|
NAMEMCE::Relay - Extends Many-Core Engine with relay capabilitiesVERSIONThis document describes MCE::Relay version 1.878SYNOPSISuse MCE::Flow; my $file = shift || \*STDIN; ## Line Count ####################################### mce_flow_f { max_workers => 4, use_slurpio => 1, init_relay => 0, }, sub { my ($mce, $slurp_ref, $chunk_id) = @_; my $line_count = ($$slurp_ref =~ tr/\n//); ## Receive and pass on updated information. my $lines_read = MCE::relay { $_ += $line_count }; }, $file; my $total_lines = MCE->relay_final; print {*STDERR} "$total_lines\n"; ## Orderly Action ################################### $| = 1; # Important, must flush output immediately. mce_flow_f { max_workers => 2, use_slurpio => 1, init_relay => 0, }, sub { my ($mce, $slurp_ref, $chunk_id) = @_; ## The relay value is relayed and remains 0. ## Writes to STDOUT orderly. MCE->relay_lock; print $$slurp_ref; MCE->relay_unlock; }, $file; DESCRIPTIONThis module enables workers to receive and pass on information orderly with zero involvement by the manager process while running. The module is loaded automatically when MCE option "init_relay" is specified.All workers (belonging to task_id 0) must participate when relaying data. Relaying is not meant for passing big data. The last worker will stall if exceeding the buffer size for the socket. Not exceeding 16 KiB - 7 is safe across all platforms. API DOCUMENTATION
Relay is enabled by specifying the init_relay option which takes a hash or array reference, or a scalar value. Relaying is orderly and driven by chunk_id when processing data, otherwise task_wid. Omitting the code block (e.g. MCE::relay) relays forward. Below, relaying multiple values via a HASH reference. use MCE::Flow max_workers => 4; mce_flow { init_relay => { p => 0, e => 0 }, }, sub { my $wid = MCE->wid; ## do work my $pass = $wid % 3; my $errs = $wid % 2; ## relay my %last_rpt = MCE::relay { $_->{p} += $pass; $_->{e} += $errs }; MCE->print("$wid: passed $pass, errors $errs\n"); return; }; my %results = MCE->relay_final; print " passed $results{p}, errors $results{e} final\n\n"; -- Output 1: passed 1, errors 1 2: passed 2, errors 0 3: passed 0, errors 1 4: passed 1, errors 0 passed 4, errors 2 final Or multiple values via an ARRAY reference. use MCE::Flow max_workers => 4; mce_flow { init_relay => [ 0, 0 ], }, sub { my $wid = MCE->wid; ## do work my $pass = $wid % 3; my $errs = $wid % 2; ## relay my @last_rpt = MCE::relay { $_->[0] += $pass; $_->[1] += $errs }; MCE->print("$wid: passed $pass, errors $errs\n"); return; }; my ($pass, $errs) = MCE->relay_final; print " passed $pass, errors $errs final\n\n"; -- Output 1: passed 1, errors 1 2: passed 2, errors 0 3: passed 0, errors 1 4: passed 1, errors 0 passed 4, errors 2 final Or simply a scalar value. use MCE::Flow max_workers => 4; mce_flow { init_relay => 0, }, sub { my $wid = MCE->wid; ## do work my $bytes_read = 1000 + ((MCE->wid % 3) * 3); ## relay my $last_offset = MCE::relay { $_ += $bytes_read }; ## output MCE->print("$wid: $bytes_read\n"); return; }; my $total = MCE->relay_final; print " $total size\n\n"; -- Output 1: 1003 2: 1006 3: 1000 4: 1003 4012 size
