… | |
… | |
392 | |
392 | |
393 | There is a slight catch to child watchers, however: you usually start them |
393 | There is a slight catch to child watchers, however: you usually start them |
394 | I<after> the child process was created, and this means the process could |
394 | I<after> the child process was created, and this means the process could |
395 | have exited already (and no SIGCHLD will be sent anymore). |
395 | have exited already (and no SIGCHLD will be sent anymore). |
396 | |
396 | |
397 | Not all event models handle this correctly (POE doesn't), but even for |
397 | Not all event models handle this correctly (neither POE nor IO::Async do, |
|
|
398 | see their AnyEvent::Impl manpages for details), but even for event models |
398 | event models that I<do> handle this correctly, they usually need to be |
399 | that I<do> handle this correctly, they usually need to be loaded before |
399 | loaded before the process exits (i.e. before you fork in the first place). |
400 | the process exits (i.e. before you fork in the first place). AnyEvent's |
|
|
401 | pure perl event loop handles all cases correctly regardless of when you |
|
|
402 | start the watcher. |
400 | |
403 | |
401 | This means you cannot create a child watcher as the very first thing in an |
404 | This means you cannot create a child watcher as the very first |
402 | AnyEvent program, you I<have> to create at least one watcher before you |
405 | thing in an AnyEvent program, you I<have> to create at least one |
403 | C<fork> the child (alternatively, you can call C<AnyEvent::detect>). |
406 | watcher before you C<fork> the child (alternatively, you can call |
|
|
407 | C<AnyEvent::detect>). |
404 | |
408 | |
405 | Example: fork a process and wait for it |
409 | Example: fork a process and wait for it |
406 | |
410 | |
407 | my $done = AnyEvent->condvar; |
411 | my $done = AnyEvent->condvar; |
408 | |
412 | |
… | |
… | |
730 | AnyEvent::Impl::Tk based on Tk, very bad choice. |
734 | AnyEvent::Impl::Tk based on Tk, very bad choice. |
731 | AnyEvent::Impl::Qt based on Qt, cannot be autoprobed (see its docs). |
735 | AnyEvent::Impl::Qt based on Qt, cannot be autoprobed (see its docs). |
732 | AnyEvent::Impl::EventLib based on Event::Lib, leaks memory and worse. |
736 | AnyEvent::Impl::EventLib based on Event::Lib, leaks memory and worse. |
733 | AnyEvent::Impl::POE based on POE, not generic enough for full support. |
737 | AnyEvent::Impl::POE based on POE, not generic enough for full support. |
734 | |
738 | |
|
|
739 | # warning, support for IO::Async is only partial, as it is too broken |
|
|
740 | # and limited toe ven support the AnyEvent API. See AnyEvent::Impl::Async. |
|
|
741 | AnyEvent::Impl::IOAsync based on IO::Async, cannot be autoprobed (see its docs). |
|
|
742 | |
735 | There is no support for WxWidgets, as WxWidgets has no support for |
743 | There is no support for WxWidgets, as WxWidgets has no support for |
736 | watching file handles. However, you can use WxWidgets through the |
744 | watching file handles. However, you can use WxWidgets through the |
737 | POE Adaptor, as POE has a Wx backend that simply polls 20 times per |
745 | POE Adaptor, as POE has a Wx backend that simply polls 20 times per |
738 | second, which was considered to be too horrible to even consider for |
