@ -1,128 +1,56 @@
libev is a high-performance event loop/event model with lots of features.
(see benchmark at http://libev.schmorp.de/bench.html)
Library Documentation: http://pod.tst.eu/http://cvs.schmorp.de/libev/ev.pod
It is modelled (very losely) after libevent and the Event perl module,
but aims to be faster and more correct, and also more featureful. And
also smaller. Yay.
ABOUT THIS DISTRIBUTION
If you downloaded the libevent+libev distribution of libev, you will
find it looks very much like libevent. In fact, the distributed libev
tarballs are indeed libevent tarballs patched up with the libev
event core, taking the evbuffer, evtag, evdns and evhttpd parts from
libevent (they use the libevent emulation inside libev). Configure and
Makefile stuff is also a more or less direct copy of libevent, and are
maintained by the libevent authors.
If you downloaded the libev distribution (without libevent), then
you only get the core parts of the library, meaning http and dns
client/server code and similar things are missing. Only the core event
loop is included.
If you are looking for an easily embeddable version, I recommend using
the libev standalone distribution or the CVS repository.
Libev is modelled (very losely) after libevent and the Event perl
module, but is faster, scales better and is more correct, and also more
featureful. And also smaller. Yay.
Some of the specialties of libev not commonly found elsewhere are:
- extensive and detailed, readable documentation (not doxygen garbage).
- fully supports fork, can detect fork in various ways and automatically
re-arms kernel mechanisms that do not support fork.
- highly optimised select, poll, epoll, kqueue and event ports backends.
- filesystem object (path) watching (with optional linux inotify support).
- wallclock-based times (using absolute time, cron-like).
- relative timers/timeouts (handle time jumps).
- fast intra-thread communication between multiple
event loops (with optional fast linux eventfd backend).
- extremely easy to embed.
- very small codebase, no bloated library.
- fully extensible by being able to plug into the event loop,
integrate other event loops, integrate other event loop users.
- very little memory use (small watchers, small event loop data).
- optional C++ interface allowing method and function callbacks
at no extra memory or runtime overhead.
- optional Perl interface with similar characteristics (capable
of running Glib/Gtk2 on libev, interfaces with Net::SNMP and
- support for other languages (multiple C++ interfaces, D, Ruby,
Python) available from third-parties.
Examples of programs that embed libev: the EV perl module,
rxvt-unicode, gvpe (GNU Virtual Private Ethernet) and deliantra
DIFFERENCES AND COMPARISON TO LIBEVENT
The comparisons below are relative to libevent-1.3e.
- multiple watchers can wait for the same event without deregistering others,
both for file descriptors as well as signals.
(registering two read events on fd 10 and unregistering one will not
break the other).
- fork() is supported and can be handled
(there is no way to recover from a fork with libevent).
- timers are handled as a priority queue (important operations are O(1))
(libevent uses a much less efficient but more complex red-black tree).
- supports absolute (wallclock-based) timers in addition to relative ones,
i.e. can schedule timers to occur after n seconds, or at a specific time.
- timers can be repeating (both absolute and relative ones).
- absolute timers can have customised rescheduling hooks (suitable for cron-like
- detects time jumps and adjusts timers
(works for both forward and backward time jumps and also for absolute timers).
- race-free signal processing
(libevent may delay processing signals till after the next event).
- more efficient epoll backend
(stopping and starting an io watcher between two loop iterations will not
result in spurious epoll_ctl calls).
- usually less calls to gettimeofday and clock_gettime
(libevent calls it on every timer event change, libev twice per iteration).
- watchers use less memory
(libevent watcher on amd64: 152 bytes, libev native: <= 56 bytes, libevent emulation: 144 bytes).
- library uses less memory
(libevent allocates large data structures wether used or not, libev
scales all its data structures dynamically).
- no hardcoded arbitrary limits
(libevent contains an off-by-one bug and sometimes hardcodes limits).
- libev separates timer, signal and io watchers from each other
(libevent combines them, but with libev you can combine them yourself
by reusing the same callback and still save memory).
- simpler design, backends are potentially much simpler
(in libevent, backends have to deal with watchers, thus the problems with
wildly different semantics between diferent backends)
(epoll backend in libevent: 366 lines no caching, libev: 90 lines full caching).
- libev handles EBADF gracefully by removing the offending fds.
- libev communicates errors to the callback, libevent to the
event adder or not at all.
- doesn't rely on nonportable BSD header files.
- an event.h compatibility header exists, and can be used to run a wide
range of libevent programs unchanged (such as evdns.c).
- win32 compatibility for the core parts.
(the backend is fd-based as documented and on other platforms,
not handle-based like libevent, and can be used for both winscoket environments
and unix-like ones).
- libev can be embedded easily with or without autoconf support into
other programs, with no changes to the source code necessary.
- the event core library (ev and event layer) compiles and works both as
C and C++.
- a simple C++ wrapper that supports methods as callbacks exists.
- a full featured and widely used perl module is available.
rxvt-unicode, gvpe (GNU Virtual Private Ethernet), the Deliantra MMORPG
server (http://www.deliantra.net/), Rubinius (a next-generation Ruby
VM), the Ebb web server, the Rev event toolkit.
- no event-like priority support at the moment (the ev priorities work
differently, but you can use idle watchers to get a similar effect).
libev was written and designed by Marc Lehmann and Emanuele Giaquinta.
The following people sent in patches or made other noteworthy
contributions to the design (if I forgot to include you, please shout
at me, it was an accident):
contributions to the design (for minor patches, see the Changes
file. If I forgot to include you, please shout at me, it was an