native libaio

[Volker Lendecke]

On Sun, Jun 13, 2010 at 05:37:24PM +0200, Volker Lendecke wrote:
> In my wip branch on gitweb.samba.org I've got code to call
> libaio (using native linux aio) for read&x. I've got some
> numbers for reading 1GB (out of cache) with smbclient on TCP
> loopback on a 4-core system. User/sys times are smbd -d 0 -i
> times.
> 
> sync I/O:
> user    0m0.072s
> sys     0m0.716s
> 817284,4 MByte/s
> 
> libaio async I/O:
> user    0m0.124s
> sys     0m0.796s
> 742617,5 MByte/s
> 
> posix AIO:
> user    0m0.300s
> sys     0m1.600s
> 461724,3 MByte/s
> 
> This means the native linux aio is in fact a hell lot faster
> than the glibc implementation :-)

Crap. Bjoern just pointed out that all this might be a pure
fallacy. In http://axboe.livejournal.com/1718.html you can
read that libaio is only asynch if we use DIRECTIO. That
sucks. Severely. I've done some benchmarking with the pure
clone-based implementation in libaio, and to me it seems
that the context switches between the main and the worker
threads really kill the performance. Well, what can you do.

Volker