746 | second, which was considered to be too horrible to even consider for |
739 | AnyEvent. Likewise, other POE backends can be used by AnyEvent by using |
747 | AnyEvent. Likewise, other POE backends can be used by AnyEvent by using |
… | |
… | |
931 | no warnings; |
939 | no warnings; |
932 | use strict qw(vars subs); |
940 | use strict qw(vars subs); |
933 | |
941 | |
934 | use Carp; |
942 | use Carp; |
935 | |
943 | |
936 | our $VERSION = 4.411; |
944 | our $VERSION = 4.412; |
937 | our $MODEL; |
945 | our $MODEL; |
938 | |
946 | |
939 | our $AUTOLOAD; |
947 | our $AUTOLOAD; |
940 | our @ISA; |
948 | our @ISA; |
941 | |
949 | |
… | |
… | |
974 | [Event::Lib:: => AnyEvent::Impl::EventLib::], # too buggy |
982 | [Event::Lib:: => AnyEvent::Impl::EventLib::], # too buggy |
975 | [Qt:: => AnyEvent::Impl::Qt::], # requires special main program |
983 | [Qt:: => AnyEvent::Impl::Qt::], # requires special main program |
976 | [POE::Kernel:: => AnyEvent::Impl::POE::], # lasciate ogni speranza |
984 | [POE::Kernel:: => AnyEvent::Impl::POE::], # lasciate ogni speranza |
977 | [Wx:: => AnyEvent::Impl::POE::], |
985 | [Wx:: => AnyEvent::Impl::POE::], |
978 | [Prima:: => AnyEvent::Impl::POE::], |
986 | [Prima:: => AnyEvent::Impl::POE::], |
|
|
987 | # IO::Async is just too broken - we would need workaorunds for its |
|
|
988 | # byzantine signal and broken child handling, among others. |
|
|
989 | # IO::Async is rather hard to detect, as it doesn't have any |
|
|
990 | # obvious default class. |
|
|
991 | # [IO::Async:: => AnyEvent::Impl::IOAsync::], # requires special main program |
|
|
992 | # [IO::Async::Loop:: => AnyEvent::Impl::IOAsync::], # requires special main program |
|
|
993 | # [IO::Async::Notifier:: => AnyEvent::Impl::IOAsync::], # requires special main program |
979 | ); |
994 | ); |
980 | |
995 | |
981 | our %method = map +($_ => 1), |
996 | our %method = map +($_ => 1), |
982 | qw(io timer time now now_update signal child idle condvar one_event DESTROY); |
997 | qw(io timer time now now_update signal child idle condvar one_event DESTROY); |
983 | |
998 | |
… | |
… | |
1075 | } |
1090 | } |
1076 | |
1091 | |
1077 | # utility function to dup a filehandle. this is used by many backends |
1092 | # utility function to dup a filehandle. this is used by many backends |
1078 | # to support binding more than one watcher per filehandle (they usually |
1093 | # to support binding more than one watcher per filehandle (they usually |
1079 | # allow only one watcher per fd, so we dup it to get a different one). |
1094 | # allow only one watcher per fd, so we dup it to get a different one). |
1080 | sub _dupfh($$$$) { |
1095 | sub _dupfh($$;$$) { |
1081 | my ($poll, $fh, $r, $w) = @_; |
1096 | my ($poll, $fh, $r, $w) = @_; |
1082 | |
1097 | |
1083 | # cygwin requires the fh mode to be matching, unix doesn't |
1098 | # cygwin requires the fh mode to be matching, unix doesn't |
1084 | my ($rw, $mode) = $poll eq "r" ? ($r, "<") |
1099 | my ($rw, $mode) = $poll eq "r" ? ($r, "<") |
1085 | : $poll eq "w" ? ($w, ">") |
1100 | : $poll eq "w" ? ($w, ">") |
… | |
… | |
1366 | =item C<PERL_ANYEVENT_STRICT> |
1381 | =item C<PERL_ANYEVENT_STRICT> |
1367 | |
1382 | |
1368 | AnyEvent does not do much argument checking by default, as thorough |
1383 | AnyEvent does not do much argument checking by default, as thorough |
