit(1); $self->exit(2, 'error: entire chunk failed', $chunk_id); $self->exit(0, 'ok', 'pebbles');
Worker immediately leaves the chunking loop or user func. Called from inside foreach, forchunk, forseq, and user_func.
my @list = (1 .. 80); $mce->forchunk(\@list, { chunk_size => 2 }, sub { my ($self, $chunk_ref, $chunk_id) = @_; $self->last if ($chunk_id > 4); my @output = (); foreach my $rec ( @{ $chunk_ref } ) { push @output, $rec, "\n"; } $self->sendto('stdout', @output); }); -- Output (each chunk above consists of 2 elements) 1 2 3 4 5 6 7 8
Worker starts the next iteration of the chunking loop. Called from inside foreach, forchunk, forseq, and user_func.
my @list = (1 .. 80); $mce->forchunk(\@list, { chunk_size => 4 }, sub { my ($self, $chunk_ref, $chunk_id) = @_; $self->next if ($chunk_id < 20); my @output = (); foreach my $rec ( @{ $chunk_ref } ) { push @output, $rec, "\n"; } $self->sendto('stdout', @output); }); -- Output (each chunk above consists of 4 elements) 77 78 79 80
The sendto method is called by workers to serialize data to standard output, standard error, or to end of file. The action is done by the main process or thread.
Release 1.100 added the ability to pass multiple arguments.
Release 1.00x supported only 1 data argument. /path/to/file is the 3rd argument for 'file'.
$self->sendto('stdout', \@array); $self->sendto('stdout', \$scalar); $self->sendto('stdout', $scalar); $self->sendto('stderr', \@array); $self->sendto('stderr', \$scalar); $self->sendto('stderr', $scalar); $self->sendto('file', \@array, '/path/to/file'); $self->sendto('file', \$scalar, '/path/to/file'); $self->sendto('file', $scalar, '/path/to/file');
Notice the syntax change for appending to a file.
$self->sendto('stdout', $arg1 [, $arg2, ... ]); $self->sendto('stderr', $arg1 [, $arg2, ... ]); $self->sendto('file:/path/to/file', $arg1 [, $arg2, ... ]);
Passing a reference is no longer necessary beginning with 1.100.
$self->sendto("stdout", @a, "\n", %h, "\n", $s, "\n");
To retain 1.00x compatibility, sendto outputs the content when a single data argument is specified and is a reference.
$self->sendto('stdout', \@array); $self->sendto('stderr', \$scalar); $self->sendto('file:/path/to/file', \@array);
Otherwise, the reference for \@array and \$scalar is shown, not the content. Basically, output matches the print statement.
## print STDOUT "hello\n", \@array, \$scalar, "\n"; $self->sendto('stdout', "hello\n", \@array, \$scalar, "\n");
A barrier sync operation in the example code means any worker must stop at this point until all workers reach this barrier. Barrier syncing is useful for many computer algorithms.
Barrier synchronization is supported for task 0 only. Note: Workers are assigned task_id 0 when omitting user_tasks. All workers assigned task_id 0 must call sync whenever barrier syncing.
Barrier synchronization (sync) was added to MCE 1.406.
use MCE 1.406; sub user_func { my ($self) = @_; my $wid = $self->wid(); $self->sendto('stdout', "a: $wid\n"); $self->sync; $self->sendto('stdout', "b: $wid\n"); $self->sync; $self->sendto('stdout', "c: $wid\n"); $self->sync; } my $mce = MCE->new( max_workers => 4, user_func => \&user_func ); $mce->run(); -- Output (without barrier synchronization) a: 1 a: 2 b: 1 b: 2 c: 1 c: 2 a: 3 b: 3 c: 3 a: 4 b: 4 c: 4 -- Output (with barrier synchronization) a: 1 a: 2 a: 4 a: 3 b: 2 b: 1 b: 3 b: 4 c: 1 c: 4 c: 2 c: 3
Consider the following example. The sync operation is done inside a for loop along with calling the do method. A stall can occur for workers calling sync the 2nd or 3rd time while other workers are sending results via the do or sendto methods. It requires yet another semaphore lock file in MCE to solve this which was not done in order to keep resources low.
Therefore, please keep this in mind when mixing $self->sync with the $self->do or $self->sendto methods inside a loop.
sub user_func { my ($self) = @_; my @result; for (1 .. 3) { ... compute algorithm ... $self->sync; ... compute algorithm ... $self->sync; $self->do('aggregate_result', \@result); ## or $self->sendto $self->sync; ## The sync operation is also needed here to ## prevent MCE from stalling. } }
MCE comes with various examples showing real-world use case scenarios on parallelizing something as small as cat (try with -n) to searching for patterns and word count aggregation.
