… | |
… | |
70 | preserving most of the advantages of fork. |
70 | preserving most of the advantages of fork. |
71 | |
71 | |
72 | It can be used to create new worker processes or new independent |
72 | It can be used to create new worker processes or new independent |
73 | subprocesses for short- and long-running jobs, process pools (e.g. for use |
73 | subprocesses for short- and long-running jobs, process pools (e.g. for use |
74 | in pre-forked servers) but also to spawn new external processes (such as |
74 | in pre-forked servers) but also to spawn new external processes (such as |
75 | CGI scripts from a webserver), which can be faster (and more well behaved) |
75 | CGI scripts from a web server), which can be faster (and more well behaved) |
76 | than using fork+exec in big processes. |
76 | than using fork+exec in big processes. |
77 | |
77 | |
78 | Special care has been taken to make this module useful from other modules, |
78 | Special care has been taken to make this module useful from other modules, |
79 | while still supporting specialised environments such as L<App::Staticperl> |
79 | while still supporting specialised environments such as L<App::Staticperl> |
80 | or L<PAR::Packer>. |
80 | or L<PAR::Packer>. |
|
|
81 | |
|
|
82 | =head1 WHAT THIS MODULE IS NOT |
|
|
83 | |
|
|
84 | This module only creates processes and lets you pass file handles and |
|
|
85 | strings to it, and run perl code. It does not implement any kind of RPC - |
|
|
86 | there is no back channel from the process back to you, and there is no RPC |
|
|
87 | or message passing going on. |
|
|
88 | |
|
|
89 | If you need some form of RPC, you can either implement it yourself |
|
|
90 | in whatever way you like, use some message-passing module such |
|
|
91 | as L<AnyEvent::MP>, some pipe such as L<AnyEvent::ZeroMQ>, use |
|
|
92 | L<AnyEvent::Handle> on both sides to send e.g. JSON or Storable messages, |
|
|
93 | and so on. |
81 | |
94 | |
82 | =head1 PROBLEM STATEMENT |
95 | =head1 PROBLEM STATEMENT |
83 | |
96 | |
84 | There are two ways to implement parallel processing on UNIX like operating |
97 | There are two ways to implement parallel processing on UNIX like operating |
85 | systems - fork and process, and fork+exec and process. They have different |
98 | systems - fork and process, and fork+exec and process. They have different |
… | |
… | |
97 | or fork+exec instead. |
110 | or fork+exec instead. |
98 | |
111 | |
99 | =item Forking usually creates a copy-on-write copy of the parent |
112 | =item Forking usually creates a copy-on-write copy of the parent |
100 | process. Memory (for example, modules or data files that have been |
113 | process. Memory (for example, modules or data files that have been |
101 | will not take additional memory). When exec'ing a new process, modules |
114 | will not take additional memory). When exec'ing a new process, modules |
102 | and data files might need to be loaded again, at extra cpu and memory |
115 | and data files might need to be loaded again, at extra CPU and memory |
103 | cost. Likewise when forking, all data structures are copied as well - if |
116 | cost. Likewise when forking, all data structures are copied as well - if |
104 | the program frees them and replaces them by new data, the child processes |
117 | the program frees them and replaces them by new data, the child processes |
105 | will retain the memory even if it isn't used. |
118 | will retain the memory even if it isn't used. |
106 | |
119 | |
107 | This module allows the main program to do a controlled fork, and allows |
120 | This module allows the main program to do a controlled fork, and allows |
… | |
… | |
119 | as template, and also tries hard to identify the correct path to the perl |
132 | as template, and also tries hard to identify the correct path to the perl |
120 | interpreter. With a cooperative main program, exec'ing the interpreter |
133 | interpreter. With a cooperative main program, exec'ing the interpreter |
121 | might not even be necessary. |
134 | might not even be necessary. |
122 | |
