Performance on QEMU IDE disks

Performance on QEMU IDE disks

[P M, Priya (STSD)]

[-- Attachment #1.1: Type: text/plain, Size: 781 bytes --]

Hi,
 
We have Intel VT enabled system loaded with Xen 3.0.2 and qemu based
FC5+linux 2.6.16 as a guest operating system. 
We did some performance tests on this guest OS and the results are
 
Sequential Write - IOPs
IO size

512B

4K

16K

32K

FC5 Native

162

163

156

147

FC5_DOM1(Linux 2.6.16 - Full Virtualization) - qemu based

1305

742

285

155

 
We have disabled the Write Cache using sdparm and hdparm utilities. But
if you see the results in Linux Full Virtualization, the Sequential
Write IOPs are better than Native performance. I am wondering how is
this possible? Is there anywhere the caching is happening in the full
virtualization case? If you have an idea about this behavior, please let
me know.
 
Thanks
Priya

[-- Attachment #1.2: Type: text/html, Size: 4436 bytes --]

[-- Attachment #2: Type: text/plain, Size: 138 bytes --]

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel

Re: Performance on QEMU IDE disks

[Steve Ofsthun]

P M, Priya (STSD) wrote:

> We have disabled the Write Cache using sdparm and hdparm utilities. But 
> if you see the results in Linux Full Virtualization, the Sequential 
> Write IOPs are better than Native performance. I am wondering how is 
> this possible? Is there anywhere the caching is happening in the full 
> virtualization case? If you have an idea about this behavior, please let 
> me know.

QEMU does indeed use the buffer cache in Domain 0.  In doing so, it also
will take advantage of Domain 0 read ahead/write behind.  The down side is
that the disk write ordering guarantees expected by DomU filesystems are
violated as well.  If XEN or Domain 0 crashes, your DomU filesystems may
be toast.

To fix this you would need to patch QEMU to used O_DIRECT when accessing
the virtual disk backing store object (block device, file, etc).  We are
currently testing just such a patch for the old QEMU device model.  I
haven't looked at the new device model to see if it already handles this.

Steve
-- 
Steve Ofsthun - Virtual Iron Software, Inc.

Re: Performance on QEMU IDE disks

[Stephen Tweedie]

Hi,

On Fri, Jul 14, 2006 at 03:41:23PM -0400, Steve Ofsthun wrote:

> To fix this you would need to patch QEMU to used O_DIRECT when accessing
> the virtual disk backing store object (block device, file, etc).  We are
> currently testing just such a patch for the old QEMU device model.

Thanks!

>  I
> haven't looked at the new device model to see if it already handles this.

I looked last week and it still had the same problem.

Note that O_DIRECT doesn't solve all of the problems either.  You
still end up with the _entire_ device model being blocked while the
disk IO is in progress, which will leads to poor VT guest performance
when trying to mix disk and CPU activity.  But I'd rather have a
performance problem than a data corrupter. :)

--Stephen