[Qemu-devel] qemu (cvs-version) performance?

[Qemu-devel] qemu (cvs-version) performance?

[Sven Köhler]

Hi,

i'd like to know, what the qemu-project aims at.
Do the authers aim at writing (together with kqemu or qv86) an
alternative to VMWare? (At least in the x86 virtualization case)

I first thought that way and i still hope that, but i was disappointed
by qemu 0.8.0 and the latest version from CVS.

I thought, that qemu-0.8.0's problem would be disk-IO. I installed
Windows 2000 and in the task-manager the CPU-Time is 100% during intense
IO. A "dir c:\ /A /S /B >NUL:" on the console causes 100% CPU-time and
most of it being that red-colored "in-kernel-time".

The DMA-mode is active according to the device-manager, since i upgraded
to qemu CVS (version from an hour ago).

So IO still seems to be very expensive. AFAIK, IDE-DMA or BusMastering
should enable the CPU to do other things while waiting for the Interrupt
that signals the finished transfer. (My knowledge about hardware is not
that good though.) But the virtual CPU is not idle and the host's CPU
isn't either.

Of course i'm using kqemu.


So what might be wrong?
Some hints:
- the disk-image(raw format) is on a reiserfs-filesystem
- the host ist kernel 2.6.15


Please don't understand me wrong.
I appreciate your work.


Greetings
  Sven

Re: [Qemu-devel] qemu (cvs-version) performance?

[Paul Brook]

> I thought, that qemu-0.8.0's problem would be disk-IO. I installed
> Windows 2000 and in the task-manager the CPU-Time is 100% during intense
> IO. A "dir c:\ /A /S /B >NUL:" on the console causes 100% CPU-time and
> most of it being that red-colored "in-kernel-time".

Qemu host IO is syncrhonous. ie. the guest is paused on the instruction that 
issues the IO until the IO completes. As for as the guest is concerned it 
happens instantly. This is all your test shows.

Paul

[Qemu-devel] Re: qemu (cvs-version) performance?

[Sven Köhler]

>> I thought, that qemu-0.8.0's problem would be disk-IO. I installed
>> Windows 2000 and in the task-manager the CPU-Time is 100% during intense
>> IO. A "dir c:\ /A /S /B >NUL:" on the console causes 100% CPU-time and
>> most of it being that red-colored "in-kernel-time".
> 
> Qemu host IO is syncrhonous. ie. the guest is paused on the instruction that 
> issues the IO until the IO completes. As for as the guest is concerned it 
> happens instantly. This is all your test shows.

I have seen thoughts about asynchronous IO for qemu. I thought that they
would have been integrated along with the DMA-patches already.

I would really see how qemu performs with asynch IO enabled. Are there
any patches out there?


Greetings
  Sven

[Qemu-devel] Re: qemu (cvs-version) performance?

[Sven Köhler]

> I have seen thoughts about asynchronous IO for qemu. I thought that they
> would have been integrated along with the DMA-patches already.

I'm sorry, i read a mail by Fabrice that async block I/O is not merged yet.

Re: [Qemu-devel] Re: qemu (cvs-version) performance?

[Jim C. Brown]

On Sat, Mar 11, 2006 at 11:35:48PM +0100, Sven K?hler wrote:
> I have seen thoughts about asynchronous IO for qemu. I thought that they
> would have been integrated along with the DMA-patches already.
> 
> I would really see how qemu performs with asynch IO enabled. Are there
> any patches out there?
> 

There was one that used threads to do IO asychonrously, but iiuc the final
patches (which will use posix? async io) are still being worked on..

You can find the patches in the mailing list archives.

> 
> Greetings
>   Sven
> 
> 
> 
> _______________________________________________
> Qemu-devel mailing list
> Qemu-devel@nongnu.org
> http://lists.nongnu.org/mailman/listinfo/qemu-devel
> 

-- 
Infinite complexity begets infinite beauty.
Infinite precision begets infinite perfection.