135 | |
123 | =item Forking might be impossible when a program is running. For example, |
136 | =item Forking might be impossible when a program is running. For example, |
124 | POSIX makes it almost impossible to fork from a multithreaded program and |
137 | POSIX makes it almost impossible to fork from a multi-threaded program and |
125 | do anything useful in the child - strictly speaking, if your perl program |
138 | do anything useful in the child - strictly speaking, if your perl program |
126 | uses posix threads (even indirectly via e.g. L<IO::AIO> or L<threads>), |
139 | uses posix threads (even indirectly via e.g. L<IO::AIO> or L<threads>), |
127 | you cannot call fork on the perl level anymore, at all. |
140 | you cannot call fork on the perl level anymore, at all. |
128 | |
141 | |
129 | This module can safely fork helper processes at any time, by caling |
142 | This module can safely fork helper processes at any time, by calling |
130 | fork+exec in C, in a POSIX-compatible way. |
143 | fork+exec in C, in a POSIX-compatible way. |
131 | |
144 | |
132 | =item Parallel processing with fork might be inconvenient or difficult |
145 | =item Parallel processing with fork might be inconvenient or difficult |
133 | to implement. For example, when a program uses an event loop and creates |
146 | to implement. For example, when a program uses an event loop and creates |
134 | watchers it becomes very hard to use the event loop from a child |
147 | watchers it becomes very hard to use the event loop from a child |
… | |
… | |
166 | needed the first time. Forking from this process shares the memory used |
179 | needed the first time. Forking from this process shares the memory used |
167 | for the perl interpreter with the new process, but loading modules takes |
180 | for the perl interpreter with the new process, but loading modules takes |
168 | time, and the memory is not shared with anything else. |
181 | time, and the memory is not shared with anything else. |
169 | |
182 | |
170 | This is ideal for when you only need one extra process of a kind, with the |
183 | This is ideal for when you only need one extra process of a kind, with the |
171 | option of starting and stipping it on demand. |
184 | option of starting and stopping it on demand. |
172 | |
185 | |
173 | Example: |
186 | Example: |
174 | |
187 | |
175 | AnyEvent::Fork |
188 | AnyEvent::Fork |
176 | ->new |
189 | ->new |
… | |
… | |
191 | modules you loaded) is shared between the processes, and each new process |
204 | modules you loaded) is shared between the processes, and each new process |
192 | consumes relatively little memory of its own. |
205 | consumes relatively little memory of its own. |
193 | |
206 | |
194 | The disadvantage of this approach is that you need to create a template |
207 | The disadvantage of this approach is that you need to create a template |
195 | process for the sole purpose of forking new processes from it, but if you |
208 | process for the sole purpose of forking new processes from it, but if you |
196 | only need a fixed number of proceses you can create them, and then destroy |
209 | only need a fixed number of processes you can create them, and then destroy |
197 | the template process. |
210 | the template process. |
198 | |
211 | |
199 | Example: |
212 | Example: |
200 | |
213 | |
201 | my $template = AnyEvent::Fork->new->require ("Some::Module"); |
214 | my $template = AnyEvent::Fork->new->require ("Some::Module"); |
… | |
… | |
272 | |
285 | |
273 | #TODO: maybe append the packet to any existing string command already in the queue |
286 | #TODO: maybe append the packet to any existing string command already in the queue |
274 | |
287 | |
275 | # ideally, we would want to use "a (w/a)*" as format string, but perl versions |
288 | # ideally, we would want to use "a (w/a)*" as format string, but perl versions |
276 | # from at least 5.8.9 to 5.16.3 are all buggy and can't unpack it. |
289 | # from at least 5.8.9 to 5.16.3 are all buggy and can't unpack it. |
277 | push @{ $self->[2] }, pack "N/a*", pack "(w/a*)*", @_; |
