Re: [Qemu-devel] Qemu speed vs vmplayer?

[Joseph Miller]

Paul Brook wrote:
> > Does anyone know the reason for the removal of the mmap()? I have used 
> > a benchmarking tool (I think it was 3D Mark05 or 3D Mark06) and the
> > memory access in the guest WinXP was slooooooow.  Does anyone have any
> > insight on making the hardware MMU function for linux-x86 host to WinXP
> > 32-bit guest, even partially? 
> >     
>
> As I already said, the code required a modified guest OS. i.e. it wouldn't 
> work with windows guests anyway.
>
> It was removed because it provided very little benefit, and wasn't worth the 
> effort to keep it working.  Its uses were very specialised, and IMHO better 
> covered by other products like UML or xen.
>
> I'm also doubtful how much benefit it gave in practice. I'm sure it would be 
> good for synthetic CPU benchmarks. However using mmap significantly increases 
> the overhead of context switches/tlb misses.
>
> To get good overall performance I suspect you're going to need closer 
> cooperation with the kernel than mmap gives you. If you really want to make 
> cross-emulation go fast I suggest working with the xen and/or kvm people to 
> integrate qemu dynamic translation into those products. In theory I'd guess 
> you should be able to plug it in as an alternative to the HVM code. I've no 
> idea how close that is to being practical.
>
>   
> > Wouldn't this be a *significant* 
> > performance enhancement for this setup which I'm sure is a common one?
> > Maybe this can be implemented for regular processes on the guest and
> > only use the softmmu for the kernel?  Would someone point me in the
> > right direction for technical information?  I have had to switch to
> > vmware free player until Qemu+KQEMU reaches a point of similar
> > performance, but I would really rather see Qemu advance further.
> >     
>
> If you're using an accelerator (eg. kqemu or kvm) this is all irelevant as 
> most code isn't run by qemu, it's virtualized by the accelerator. qemu just 
> does the IO emulation.
>
> Paul
>   
OK, so mmap is not the way to increase some speed.  What needs to be 
done to provide a higher Qemu+KQEMU performance, comparable to 
VMPlayer?  How does VMPlayer manage to be so much faster than Qemu?  Is 
this simply an I/O bottleneck?  How would I go about finding out what 
the differences are and how we can improve Qemu+KQEMU performance?

-Joseph