barrier_sync.pl A barrier sync demonstration. cat.pl Concatenation script, similar to the cat binary. egrep.pl Egrep script, similar to the egrep binary. wc.pl Word count script, similar to the wc binary. findnull.pl A parallel driven script to report lines containing null fields. It's many times faster than the binary egrep command. Try against a large file containing very long lines. foreach.pl, forseq.pl, forchunk.pl These take the same sqrt example from Parallel::Loops and measures the overhead of the engine. The number indicates the size of @input which can be submitted and results displayed in under 1 second. Parallel::Loops is based on Parallel::ForkManager. MCE utilizes a pool of workers. Parallel::Loops: 600 Forking each @input is expensive MCE foreach....: 21,500 Sends result after each @input MCE forseq.....: 64,000 Loops through sequence of numbers MCE forchunk...: 450,000 Chunking reduces overhead 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. Also included are 2 plain Perl examples. scaling_pings.pl Perform ping test and report back failing IPs to standard output. seq_demo.pl A demonstration of the new sequence option appearing in MCE 1.3. Run with seq_demo.pl | sort 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.
Imagine a long running process and wanting to parallelize an array against a pool of workers. Note: The sequence option can be used if simply wanting to loop through a sequence of numbers in parallel one number at a time.
Below, a callback function for displaying results is used. The logic shows how one can display results immediately while still preserving the output order as if processing serially. The %result hash is a temporary cache to store results for out-of-order replies.
my @input_data = (0 .. 18000 - 1); my $max_workers = 3; my $order_id = 1; my %result; ## Callback function for displaying results. sub display_result { my ($wk_result, $chunk_id) = @_; $result{$chunk_id} = $wk_result; while (1) { last unless (exists $result{$order_id}); printf "i: %d sqrt(i): %f\n", $input_data[$order_id - 1], $result{$order_id}; delete $result{$order_id}; $order_id++; } } ## Compute via MCE. my $mce = MCE->new( input_data => \@input_data, max_workers => $max_workers, chunk_size => 1, user_func => sub { my ($self, $chunk_ref, $chunk_id) = @_; my $wk_result = sqrt($chunk_ref->[0]); $self->do('display_result', $wk_result, $chunk_id); } ); $mce->run();
## Compute via MCE. Foreach implies chunk_size => 1. my $mce = MCE->new( max_workers => $max_workers ); ## Worker calls code block passing a reference to an array containing ## one item. Use $chunk_ref->[0] to retrieve the single element. $mce->foreach(\@input_data, sub { my ($self, $chunk_ref, $chunk_id) = @_; my $wk_result = sqrt($chunk_ref->[0]); $self->do('display_result', $wk_result, $chunk_id); });
Chunking reduces overhead many folds. Instead of passing a single item from @input_data, a chunk of $chunk_size is sent to the next available worker. The sequence option can be used as well if simply wanting to loop through a sequence of numbers with chunking applied in parallel.
my @input_data = (0 .. 385000 - 1); my $max_workers = 3; my $chunk_size = 500; my $order_id = 1; my %result; ## Callback function for displaying results. sub display_result { my ($wk_result, $chunk_id) = @_; $result{$chunk_id} = $wk_result; while (1) { last unless (exists $result{$order_id}); my $i = ($order_id - 1) * $chunk_size; foreach ( @{ $result{$order_id} } ) { printf "i: %d sqrt(i): %f\n", $input_data[$i++], $_; } delete $result{$order_id}; $order_id++; } } ## Compute via MCE. my $mce = MCE->new( input_data => \@input_data, max_workers => $max_workers, chunk_size => $chunk_size, user_func => sub { my ($self, $chunk_ref, $chunk_id) = @_; my @wk_result; foreach ( @{ $chunk_ref } ) { push @wk_result, sqrt($_); } $self->do('display_result', \@wk_result, $chunk_id); } ); $mce->run();
## Compute via MCE. my $mce = MCE->new( max_workers => $max_workers, chunk_size => $chunk_size ); ## Below, $chunk_ref is a reference to an array containing the next ## $chunk_size items from @input_data. $mce->forchunk(\@input_data, sub { my ($self, $chunk_ref, $chunk_id) = @_; my @wk_result; foreach ( @{ $chunk_ref } ) { push @wk_result, sqrt($_); } $self->do('display_result', \@wk_result, $chunk_id); });
One may specify the sequence option per each task. The following is taken directly from the seq_demo.pl example. Think of the following demonstration as having 3 mini-MCEs running simultaneously in parallel. Chunking can also be configured independently as well.