290 | push @{ $self->[2] }, pack "L/a*", pack "(w/a*)*", @_; |
278 | |
291 | |
279 | $self->[3] ||= AE::io $self->[1], 1, sub { |
292 | $self->[3] ||= AE::io $self->[1], 1, sub { |
280 | # send the next "thing" in the queue - either a reference to an fh, |
293 | # send the next "thing" in the queue - either a reference to an fh, |
281 | # or a plain string. |
294 | # or a plain string. |
282 | |
295 | |
… | |
… | |
329 | if ($pid eq 0) { |
342 | if ($pid eq 0) { |
330 | require AnyEvent::Fork::Serve; |
343 | require AnyEvent::Fork::Serve; |
331 | $AnyEvent::Fork::Serve::OWNER = $parent; |
344 | $AnyEvent::Fork::Serve::OWNER = $parent; |
332 | close $fh; |
345 | close $fh; |
333 | $0 = "$_[1] of $parent"; |
346 | $0 = "$_[1] of $parent"; |
|
|
347 | $SIG{CHLD} = 'IGNORE'; |
334 | AnyEvent::Fork::Serve::serve ($slave); |
348 | AnyEvent::Fork::Serve::serve ($slave); |
335 | exit 0; |
349 | exit 0; |
336 | } elsif (!$pid) { |
350 | } elsif (!$pid) { |
337 | die "AnyEvent::Fork::Early/Template: unable to fork template process: $!"; |
351 | die "AnyEvent::Fork::Early/Template: unable to fork template process: $!"; |
338 | } |
352 | } |
… | |
… | |
396 | reduces the amount of memory sharing that is possible, and is also slower. |
410 | reduces the amount of memory sharing that is possible, and is also slower. |
397 | |
411 | |
398 | You should use C<new> whenever possible, except when having a template |
412 | You should use C<new> whenever possible, except when having a template |
399 | process around is unacceptable. |
413 | process around is unacceptable. |
400 | |
414 | |
401 | The path to the perl interpreter is divined usign various methods - first |
415 | The path to the perl interpreter is divined using various methods - first |
402 | C<$^X> is investigated to see if the path ends with something that sounds |
416 | C<$^X> is investigated to see if the path ends with something that sounds |
403 | as if it were the perl interpreter. Failing this, the module falls back to |
417 | as if it were the perl interpreter. Failing this, the module falls back to |
404 | using C<$Config::Config{perlpath}>. |
418 | using C<$Config::Config{perlpath}>. |
405 | |
419 | |
406 | =cut |
420 | =cut |
… | |
… | |
501 | accomplished by simply not storing the file handles anywhere after passing |
515 | accomplished by simply not storing the file handles anywhere after passing |
502 | them to this method. |
516 | them to this method. |
503 | |
517 | |
504 | Returns the process object for easy chaining of method calls. |
518 | Returns the process object for easy chaining of method calls. |
505 | |
519 | |
506 | Example: pass an fh to a process, and release it without closing. it will |
520 | Example: pass a file handle to a process, and release it without |
507 | be closed automatically when it is no longer used. |
521 | closing. It will be closed automatically when it is no longer used. |
508 | |
522 | |
509 | $proc->send_fh ($my_fh); |
523 | $proc->send_fh ($my_fh); |
510 | undef $my_fh; # free the reference if you want, but DO NOT CLOSE IT |
524 | undef $my_fh; # free the reference if you want, but DO NOT CLOSE IT |
511 | |
525 | |
512 | =cut |
526 | =cut |
… | |
… | |
525 | =item $proc = $proc->send_arg ($string, ...) |
539 | =item $proc = $proc->send_arg ($string, ...) |
526 | |
540 | |
527 | Send one or more argument strings to the process, to prepare a call to |
541 | Send one or more argument strings to the process, to prepare a call to |
528 | C<run>. The strings can be any octet string. |
542 | C<run>. The strings can be any octet string. |
529 | |
543 | |
530 | Returns the process object for easy chaining of emthod calls. |
544 | Returns the process object for easy chaining of method calls. |
531 | |
545 | |
532 | =cut |
546 | =cut |
533 | |
547 | |
534 | sub send_arg { |
548 | sub send_arg { |
535 | my ($self, @arg) = @_; |
549 | my ($self, @arg) = @_; |
… | |
… | |
556 | If the communication socket isn't used, it should be closed on both sides, |
570 | If the communication socket isn't used, it should be closed on both sides, |
