ViewVC Help
View File | Revision Log | Show Annotations | Download File
/cvs/AnyEvent/lib/AnyEvent.pm
(Generate patch)

Comparing AnyEvent/lib/AnyEvent.pm (file contents):
Revision 1.67 by root, Fri Apr 25 06:58:38 2008 UTC vs.
Revision 1.77 by root, Fri Apr 25 09:00:37 2008 UTC

… …
864 864
865 865
866=head1 BENCHMARK 866=head1 BENCHMARK
867 867
868To give you an idea of the performance and overheads that AnyEvent adds 868To give you an idea of the performance and overheads that AnyEvent adds
869over the backends directly, here is a benchmark of various supported event 869over the event loops themselves (and to give you an impression of the
870speed of various event loops), here is a benchmark of various supported
870models natively and with anyevent. The benchmark creates a lot of timers 871event models natively and with anyevent. The benchmark creates a lot of
871(with a zero timeout) and io watchers (watching STDOUT, a pty, to become 872timers (with a zero timeout) and io watchers (watching STDOUT, a pty, to
872writable, which it is), lets them fire exactly once and destroys them 873become writable, which it is), lets them fire exactly once and destroys
873again. 874them again.
874 875
875Explanation of the fields: 876Rewriting the benchmark to use many different sockets instead of using
877the same filehandle for all io watchers results in a much longer runtime
878(socket creation is expensive), but qualitatively the same figures, so it
879was not used.
876 880
881=head2 Explanation of the columns
882
877I<watcher> is the number of event watchers created/destroyed. Sicne 883I<watcher> is the number of event watchers created/destroyed. Since
878different event models have vastly different performance each backend was 884different event models feature vastly different performances, each event
879handed a number of watchers so that overall runtime is acceptable and 885loop was given a number of watchers so that overall runtime is acceptable
880similar to all backends (and keep them from crashing). 886and similar between tested event loop (and keep them from crashing): Glib
887would probably take thousands of years if asked to process the same number
888of watchers as EV in this benchmark.
881 889
882I<bytes> is the number of bytes (as measured by resident set size) used by 890I<bytes> is the number of bytes (as measured by the resident set size,
883each watcher. 891RSS) consumed by each watcher. This method of measuring captures both C
892and Perl-based overheads.
884 893
885I<create> is the time, in microseconds, to create a single watcher. 894I<create> is the time, in microseconds (millionths of seconds), that it
895takes to create a single watcher. The callback is a closure shared between
896all watchers, to avoid adding memory overhead. That means closure creation
897and memory usage is not included in the figures.
886 898
887I<invoke> is the time, in microseconds, used to invoke a simple callback 899I<invoke> is the time, in microseconds, used to invoke a simple
888that simply counts down. 900callback. The callback simply counts down a Perl variable and after it was
901invoked "watcher" times, it would C<< ->broadcast >> a condvar once to
902signal the end of this phase.
889 903
890I<destroy> is the time, in microseconds, to destroy a single watcher. 904I<destroy> is the time, in microseconds, that it takes to destroy a single
905watcher.
891 906
907=head2 Results
908
892 name watcher bytes create invoke destroy comment 909 name watchers bytes create invoke destroy comment
893 EV/EV 400000 244 0.56 0.46 0.31 EV native interface 910 EV/EV 400000 244 0.56 0.46 0.31 EV native interface
894 EV/Any 100000 610 3.52 0.91 0.75 911 EV/Any 100000 610 3.52 0.91 0.75 EV + AnyEvent watchers
895 CoroEV/Any 100000 610 3.49 0.92 0.75 coroutines + Coro::Signal 912 CoroEV/Any 100000 610 3.49 0.92 0.75 coroutines + Coro::Signal
896 Perl/Any 10000 654 4.64 1.22 0.77 pure perl implementation 913 Perl/Any 100000 513 4.91 0.92 1.15 pure perl implementation
897 Event/Event 10000 523 28.05 21.38 5.22 Event native interface 914 Event/Event 16000 523 28.05 21.38 0.86 Event native interface
898 Event/Any 10000 943 34.43 20.48 1.39 915 Event/Any 16000 943 34.43 20.48 1.39 Event + AnyEvent watchers
899 Glib/Any 16000 1357 96.99 12.55 55.51 quadratic behaviour 916 Glib/Any 16000 1357 96.99 12.55 55.51 quadratic behaviour
900 Tk/Any 2000 1855 27.01 66.61 14.03 SEGV with >> 2000 watchers 917 Tk/Any 2000 1855 27.01 66.61 14.03 SEGV with >> 2000 watchers
918 POE/Event 2000 6644 108.15 768.19 14.33 via POE::Loop::Event
901 POE/Select 2000 6343 94.69 807.65 562.69 POE::Loop::Select 919 POE/Select 2000 6343 94.69 807.65 562.69 via POE::Loop::Select
902 POE/Event 2000 6644 108.15 768.19 14.33 POE::Loop::Event
903 920
904Discussion: The benchmark does I<not> bench scalability of the 921=head2 Discussion
922
923The benchmark does I<not> measure scalability of the event loop very
905backend. For example a select-based backend (such as the pureperl one) can 924well. For example, a select-based event loop (such as the pure perl one)