1369 | argument checking is very costly. Setting this variable to a true value |
1384 | argument checking is very costly. Setting this variable to a true value |
1370 | will cause AnyEvent to load C<AnyEvent::Strict> and then to thoroughly |
1385 | will cause AnyEvent to load C<AnyEvent::Strict> and then to thoroughly |
1371 | check the arguments passed to most method calls. If it finds any problems |
1386 | check the arguments passed to most method calls. If it finds any problems, |
1372 | it will croak. |
1387 | it will croak. |
1373 | |
1388 | |
1374 | In other words, enables "strict" mode. |
1389 | In other words, enables "strict" mode. |
1375 | |
1390 | |
1376 | Unlike C<use strict>, it is definitely recommended ot keep it off in |
1391 | Unlike C<use strict>, it is definitely recommended to keep it off in |
1377 | production. Keeping C<PERL_ANYEVENT_STRICT=1> in your environment while |
1392 | production. Keeping C<PERL_ANYEVENT_STRICT=1> in your environment while |
1378 | developing programs can be very useful, however. |
1393 | developing programs can be very useful, however. |
1379 | |
1394 | |
1380 | =item C<PERL_ANYEVENT_MODEL> |
1395 | =item C<PERL_ANYEVENT_MODEL> |
1381 | |
1396 | |
… | |
… | |
1680 | EV/Any 100000 224 2.88 0.34 0.27 EV + AnyEvent watchers |
1695 | EV/Any 100000 224 2.88 0.34 0.27 EV + AnyEvent watchers |
1681 | CoroEV/Any 100000 224 2.85 0.35 0.28 coroutines + Coro::Signal |
1696 | CoroEV/Any 100000 224 2.85 0.35 0.28 coroutines + Coro::Signal |
1682 | Perl/Any 100000 452 4.13 0.73 0.95 pure perl implementation |
1697 | Perl/Any 100000 452 4.13 0.73 0.95 pure perl implementation |
1683 | Event/Event 16000 517 32.20 31.80 0.81 Event native interface |
1698 | Event/Event 16000 517 32.20 31.80 0.81 Event native interface |
1684 | Event/Any 16000 590 35.85 31.55 1.06 Event + AnyEvent watchers |
1699 | Event/Any 16000 590 35.85 31.55 1.06 Event + AnyEvent watchers |
|
|
1700 | IOAsync/Any 16000 989 38.10 32.77 11.13 via IO::Async::Loop::IO_Poll |
|
|
1701 | IOAsync/Any 16000 990 37.59 29.50 10.61 via IO::Async::Loop::Epoll |
1685 | Glib/Any 16000 1357 102.33 12.31 51.00 quadratic behaviour |
1702 | Glib/Any 16000 1357 102.33 12.31 51.00 quadratic behaviour |
1686 | Tk/Any 2000 1860 27.20 66.31 14.00 SEGV with >> 2000 watchers |
1703 | Tk/Any 2000 1860 27.20 66.31 14.00 SEGV with >> 2000 watchers |
1687 | POE/Event 2000 6328 109.99 751.67 14.02 via POE::Loop::Event |
1704 | POE/Event 2000 6328 109.99 751.67 14.02 via POE::Loop::Event |
1688 | POE/Select 2000 6027 94.54 809.13 579.80 via POE::Loop::Select |
1705 | POE/Select 2000 6027 94.54 809.13 579.80 via POE::Loop::Select |
1689 | |
1706 | |
… | |
… | |
1718 | performance becomes really bad with lots of file descriptors (and few of |
1735 | performance becomes really bad with lots of file descriptors (and few of |
1719 | them active), of course, but this was not subject of this benchmark. |
1736 | them active), of course, but this was not subject of this benchmark. |
1720 | |
1737 | |
1721 | The C<Event> module has a relatively high setup and callback invocation |
1738 | The C<Event> module has a relatively high setup and callback invocation |
1722 | cost, but overall scores in on the third place. |
1739 | cost, but overall scores in on the third place. |
|
|
1740 | |
|
|
1741 | C<IO::Async> performs admirably well, about on par with C<Event>, even |
|
|