557 | to save on kernel memory. |
571 | to save on kernel memory. |
558 | |
572 | |
559 | The socket is non-blocking in the parent, and blocking in the newly |
573 | The socket is non-blocking in the parent, and blocking in the newly |
560 | created process. The close-on-exec flag is set on both. Even if not used |
574 | created process. The close-on-exec flag is set on both. Even if not used |
561 | otherwise, the socket can be a good indicator for the existance of the |
575 | otherwise, the socket can be a good indicator for the existence of the |
562 | process - if the other process exits, you get a readable event on it, |
576 | process - if the other process exits, you get a readable event on it, |
563 | because exiting the process closes the socket (if it didn't create any |
577 | because exiting the process closes the socket (if it didn't create any |
564 | children using fork). |
578 | children using fork). |
565 | |
579 | |
566 | Example: create a template for a process pool, pass a few strings, some |
580 | Example: create a template for a process pool, pass a few strings, some |
… | |
… | |
603 | $self->_cmd (r => $func); |
617 | $self->_cmd (r => $func); |
604 | } |
618 | } |
605 | |
619 | |
606 | =back |
620 | =back |
607 | |
621 | |
|
|
622 | =head1 PERFORMANCE |
|
|
623 | |
|
|
624 | Now for some unscientific benchmark numbers (all done on an amd64 |
|
|
625 | GNU/Linux box). These are intended to give you an idea of the relative |
|
|
626 | performance you can expect. |
|
|
627 | |
|
|
628 | OK, so, I ran a simple benchmark that creates a socket pair, forks, calls |
|
|
629 | exit in the child and waits for the socket to close in the parent. I did |
|
|
630 | load AnyEvent, EV and AnyEvent::Fork, for a total process size of 6312kB. |
|
|
631 | |
|
|
632 | 2079 new processes per second, using socketpair + fork manually |
|
|
633 | |
|
|
634 | Then I did the same thing, but instead of calling fork, I called |
|
|
635 | AnyEvent::Fork->new->run ("CORE::exit") and then again waited for the |
|
|
636 | socket form the child to close on exit. This does the same thing as manual |
|
|
637 | socket pair + fork, except that what is forked is the template process |
|
|
638 | (2440kB), and the socket needs to be passed to the server at the other end |
|
|
639 | of the socket first. |
|
|
640 | |
|
|
641 | 2307 new processes per second, using AnyEvent::Fork->new |
|
|
642 | |
|
|
643 | And finally, using C<new_exec> instead C<new>, using vforks+execs to exec |
|
|
644 | a new perl interpreter and compile the small server each time, I get: |
|
|
645 | |
|
|
646 | 479 vfork+execs per second, using AnyEvent::Fork->new_exec |
|
|
647 | |
|
|
648 | So how can C<< AnyEvent->new >> be faster than a standard fork, even |
|
|
649 | though it uses the same operations, but adds a lot of overhead? |
|
|
650 | |
|
|
651 | The difference is simply the process size: forking the 6MB process takes |
|
|
652 | so much longer than forking the 2.5MB template process that the overhead |
|
|
653 | introduced is canceled out. |
|
|
654 | |
|
|
655 | If the benchmark process grows, the normal fork becomes even slower: |
|
|
656 | |
|
|
657 | 1340 new processes, manual fork in a 20MB process |
|
|
658 | 731 new processes, manual fork in a 200MB process |
|
|
659 | 235 new processes, manual fork in a 2000MB process |
|
|
660 | |
|
|
661 | What that means (to me) is that I can use this module without having a |
|
|
662 | very bad conscience because of the extra overhead required to start new |
|
|
663 | processes. |
|
|
664 | |
608 | =head1 TYPICAL PROBLEMS |
665 | =head1 TYPICAL PROBLEMS |
609 | |
666 | |
610 | This section lists typical problems that remain. I hope by recognising |
667 | This section lists typical problems that remain. I hope by recognising |
611 | them, most can be avoided. |
668 | them, most can be avoided. |
612 | |
669 | |
613 | =over 4 |
670 | =over 4 |
|
|
671 | |
|
|
672 | =item exit runs destructors |
614 | |
673 | |
615 | =item "leaked" file descriptors for exec'ed processes |