Call this method to obtain the final relay value(s) after running. See included example findnull.pl for another use case. use MCE max_workers => 4; my $mce = MCE->new( init_relay => [ 0, 100 ], ## initial values (two counters) user_func => sub { my ($mce) = @_; ## do work my ($acc1, $acc2) = (10, 20); ## relay to next worker MCE::relay { $_->[0] += $acc1; $_->[1] += $acc2 }; return; } )->run; my ($cnt1, $cnt2) = $mce->relay_final; print "$cnt1 : $cnt2\n"; -- Output 40 : 180
Call this method to obtain the next relay value before relaying. This allows serial-code to be processed orderly between workers. The following is a parallel demonstration for the fasta-benchmark on the web. # perl fasta.pl 25000000 # The Computer Language Benchmarks game # http://benchmarksgame.alioth.debian.org/ # # contributed by Barry Walsh # port of fasta.rb #6 # # MCE::Flow version by Mario Roy # requires MCE 1.807+ # requires MCE::Shared 1.806+ use strict; use warnings; use feature 'say'; use MCE::Flow; use MCE::Shared; use MCE::Candy; use constant IM => 139968; use constant IA => 3877; use constant IC => 29573; my $LAST = MCE::Shared->scalar( 42 ); my $alu = 'GGCCGGGCGCGGTGGCTCACGCCTGTAATCCCAGCACTTTGG' . 'GAGGCCGAGGCGGGCGGATCACCTGAGGTCAGGAGTTCGAGA' . 'CCAGCCTGGCCAACATGGTGAAACCCCGTCTCTACTAAAAAT' . 'ACAAAAATTAGCCGGGCGTGGTGGCGCGCGCCTGTAATCCCA' . 'GCTACTCGGGAGGCTGAGGCAGGAGAATCGCTTGAACCCGGG' . 'AGGCGGAGGTTGCAGTGAGCCGAGATCGCGCCACTGCACTCC' . 'AGCCTGGGCGACAGAGCGAGACTCCGTCTCAAAAA'; my $iub = [ [ 'a', 0.27 ], [ 'c', 0.12 ], [ 'g', 0.12 ], [ 't', 0.27 ], [ 'B', 0.02 ], [ 'D', 0.02 ], [ 'H', 0.02 ], [ 'K', 0.02 ], [ 'M', 0.02 ], [ 'N', 0.02 ], [ 'R', 0.02 ], [ 'S', 0.02 ], [ 'V', 0.02 ], [ 'W', 0.02 ], [ 'Y', 0.02 ] ]; my $homosapiens = [ [ 'a', 0.3029549426680 ], [ 'c', 0.1979883004921 ], [ 'g', 0.1975473066391 ], [ 't', 0.3015094502008 ] ]; sub make_repeat_fasta { my ( $src, $n ) = @_; my $width = qr/(.{1,60})/; my $l = length $src; my $s = $src x ( ($n / $l) + 1 ); substr( $s, $n, $l ) = ''; while ( $s =~ m/$width/g ) { say $1 } } sub make_random_fasta { my ( $table, $n ) = @_; my $rand = undef; my $width = 60; my $prob = 0.0; my $output = ''; my ( $c1, $c2, $last ); $_->[1] = ( $prob += $_->[1] ) for @$table; $c1 = '$rand = ( $last = ( $last * IA + IC ) % IM ) / IM;'; $c1 .= "\$output .= '$_->[0]', next if $_->[1] > \$rand;\n" for @$table; my $seq = MCE::Shared->sequence( { chunk_size => 2000, bounds_only => 1 }, 1, $n / $width ); my $code1 = q{ while ( 1 ) { # -------------------------------------------- # Process code orderly between workers. # -------------------------------------------- my $chunk_id = MCE->relay_recv; my ( $begin, $end ) = $seq->next; MCE->relay, last if ( !defined $begin ); my $last = $LAST->get; my $temp = $last; # Pre-compute $LAST value for the next worker for ( 1 .. ( $end - $begin + 1 ) * $width ) { $temp = ( $temp * IA + IC ) % IM; } $LAST->set( $temp ); # Increment chunk_id value MCE->relay( sub { $_ += 1 } ); # -------------------------------------------- # Also run code in parallel between workers. # -------------------------------------------- for ( $begin .. $end ) { for ( 1 .. $width ) { !C! } $output .= "\n"; } # -------------------------------------------- # Display orderly. # -------------------------------------------- MCE->gather( $chunk_id, $output ); $output = ''; } }; $code1 =~ s/!C!