Re: [ANNOUNCE] Native Linux KVM tool v2

[Ingo Molnar <mingo@elte.hu>]

* Pekka Enberg <penberg@kernel.org> wrote:

> Hi Ingo,
> 
> On Thu, Jun 16, 2011 at 10:24 AM, Ingo Molnar <mingo@elte.hu> wrote:
> >  - executing AIO in the vcpu thread eats up precious vcpu execution
> >   time: combined QCOW2 throughput would be limited by a single
> >   core's performance, and any time spent on QCOW2 processing would
> >   not be spent running the guest CPU. (In such a model we certainly
> >   couldnt do more intelligent, CPU-intense storage solutions like on
> >   the fly compress/decompress of QCOW2 data.)
> 
> Most image formats have optional on-the-fly 
> compression/decompression so we'd need to keep the current I/O 
> thread scheme anyway.

Yeah - although high-performance setups will probably not use that.

> > I'd only consider KAIO it if it provides some *real* measurable 
> > performance advantage of at least 10% in some important usecase. 
> > A few percent probably wouldnt be worth it.
> 
> I've only been following AIO kernel development from the sidelines 
> but I really haven't seen any reports of significant gains over 
> read()/write() from a thread pool. Are there any such reports?

I've measured such gains myself a couple of years ago, using an 
Oracle DB and a well-known OLTP benchmark, on a 64-way system.

I also profiled+tuned the kernel-side AIO implementation to be more 
scalable so i'm reasonably certain that the gains exist, and they 
were above 10%.

So the kaio gains existed back then but they needed sane userspace 
(POSIX AIO with signal notification sucks) and needed a well-tuned 
in-kernel implementation as well. (the current AIO code might have 
bitrotted)

Also, synchronous read()/write() [and scheduler() :-)] scalability 
improvements have not stopped in the past few years so the 
performance picture might have shifted in favor of a thread pool.

Thanks,

	Ingo