674 | =item "leaked" file descriptors for exec'ed processes |
616 | |
675 | |
617 | POSIX systems inherit file descriptors by default when exec'ing a new |
676 | POSIX systems inherit file descriptors by default when exec'ing a new |
618 | process. While perl itself laudably sets the close-on-exec flags on new |
677 | process. While perl itself laudably sets the close-on-exec flags on new |
619 | file handles, most C libraries don't care, and even if all cared, it's |
678 | file handles, most C libraries don't care, and even if all cared, it's |
620 | often not possible to set the flag in a race-free manner. |
679 | often not possible to set the flag in a race-free manner. |
621 | |
680 | |
622 | That means some file descriptors can leak through. And since it isn't |
681 | That means some file descriptors can leak through. And since it isn't |
623 | possible to know which file descriptors are "good" and "neccessary" (or |
682 | possible to know which file descriptors are "good" and "necessary" (or |
624 | even to know which file descreiptors are open), there is no good way to |
683 | even to know which file descriptors are open), there is no good way to |
625 | close the ones that might harm. |
684 | close the ones that might harm. |
626 | |
685 | |
627 | As an example of what "harm" can be done consider a web server that |
686 | As an example of what "harm" can be done consider a web server that |
628 | accepts connections and afterwards some module uses AnyEvent::Fork for the |
687 | accepts connections and afterwards some module uses AnyEvent::Fork for the |
629 | first time, causing it to fork and exec a new process, which might inherit |
688 | first time, causing it to fork and exec a new process, which might inherit |
… | |
… | |
637 | well before many random file descriptors are open. |
696 | well before many random file descriptors are open. |
638 | |
697 | |
639 | In general, the solution for these kind of problems is to fix the |
698 | In general, the solution for these kind of problems is to fix the |
640 | libraries or the code that leaks those file descriptors. |
699 | libraries or the code that leaks those file descriptors. |
641 | |
700 | |
642 | Fortunately, most of these lekaed descriptors do no harm, other than |
701 | Fortunately, most of these leaked descriptors do no harm, other than |
643 | sitting on some resources. |
702 | sitting on some resources. |
644 | |
703 | |
645 | =item "leaked" file descriptors for fork'ed processes |
704 | =item "leaked" file descriptors for fork'ed processes |
646 | |
705 | |
647 | Normally, L<AnyEvent::Fork> does start new processes by exec'ing them, |
706 | Normally, L<AnyEvent::Fork> does start new processes by exec'ing them, |
… | |
… | |
667 | |
726 | |
668 | Native win32 perls are somewhat supported (AnyEvent::Fork::Early is a nop, |
727 | Native win32 perls are somewhat supported (AnyEvent::Fork::Early is a nop, |
669 | and ::Template is not going to work), and it cost a lot of blood and sweat |
728 | and ::Template is not going to work), and it cost a lot of blood and sweat |
670 | to make it so, mostly due to the bloody broken perl that nobody seems to |
729 | to make it so, mostly due to the bloody broken perl that nobody seems to |
671 | care about. The fork emulation is a bad joke - I have yet to see something |
730 | care about. The fork emulation is a bad joke - I have yet to see something |
672 | useful that you cna do with it without running into memory corruption |
731 | useful that you can do with it without running into memory corruption |
673 | issues or other braindamage. Hrrrr. |
732 | issues or other braindamage. Hrrrr. |
674 | |
733 | |
675 | Cygwin perl is not supported at the moment, as it should implement fd |
734 | Cygwin perl is not supported at the moment, as it should implement fd |
676 | passing, but doesn't, and rolling my own is hard, as cygwin doesn't |
735 | passing, but doesn't, and rolling my own is hard, as cygwin doesn't |
677 | support enough functionality to do it. |
736 | support enough functionality to do it. |