/$c1/g; MCE::Flow->init( max_workers => 4, ## MCE::Util->get_ncpu || 4, gather => MCE::Candy::out_iter_fh( \*STDOUT ), init_relay => 1, use_threads => 0, ); MCE::Flow->run( sub { eval $code1 } ); MCE::Flow->finish; $last = $LAST->get; $c2 = '$rand = ( $last = ( $last * IA + IC ) % IM ) / IM;'; $c2 .= "print('$_->[0]'), next if $_->[1] > \$rand;\n" for @$table; my $code2 = q{ if ( $n % $width != 0 ) { for ( 1 .. $n % $width ) { !C! } print "\n"; } }; $code2 =~ s/!C!/$c2/g; eval $code2; $LAST->set( $last ); } my $n = $ARGV[0] || 27; say ">ONE Homo sapiens alu"; make_repeat_fasta( $alu, $n * 2 ); say ">TWO IUB ambiguity codes"; make_random_fasta( $iub, $n * 3 ); say ">THREE Homo sapiens frequency"; make_random_fasta( $homosapiens, $n * 5 );
The "relay_lock" and "relay_unlock" methods, added to MCE 1.807, are aliases for "relay_recv" and "relay" respectively. Together, they allow one to perform an exclusive action prior to actual relaying of data. Relaying is driven by "chunk_id" or "task_wid" when not processing input, as seen here. MCE->new( max_workers => 8, init_relay => 0, user_func => sub { MCE->relay_lock; MCE->say("wid: ", MCE->task_wid); MCE->relay_unlock( sub { $_ += 2; }); } )->run; MCE->say("sum: ", MCE->relay_final); __END__ wid: 1 wid: 2 wid: 3 wid: 4 wid: 5 wid: 6 wid: 7 wid: 8 sum: 16 Described above, "relay" takes a code block and combines "relay_lock" and "relay_unlock" into a single call. To make this more interesting, I define "init_relay" to a hash containing two key-value pairs. MCE->new( max_workers => 8, init_relay => { count => 0, total => 0 }, user_func => sub { MCE->relay_lock; MCE->say("wid: ", MCE->task_wid); MCE->relay_unlock( sub { $_->{count} += 1; $_->{total} += 2; }); } )->run; my %results = MCE->relay_final; MCE->say("count: ", $results{count}); MCE->say("total: ", $results{total}); __END__ wid: 1 wid: 2 wid: 3 wid: 4 wid: 5 wid: 6 wid: 7 wid: 8 count: 8 total: 16 Below, "user_func" is taken from the "cat.pl" MCE example. Incrementing the count is done only when the "-n" switch is passed to the script. Otherwise, output is displaced orderly and not necessary to update the $_ value if exclusive locking is all you need. user_func => sub { my ($mce, $chunk_ref, $chunk_id) = @_; if ($n_flag) { ## Relays the total lines read. my $output = ''; my $line_count = ($$chunk_ref =~ tr/\n//); my $lines_read = MCE::relay { $_ += $line_count }; open my $fh, '<', $chunk_ref; $output .= sprintf "%6d\t%s", ++$lines_read, $_ while (<$fh>); close $fh; $output .= ":$chunk_id"; MCE->do('display_chunk', $output); } else { ## The following is another way to have ordered output. Workers ## write directly to STDOUT exclusively without any involvement ## from the manager process. The statement(s) between relay_lock ## and relay_unlock run serially and most important orderly. MCE->relay_lock; # alias for MCE->relay_recv print $$chunk_ref; # ensure $| = 1 in script MCE->relay_unlock; # alias for MCE->relay } return; } The following is a variant of the fasta-benchmark demonstration shown above. Here, workers write exclusively and orderly to "STDOUT". # perl fasta.pl 25000000 # The Computer Language Benchmarks game # http://benchmarksgame.alioth.debian.org/ # # contributed by Barry Walsh # port of fasta.rb #6 # # MCE::Flow version by Mario Roy # requires MCE 1.807+ # requires MCE::Shared 1.806+ use strict; use warnings; use feature 'say'; use MCE::Flow; use MCE::Shared; use constant IM => 139968; use constant IA => 3877; use constant IC => 29573; my $LAST = MCE::Shared->scalar( 42 ); my $alu = 'GGCCGGGCGCGGTGGCTCACGCCTGTAATCCCAGCACTTTGG' . 