Re: [Qemu-devel] Anyone seeing huge slowdown launching qemu with Linux 2.6.35?

[Gleb Natapov <gleb@redhat.com>]

On Wed, Aug 04, 2010 at 05:59:40PM +0200, Alexander Graf wrote:
> 
> On 04.08.2010, at 17:48, Gleb Natapov wrote:
> 
> > On Wed, Aug 04, 2010 at 05:31:12PM +0200, Alexander Graf wrote:
> >> 
> >> On 04.08.2010, at 17:25, Gleb Natapov wrote:
> >> 
> >>> On Wed, Aug 04, 2010 at 09:57:17AM -0500, Anthony Liguori wrote:
> >>>> On 08/04/2010 09:51 AM, David S. Ahern wrote:
> >>>>> 
> >>>>> On 08/03/10 12:43, Avi Kivity wrote:
> >>>>>> libguestfs does not depend on an x86 architectural feature.
> >>>>>> qemu-system-x86_64 emulates a PC, and PCs don't have -kernel.  We should
> >>>>>> discourage people from depending on this interface for production use.
> >>>>> That is a feature of qemu - and an important one to me as well. Why
> >>>>> should it be discouraged? You end up at the same place -- a running
> >>>>> kernel and in-ram filesystem; why require going through a bootloader
> >>>>> just because the hardware case needs it?
> >>>> 
> >>>> It's smoke and mirrors.  We're still providing a boot loader it's
> >>>> just a little tiny one that we've written soley for this purpose.
> >>>> 
> >>>> And it works fine for production use.  The question is whether we
> >>>> ought to be aggressively optimizing it for large initrd sizes.  To
> >>>> be honest, after a lot of discussion of possibilities, I've come to
> >>>> the conclusion that it's just not worth it.
> >>>> 
> >>>> There are better ways like using string I/O and optimizing the PIO
> >>>> path in the kernel.  That should cut down the 1s slow down with a
> >>>> 100MB initrd by a bit.  But honestly, shaving a couple hundred ms
> >>>> further off the initrd load is just not worth it using the current
> >>>> model.
> >>>> 
> >>> The slow down is not 1s any more. String PIO emulation had many bugs
> >>> that were fixed in 2.6.35. I verified how much time it took to load 100M
> >>> via fw_cfg interface on older kernel and on 2.6.35. On older kernels on
> >>> my machine it took ~2-3 second on 2.6.35 it took 26s. Some optimizations
> >>> that was already committed make it 20s. I have some code prototype that
> >>> makes it 11s. I don't see how we can get below that, surely not back to
> >>> ~2-3sec.
> >> 
> >> What exactly is the reason for the slowdown? It can't be only boundary and permission checks, right?
> >> 
> >> 
> > The big part of slowdown right now is that write into memory is done
> > for each byte. It means for each byte we call kvm_write_guest() and
> > kvm_mmu_pte_write(). The second call is needed in case memory, instruction
> > is trying to write to, is shadowed. Previously we didn't checked for
> > that at all. This can be mitigated by introducing write cache and do
> > combined writes into the memory and unshadow the page if there is more
> > then one write into it. This optimization saves ~10secs. Currently string
> 
> Ok, so you tackled that bit already.
> 
> > emulation enter guest from time to time to check if event injection is
> > needed and read from userspace is done in 1K chunks, not 4K like it was,
> > but when I made reads to be 4K and disabled guest reentry I haven't seen
> > any speed improvements worth talking about.
> 
> So what are we wasting those 10 seconds on then? Does perf tell you anything useful?
> 
Not 10, but 7-8 seconds.

After applying cache fix nothing definite as far as I remember (I ran it last time
almost 2 week ago, need to rerun). Code always go through emulator now
and check direction flags to update SI/DI accordingly. Emulator is a big
switch and it calls various callbacks that may also slow things down.

--
			Gleb.