906never compete with a backend using epoll. In this benchmark, only a single 925can never compete with an event loop that uses epoll when the number of
907filehandle is used. 926file descriptors grows high. In this benchmark, only a single filehandle
927is used (although some of the AnyEvent adaptors dup() its file descriptor
928to worka round bugs).
908 929
909EV is the sole leader regarding speed and memory use, which are both 930C<EV> is the sole leader regarding speed and memory use, which are both
910maximal/minimal. Even when going through AnyEvent, there is only one event 931maximal/minimal, respectively. Even when going through AnyEvent, there are
911loop that uses less memory (the Event module natively), and no faster 932only two event loops that use slightly less memory (the C<Event> module
912event model. 933natively and the pure perl backend), and no faster event models, not even
934C<Event> natively.
913 935
914The pure perl implementation is hit in a few sweet spots (both the 936The pure perl implementation is hit in a few sweet spots (both the
915zero timeout and the use of a single fd hit optimisations in the perl 937zero timeout and the use of a single fd hit optimisations in the perl
916interpreter and the backend itself), but it shows that it adds very little 938interpreter and the backend itself, and all watchers become ready at the
939same time). Nevertheless this shows that it adds very little overhead in
917overhead in itself. Like any select-based backend it's performance becomes 940itself. Like any select-based backend its performance becomes really bad
918really bad with lots of file descriptors. 941with lots of file descriptors (and few of them active), of course, but
942this was not subject of this benchmark.
919 943
920The Event module has a relatively high setup and callback invocation cost, 944The C<Event> module has a relatively high setup and callback invocation cost,
921but overall scores on the third place. 945but overall scores on the third place.
922 946
923Glib has a little higher memory cost, a bit fster callback invocation and 947C<Glib>'s memory usage is quite a bit bit higher, but it features a
924has a similar speed as Event. 948faster callback invocation and overall ends up in the same class as
949C<Event>. However, Glib scales extremely badly, doubling the number of
950watchers increases the processing time by more than a factor of four,
951making it completely unusable when using larger numbers of watchers
952(note that only a single file descriptor was used in the benchmark, so
953inefficiencies of C<poll> do not account for this).
925 954
926The Tk backend works relatively well, the fact that it crashes with 955The C<Tk> adaptor works relatively well. The fact that it crashes with
927more than 2000 watchers is a big setback, however, as correctness takes 956more than 2000 watchers is a big setback, however, as correctness takes
928precedence over speed. 957precedence over speed. Nevertheless, its performance is surprising, as the
958file descriptor is dup()ed for each watcher. This shows that the dup()
959employed by some adaptors is not a big performance issue (it does incur a
960hidden memory cost inside the kernel, though, that is not reflected in the
961figures above).
929 962
930POE, regardless of backend (wether it's pure perl select backend or the 963C<POE>, regardless of underlying event loop (wether using its pure perl
931Event backend) shows abysmal performance and memory usage: Watchers use 964select-based backend or the Event module) shows abysmal performance and
932almost 30 times as much memory as EV watchers, and 10 times as much memory 965memory usage: Watchers use almost 30 times as much memory as EV watchers,
933as both Event or EV via AnyEvent. 966and 10 times as much memory as both Event or EV via AnyEvent. Watcher
967invocation is almost 700 times slower than with AnyEvent's pure perl
968implementation. The design of the POE adaptor class in AnyEvent can not
969really account for this, as session creation overhead is small compared
970to execution of the state machine, which is coded pretty optimally within
971L<AnyEvent::Impl::POE>. POE simply seems to be abysmally slow.
934 972
973=head2 Summary
974
935Summary: using EV through AnyEvent is faster than any other event 975Using EV through AnyEvent is faster than any other event loop, but most
936loop. The overhead AnyEvent adds can be very small, and you should avoid 976event loops have acceptable performance with or without AnyEvent.
937POE like the plague if you want performance or reasonable memory usage. 977
978The overhead AnyEvent adds is usually much smaller than the overhead of
979the actual event loop, only with extremely fast event loops such as the EV
980adds AnyEvent significant overhead.
981
982And you should simply avoid POE like the plague if you want performance or
983reasonable memory usage.
938 984
939 985
940=head1 FORK 986=head1 FORK
941 987
942Most event libraries are not fork-safe. The ones who are usually are 988Most event libraries are not fork-safe. The ones who are usually are

Diff Legend

– Removed lines
+ Added lines
< Changed lines
> Changed lines