On Wed, Nov 28, 2012 at 08:43:24AM -0800, Kent Overstreet wrote:

Bunch of performance improvements and cleanups Zach Brown and I have
been working on. The code should be pretty solid at this point, though
it could of course use more review and testing.

Thanks for sending these out.  I have some initial review comments
that'll follow, but I'm running out of steam today.  I'll continue
tomorrow.

The results in my testing are pretty impressive, particularly when an
ioctx is being shared between multiple threads. In my crappy synthetic
benchmark, with 4 threads submitting and one thread reaping completions,
I saw overhead in the aio code go from ~50% (mostly ioctx lock
contention) to low single digits. Performance with ioctx per thread
improved too, but I'd have to rerun those benchmarks.

You should probably mention that those four threads were *spinning* on
io_submit() :).  I'm still guessing that this unreasonably inflated the
contention amongst submitters and that without this inflation we might
not find the per-cpu ioctx refcounts worth the trouble.

Performance wise, the end result of this patch series is that submitting
a kiocb writes to _no_ shared cachelines - the penalty for sharing an
ioctx is gone there. There's still going to be some cacheline contention
when we deliver the completions to the aio ringbuffer (at least if you
have interrupts being delivered on multiple cores, which for high end
stuff you do) but I have a couple more patches not in this series that
implement coalescing for that (by taking advantage of interrupt
coalescing). With that, there's basically no bottlenecks or performance
issues to speak of in the aio code.

Yeah, this is good stuff.  Thanks for pushing it.

We should mention Jens' omnibus patch that also took on these problems:

  http://git.kernel.dk/?p=linux-block.git;a=commit;h=6b6723fc3e4f24dbd80526df935ca115ead578c6

- z