1742 | when using its pure perl backend. |
1723 | |
1743 | |
1724 | C<Glib>'s memory usage is quite a bit higher, but it features a |
1744 | C<Glib>'s memory usage is quite a bit higher, but it features a |
1725 | faster callback invocation and overall ends up in the same class as |
1745 | faster callback invocation and overall ends up in the same class as |
1726 | C<Event>. However, Glib scales extremely badly, doubling the number of |
1746 | C<Event>. However, Glib scales extremely badly, doubling the number of |
1727 | watchers increases the processing time by more than a factor of four, |
1747 | watchers increases the processing time by more than a factor of four, |
… | |
… | |
1805 | it to another server. This includes deleting the old timeout and creating |
1825 | it to another server. This includes deleting the old timeout and creating |
1806 | a new one that moves the timeout into the future. |
1826 | a new one that moves the timeout into the future. |
1807 | |
1827 | |
1808 | =head3 Results |
1828 | =head3 Results |
1809 | |
1829 | |
1810 | name sockets create request |
1830 | name sockets create request |
1811 | EV 20000 69.01 11.16 |
1831 | EV 20000 69.01 11.16 |
1812 | Perl 20000 73.32 35.87 |
1832 | Perl 20000 73.32 35.87 |
|
|
1833 | IOAsync 20000 157.00 98.14 epoll |
|
|
1834 | IOAsync 20000 159.31 616.06 poll |
1813 | Event 20000 212.62 257.32 |
1835 | Event 20000 212.62 257.32 |
1814 | Glib 20000 651.16 1896.30 |
1836 | Glib 20000 651.16 1896.30 |
1815 | POE 20000 349.67 12317.24 uses POE::Loop::Event |
1837 | POE 20000 349.67 12317.24 uses POE::Loop::Event |
1816 | |
1838 | |
1817 | =head3 Discussion |
1839 | =head3 Discussion |
1818 | |
1840 | |
1819 | This benchmark I<does> measure scalability and overall performance of the |
1841 | This benchmark I<does> measure scalability and overall performance of the |
1820 | particular event loop. |
1842 | particular event loop. |
… | |
… | |
1822 | EV is again fastest. Since it is using epoll on my system, the setup time |
1844 | EV is again fastest. Since it is using epoll on my system, the setup time |
1823 | is relatively high, though. |
1845 | is relatively high, though. |
1824 | |
1846 | |
1825 | Perl surprisingly comes second. It is much faster than the C-based event |
1847 | Perl surprisingly comes second. It is much faster than the C-based event |
1826 | loops Event and Glib. |
1848 | loops Event and Glib. |
|
|
1849 | |
|
|
1850 | IO::Async performs very well when using its epoll backend, and still quite |
|
|
1851 | good compared to Glib when using its pure perl backend. |
1827 | |
1852 | |
1828 | Event suffers from high setup time as well (look at its code and you will |
1853 | Event suffers from high setup time as well (look at its code and you will |
1829 | understand why). Callback invocation also has a high overhead compared to |
1854 | understand why). Callback invocation also has a high overhead compared to |
1830 | the C<< $_->() for .. >>-style loop that the Perl event loop uses. Event |
1855 | the C<< $_->() for .. >>-style loop that the Perl event loop uses. Event |
1831 | uses select or poll in basically all documented configurations. |
1856 | uses select or poll in basically all documented configurations. |
… | |
… | |
1894 | =item * C-based event loops perform very well with small number of |