'GAGGCCGAGGCGGGCGGATCACCTGAGGTCAGGAGTTCGAGA' . 'CCAGCCTGGCCAACATGGTGAAACCCCGTCTCTACTAAAAAT' . 'ACAAAAATTAGCCGGGCGTGGTGGCGCGCGCCTGTAATCCCA' . 'GCTACTCGGGAGGCTGAGGCAGGAGAATCGCTTGAACCCGGG' . 'AGGCGGAGGTTGCAGTGAGCCGAGATCGCGCCACTGCACTCC' . 'AGCCTGGGCGACAGAGCGAGACTCCGTCTCAAAAA'; my $iub = [ [ 'a', 0.27 ], [ 'c', 0.12 ], [ 'g', 0.12 ], [ 't', 0.27 ], [ 'B', 0.02 ], [ 'D', 0.02 ], [ 'H', 0.02 ], [ 'K', 0.02 ], [ 'M', 0.02 ], [ 'N', 0.02 ], [ 'R', 0.02 ], [ 'S', 0.02 ], [ 'V', 0.02 ], [ 'W', 0.02 ], [ 'Y', 0.02 ] ]; my $homosapiens = [ [ 'a', 0.3029549426680 ], [ 'c', 0.1979883004921 ], [ 'g', 0.1975473066391 ], [ 't', 0.3015094502008 ] ]; sub make_repeat_fasta { my ( $src, $n ) = @_; my $width = qr/(.{1,60})/; my $l = length $src; my $s = $src x ( ($n / $l) + 1 ); substr( $s, $n, $l ) = ''; while ( $s =~ m/$width/g ) { say $1 } } sub make_random_fasta { my ( $table, $n ) = @_; my $rand = undef; my $width = 60; my $prob = 0.0; my $output = ''; my ( $c1, $c2, $last ); $_->[1] = ( $prob += $_->[1] ) for @$table; $c1 = '$rand = ( $last = ( $last * IA + IC ) % IM ) / IM;'; $c1 .= "\$output .= '$_->[0]', next if $_->[1] > \$rand;\n" for @$table; my $seq = MCE::Shared->sequence( { chunk_size => 2000, bounds_only => 1 }, 1, $n / $width ); my $code1 = q{ $| = 1; # Important, must flush output immediately. while ( 1 ) { # -------------------------------------------- # Process code orderly between workers. # -------------------------------------------- MCE->relay_lock; my ( $begin, $end ) = $seq->next; print( $output ), $output = '' if ( length $output ); MCE->relay_unlock, last if ( !defined $begin ); my $last = $LAST->get; my $temp = $last; # Pre-compute $LAST value for the next worker for ( 1 .. ( $end - $begin + 1 ) * $width ) { $temp = ( $temp * IA + IC ) % IM; } $LAST->set( $temp ); MCE->relay_unlock; # -------------------------------------------- # Also run code in parallel. # -------------------------------------------- for ( $begin .. $end ) { for ( 1 .. $width ) { !C! } $output .= "\n"; } } }; $code1 =~ s/!C!/$c1/g; MCE::Flow->init( max_workers => 4, ## MCE::Util->get_ncpu || 4, init_relay => 0, use_threads => 0, ); MCE::Flow->run( sub { eval $code1 } ); MCE::Flow->finish; $last = $LAST->get; $c2 = '$rand = ( $last = ( $last * IA + IC ) % IM ) / IM;'; $c2 .= "print('$_->[0]'), next if $_->[1] > \$rand;\n" for @$table; my $code2 = q{ if ( $n % $width != 0 ) { for ( 1 .. $n % $width ) { !C! } print "\n"; } }; $code2 =~ s/!C!/$c2/g; eval $code2; $LAST->set( $last ); } my $n = $ARGV[0] || 27; say ">ONE Homo sapiens alu"; make_repeat_fasta( $alu, $n * 2 ); say ">TWO IUB ambiguity codes"; make_random_fasta( $iub, $n * 3 ); say ">THREE Homo sapiens frequency"; make_random_fasta( $homosapiens, $n * 5 ); GATHER AND RELAY DEMONSTRATIONSI received a request from John Martel to process a large flat file and expand each record to many records based on splitting out items in field 4 delimited by semicolons. Each row in the output is given a unique ID starting with one while preserving output order.
