[Qemu-devel] testing qemu svn r5890 on FreeBSD - virtio, and a patch enabling -clock dynticks

[Juergen Lock]

In article <493B35AB.1050301 at codemonkey.ws> you write:
>Juergen Lock wrote:
>> Hi!
>>
>>  Jung-uk Kim sent me a patch to enable -clock dynticks on FreeBSD hosts
>> (the configure check is mine, only FreeBSD >= 7.x has posix timers that
>> this uses), I'll append it below.
>>
>>  This is the experimental qemu-devel port update I used:
>> 	http://people.freebsd.org/~nox/qemu/qemu-devel-20081206.patch
>> As already mentioned I had to add a missing `#include <sys/uio.h>'
>> (files/patch-qemu-common.h), as also posted here:
>> 	http://lists.gnu.org/archive/html/qemu-devel/2008-12/msg00216.html
>>
>>  I only had one (type of) guest that actually had virtio drivers (three
>> versions of sidux isos), and the speed difference between virtio-blk and
>> scsi was small.  (I tested dd bs=64k count=500 </dev/vda >/dev/null and
>> similar with a raw image, both scsi and virtio were always faster than ide.)
>> I noted tho that even virtio there was not half as fast as ide (and scsi)
>> on KNOPPIX_V5.3.1DVD-2008-03-26-EN.iso, so either overhead has increased
>> greatly from 2.6.24.4 to 2.6.26, or this has something to do with
>> the sidux kernel using CONFIG_NO_HZ and the Knoppix one (apparently) not
>> and qemu (possibly, I also suspected that with the usb slowness) not
>> handling CONFIG_NO_HZ guests too well.  scsi on a FreeBSD
>> 7.1-BETA-i386-livefs.iso guest btw was even yet (noticeably) faster than
>> on the Knoppix iso.  It will be interesting how virtio-net will fare once
>> that gets committed...
>>   
>
>I don't have much experience with perf benchmarking and TCG.  TCG may 
>has interesting side effects.  For instance, it's more expensive to do 
>things in the guest instead of the host so the emulation overhead of 
>IDE/SCSI shouldn't matter much.
>
Actually most of those tests were with -kernel-kqemu sice that is what
I usually run linux guests with.

>A straight dd test is not the best test BTW.  If you want to measure 
>performance, you should use O_DIRECT in the guest (iflag=direct)

 Hmm, how many guest processes will use that?  Or is that what
fses end up doing when the guest does things like, say, cp'ing files
around?

> and 
>probably O_DIRECT in the host (cache=none).

 Ok, will do next time...

 Thanx,
	Juergen