1919 | =item * C-based event loops perform very well with small number of |
1895 | watchers, as the management overhead dominates. |
1920 | watchers, as the management overhead dominates. |
1896 | |
1921 | |
1897 | =back |
1922 | =back |
1898 | |
1923 | |
|
|
1924 | =head2 THE IO::Lambda BENCHMARK |
|
|
1925 | |
|
|
1926 | Recently I was told about the benchmark in the IO::Lambda manpage, which |
|
|
1927 | could be misinterpreted to make AnyEvent look bad. In fact, the benchmark |
|
|
1928 | simply compares IO::Lambda with POE, and IO::Lambda looks better (which |
|
|
1929 | shouldn't come as a surprise to anybody). As such, the benchmark is |
|
|
1930 | fine, and mostly shows that the AnyEvent backend from IO::Lambda isn't |
|
|
1931 | very optimal. But how would AnyEvent compare when used without the extra |
|
|
1932 | baggage? To explore this, I wrote the equivalent benchmark for AnyEvent. |
|
|
1933 | |
|
|
1934 | The benchmark itself creates an echo-server, and then, for 500 times, |
|
|
1935 | connects to the echo server, sends a line, waits for the reply, and then |
|
|
1936 | creates the next connection. This is a rather bad benchmark, as it doesn't |
|
|
1937 | test the efficiency of the framework or much non-blocking I/O, but it is a |
|
|
1938 | benchmark nevertheless. |
|
|
1939 | |
|
|
1940 | name runtime |
|
|
1941 | Lambda/select 0.330 sec |
|
|
1942 | + optimized 0.122 sec |
|
|
1943 | Lambda/AnyEvent 0.327 sec |
|
|
1944 | + optimized 0.138 sec |
|
|
1945 | Raw sockets/select 0.077 sec |
|
|
1946 | POE/select, components 0.662 sec |
|
|
1947 | POE/select, raw sockets 0.226 sec |
|
|
1948 | POE/select, optimized 0.404 sec |
|
|
1949 | |
|
|
1950 | AnyEvent/select/nb 0.085 sec |
|
|
1951 | AnyEvent/EV/nb 0.068 sec |
|
|
1952 | +state machine 0.134 sec |
|
|
1953 | |
|
|
1954 | The benchmark is also a bit unfair (my fault): the IO::Lambda/POE |
|
|
1955 | benchmarks actually make blocking connects and use 100% blocking I/O, |
|
|
1956 | defeating the purpose of an event-based solution. All of the newly |
|
|
1957 | written AnyEvent benchmarks use 100% non-blocking connects (using |
|
|
1958 | AnyEvent::Socket::tcp_connect and the asynchronous pure perl DNS |
|
|
1959 | resolver), so AnyEvent is at a disadvantage here, as non-blocking connects |
|
|
1960 | generally require a lot more bookkeeping and event handling than blocking |
|
|
1961 | connects (which involve a single syscall only). |
|
|
1962 | |
|
|
1963 | The last AnyEvent benchmark additionally uses L<AnyEvent::Handle>, which |
|
|
1964 | offers similar expressive power as POE and IO::Lambda, using conventional |
|
|
1965 | Perl syntax. This means that both the echo server and the client are 100% |
|
|
1966 | non-blocking, further placing it at a disadvantage. |
|
|
1967 | |
|
|
1968 | As you can see, the AnyEvent + EV combination even beats the |
|
|
1969 | hand-optimised "raw sockets benchmark", while AnyEvent + its pure perl |
|
|
1970 | backend easily beats IO::Lambda and POE. |
|
|
1971 | |
|
|
1972 | And even the 100% non-blocking version written using the high-level (and |
|
|
1973 | slow :) L<AnyEvent::Handle> abstraction beats both POE and IO::Lambda by a |
|
|
1974 | large margin, even though it does all of DNS, tcp-connect and socket I/O |
|
|
1975 | in a non-blocking way. |
|
|
1976 | |
|
|
1977 | The two AnyEvent benchmarks programs can be found as F<eg/ae0.pl> and |