use MCE; ## Run with seq_demo.pl | sort sub user_func { my ($self, $seq_n, $chunk_id) = @_; my $wid = $self->wid(); my $task_id = $self->task_id(); my $task_wid = $self->task_wid(); if (ref $seq_n eq 'ARRAY') { ## Received the next "chunked" sequence of numbers ## e.g. when chunk_size > 1, $seq_n will be an array ref above foreach (@{ $seq_n }) { printf( "task_id %d: seq_n %s: chunk_id %d: wid %d: task_wid %d\n", $task_id, $_, $chunk_id, $wid, $task_wid ); } } else { 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 ); } } ## Each task can be configured independently. 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 }] ); $mce->run(); -- Output task_id 0: seq_n 11: chunk_id 1: wid 1: task_wid 1 task_id 0: seq_n 12: chunk_id 2: wid 2: task_wid 2 task_id 0: seq_n 13: chunk_id 3: wid 1: task_wid 1 task_id 0: seq_n 14: chunk_id 4: wid 2: task_wid 2 task_id 0: seq_n 15: chunk_id 5: wid 1: task_wid 1 task_id 0: seq_n 16: chunk_id 6: wid 2: task_wid 2 task_id 0: seq_n 17: chunk_id 7: wid 1: task_wid 1 task_id 0: seq_n 18: chunk_id 8: wid 2: task_wid 2 task_id 0: seq_n 19: chunk_id 9: wid 1: task_wid 1 task_id 1: seq_n 21: chunk_id 1: wid 3: task_wid 1 task_id 1: seq_n 22: chunk_id 1: wid 3: task_wid 1 task_id 1: seq_n 23: chunk_id 1: wid 3: task_wid 1 task_id 1: seq_n 24: chunk_id 1: wid 3: task_wid 1 task_id 1: seq_n 25: chunk_id 1: wid 3: task_wid 1 task_id 1: seq_n 26: chunk_id 2: wid 4: task_wid 2 task_id 1: seq_n 27: chunk_id 2: wid 4: task_wid 2 task_id 1: seq_n 28: chunk_id 2: wid 4: task_wid 2 task_id 1: seq_n 29: chunk_id 2: wid 4: task_wid 2 task_id 2: seq_n 31: chunk_id 1: wid 5: task_wid 1 task_id 2: seq_n 32: chunk_id 1: wid 5: task_wid 1 task_id 2: seq_n 33: chunk_id 1: wid 5: task_wid 1 task_id 2: seq_n 34: chunk_id 2: wid 6: task_wid 2 task_id 2: seq_n 35: chunk_id 2: wid 6: task_wid 2 task_id 2: seq_n 36: chunk_id 2: wid 6: task_wid 2 task_id 2: seq_n 37: chunk_id 3: wid 5: task_wid 1 task_id 2: seq_n 38: chunk_id 3: wid 5: task_wid 1 task_id 2: seq_n 39: chunk_id 3: wid 5: task_wid 1
Both input_data and sequence options are optional in MCE. One can simply use MCE to parallelize multiple workers. The "do" & "sendto" methods can be used to pass data back to the manager process. One doesn't have to wait until the worker has completed processing to pass data back. Both "do" & "sendto" methods are processed serially by the main process on a first come, first serve basis. All 4 workers run in parallel for the demonstration below.
use MCE; sub report_stats { my ($wid, $msg, $hash_ref) = @_; print "Worker $wid says $msg: ", $hash_ref->{'counter'}, "\n"; } my $mce = MCE->new( max_workers => 4, user_func => sub { my ($self) = @_; my $wid = $self->wid(); if ($wid == 1) { my %hash = ('counter' => 0); while (1) { $hash{'counter'} += 1; $self->do('report_stats', $wid, 'Hello there', \%hash); last if ($hash{'counter'} == 4); sleep 2; } } else { my %hash = ('counter' => 0); while (1) { $hash{'counter'} += 1; $self->do('report_stats', $wid, 'Welcome ...', \%hash); last if ($hash{'counter'} == 2); sleep 4; } } $self->sendto('stdout', "Worker $wid is exiting\n"); } ); $mce->run; Worker 2 gets there first in 2nd output below. $ ./demo.pl Worker 1 says Hello there: 1 Worker 2 says Welcome ...: 1 Worker 3 says Welcome ...: 1 Worker 4 says Welcome ...: 1 Worker 1 says Hello there: 2 Worker 2 says Welcome ...: 2 Worker 3 says Welcome ...: 2 Worker 1 says Hello there: 3 Worker 2 is exiting Worker 3 is exiting Worker 4 says Welcome ...: 2 Worker 4 is exiting Worker 1 says Hello there: 4 Worker 1 is exiting $ ./demo.pl Worker 2 says Welcome ...: 1 Worker 1 says Hello there: 1 Worker 4 says Welcome ...: 1 Worker 3 says Welcome ...: 1 Worker 1 says Hello 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 Hello there: 3 Worker 3 is exiting Worker 1 says Hello there: 4 Worker 1 is exiting
Perl 5.8.0 or later. The PDL::IO::Storable module is required in scripts parallelizing PDL with MCE.
MCE::Signal, MCE::Util
The source is hosted at http://code.google.com/p/many-core-engine-perl/
Mario E. Roy, <marioeroy AT gmail DOT com>
Copyright (C) 2012-2013 by Mario E. Roy
This program is free software; you can redistribute it and/or modify it under the terms of either: the GNU General Public License as published by the Free Software Foundation; or the Artistic License.
See http://dev.perl.org/licenses/ for more information.
To install MCE, copy and paste the appropriate command in to your terminal.
cpanm
cpanm MCE
CPAN shell
perl -MCPAN -e shell install MCE
For more information on module installation, please visit the detailed CPAN module installation guide.