This example configures a custom function for preserving output order. Unfortunately, the sprintf function alone involves extra CPU time causing the manager process to fall behind. Thus, workers may idle while waiting for the manager process to respond to the gather request. use strict; use warnings; use MCE::Loop; my $infile = shift or die "Usage: $0 infile\n"; my $newfile = 'output.dat'; open my $fh_out, '>', $newfile or die "open error $newfile: $!\n"; sub preserve_order { my ($fh) = @_; my ($order_id, $start_idx, $idx, %tmp) = (1, 1); return sub { my ($chunk_id, $aref) = @_; $tmp{ $chunk_id } = $aref; while ( my $aref = delete $tmp{ $order_id } ) { foreach my $line ( @{ $aref } ) { $idx = sprintf "%015d", $start_idx++; print $fh $idx, $line; } $order_id++; } } } MCE::Loop->init( chunk_size => 'auto', max_workers => 3, gather => preserve_order($fh_out) ); mce_loop_f { my ($mce, $chunk_ref, $chunk_id) = @_; my @buf; foreach my $line (@{ $chunk_ref }) { $line =~ s/\r//g; chomp $line; my ($f1,$f2,$f3,$items,$f5,$f6,$f7) = split /\|/, $line; my @items_array = split /;/, $items; foreach my $item (@items_array) { push @buf, "|$item|$f1|$f2|$f3|$f5|$f6|$f7\n"; } } MCE->gather($chunk_id, \@buf); } $infile; MCE::Loop->finish(); close $fh_out;
In this example, workers obtain the current ID value and increment/relay for the next worker, ordered by chunk ID behind the scene. Workers call sprintf in parallel, allowing the manager process (out_iter_fh) to accommodate up to 32 workers and not fall behind. Relay accounts for the worker handling the next chunk_id value. Therefore, do not call relay more than once per chunk. Doing so will cause IPC to stall. use strict; use warnings; use MCE::Loop; use MCE::Candy; my $infile = shift or die "Usage: $0 infile\n"; my $newfile = 'output.dat'; open my $fh_out, '>', $newfile or die "open error $newfile: $!\n"; MCE::Loop->init( chunk_size => 'auto', max_workers => 8, gather => MCE::Candy::out_iter_fh($fh_out), init_relay => 1 ); mce_loop_f { my ($mce, $chunk_ref, $chunk_id) = @_; my @lines; foreach my $line (@{ $chunk_ref }) { $line =~ s/\r//g; chomp $line; my ($f1,$f2,$f3,$items,$f5,$f6,$f7) = split /\|/, $line; my @items_array = split /;/, $items; foreach my $item (@items_array) { push @lines, "$item|$f1|$f2|$f3|$f5|$f6|$f7\n"; } } my $idx = MCE::relay { $_ += scalar @lines }; my $buf = ''; foreach my $line ( @lines ) { $buf .= sprintf "%015d|%s", $idx++, $line } MCE->gather($chunk_id, $buf); } $infile; MCE::Loop->finish(); close $fh_out; INDEXMCE, MCE::CoreAUTHORMario E. Roy, <marioeroy AT gmail DOT com>
Visit the GSP FreeBSD Man Page Interface. |