On Fri, Sep 9, 2011 at 12:32 AM, Scott Lamb <sl...@slamb.org> wrote: > I see your point, fwiw. I just wish it were possible to do much better > than that, but I don't think it is given the available OS interfaces. > One could imagine a bufferevent implementation that used aio, though > I'm not sure what OSs have a satisfactory aio implementation...
It looks like Linux's might be okay. At least the io_* functions defined in <libaio.h> appear to work fine. This is a completely Linux-specific interface, but despite what some webpages say, it doesn't seem to require anything special on a modern system (in particular, O_DIRECT is not required and so neither are aligned operations). You can also poll for these events with epoll through io_set_eventfd. The manpages have some ambiguity about if the standard aio interface is built on top of this or still the thread-based implementation in glibc, and I haven't tried it to find out. According to the respective kqueue manpages, it seems like FreeBSD may also have a suitable aio implementation, but not OS X. So it'd be possible to build a pleasantly abstract libevent api for asynchronous pread(), pwrite(), fsync(), fdatasync() operations which is implementable on Linux and FreeBSD (kqueue only), and on top of that build a (more limited) bufferevent implementation as well if you like. Anything else is missing, though - open(), close(), fstat(), ftruncate(), sync_file_range(), etc. just have the blocking versions, which is a shame. -- Scott Lamb <http://www.slamb.org/> *********************************************************************** To unsubscribe, send an e-mail to majord...@freehaven.net with unsubscribe libevent-users in the body.