|
|
1978 | F<eg/ae2.pl> in the AnyEvent distribution, the remaining benchmarks are |
|
|
1979 | part of the IO::lambda distribution and were used without any changes. |
|
|
1980 | |
1899 | |
1981 | |
1900 | =head1 SIGNALS |
1982 | =head1 SIGNALS |
1901 | |
1983 | |
1902 | AnyEvent currently installs handlers for these signals: |
1984 | AnyEvent currently installs handlers for these signals: |
1903 | |
1985 | |
… | |
… | |
1906 | =item SIGCHLD |
1988 | =item SIGCHLD |
1907 | |
1989 | |
1908 | A handler for C<SIGCHLD> is installed by AnyEvent's child watcher |
1990 | A handler for C<SIGCHLD> is installed by AnyEvent's child watcher |
1909 | emulation for event loops that do not support them natively. Also, some |
1991 | emulation for event loops that do not support them natively. Also, some |
1910 | event loops install a similar handler. |
1992 | event loops install a similar handler. |
|
|
1993 | |
|
|
1994 | If, when AnyEvent is loaded, SIGCHLD is set to IGNORE, then AnyEvent will |
|
|
1995 | reset it to default, to avoid losing child exit statuses. |
1911 | |
1996 | |
1912 | =item SIGPIPE |
1997 | =item SIGPIPE |
1913 | |
1998 | |
1914 | A no-op handler is installed for C<SIGPIPE> when C<$SIG{PIPE}> is C<undef> |
1999 | A no-op handler is installed for C<SIGPIPE> when C<$SIG{PIPE}> is C<undef> |
1915 | when AnyEvent gets loaded. |
2000 | when AnyEvent gets loaded. |
… | |
… | |
1927 | |
2012 | |
1928 | =back |
2013 | =back |
1929 | |
2014 | |
1930 | =cut |
2015 | =cut |
1931 | |
2016 | |
|
|
2017 | undef $SIG{CHLD} |
|
|
2018 | if $SIG{CHLD} eq 'IGNORE'; |
|
|
2019 | |
1932 | $SIG{PIPE} = sub { } |
2020 | $SIG{PIPE} = sub { } |
1933 | unless defined $SIG{PIPE}; |
2021 | unless defined $SIG{PIPE}; |
1934 | |
|
|
1935 | |
2022 | |
1936 | =head1 FORK |
2023 | =head1 FORK |
1937 | |
2024 | |
1938 | Most event libraries are not fork-safe. The ones who are usually are |
2025 | Most event libraries are not fork-safe. The ones who are usually are |
1939 | because they rely on inefficient but fork-safe C<select> or C<poll> |
2026 | because they rely on inefficient but fork-safe C<select> or C<poll> |
… | |
… | |
1962 | Similar considerations apply to $ENV{PERL_ANYEVENT_VERBOSE}, as that can |
2049 | Similar considerations apply to $ENV{PERL_ANYEVENT_VERBOSE}, as that can |
1963 | be used to probe what backend is used and gain other information (which is |
2050 | be used to probe what backend is used and gain other information (which is |
1964 | probably even less useful to an attacker than PERL_ANYEVENT_MODEL), and |
2051 | probably even less useful to an attacker than PERL_ANYEVENT_MODEL), and |
1965 | $ENV{PERL_ANYEVENT_STRICT}. |
2052 | $ENV{PERL_ANYEVENT_STRICT}. |
1966 | |
2053 | |
|
|
2054 | Note that AnyEvent will remove I<all> environment variables starting with |
|
|
2055 | C<PERL_ANYEVENT_> from C<%ENV> when it is loaded while taint mode is |
|
|
2056 | enabled. |
|
|
2057 | |
1967 | |
2058 | |
1968 | =head1 BUGS |
2059 | =head1 BUGS |
1969 | |
2060 | |
1970 | Perl 5.8 has numerous memleaks that sometimes hit this module and are hard |
2061 | Perl 5.8 has numerous memleaks that sometimes hit this module and are hard |
1971 | to work around. If you suffer from memleaks, first upgrade to Perl 5.10 |
2062 | to work around. If you suffer from memleaks, first upgrade to Perl 5.10 |