Re: [Qemu-devel] [PATCH] CPU consumption optimization of 'qemu-img convert' using bdrv_is_allocated()

[Stefan Hajnoczi]

On Wed, Jun 15, 2011 at 2:14 PM, Dmitry Konishchev <address@hidden> wrote:
> On Wed, Jun 15, 2011 at 4:02 PM, Stefan Hajnoczi <address@hidden> wrote:
>> We need to fully understand performance before applying optimizations
>> on top.  Otherwise it is possible to paper over a problem while
>> leaving the root cause unsolved.  Avoiding lseek(2) is very important,
>> not just for qemu-img but also for running VMs.  lseek(2) should only
>> be invoked if the image is growable/removable.
>>
>> When I run a VM from a virtio-blk raw image I see no lseek(2) calls.
>>
>> On the host:
>> strace -p $pid_of_qemu -f
>>
>> Inside the guest:
>> dd if=/dev/vda of=/dev/null iflag=direct
>>
>> I see pread(2) from the posix-aio-compat.c worker threads but no
>> lseek(2).  The total_sectors cached value is being used.
>>
>> Does strace(1) show lseek(2) on your host?
>
> No, I don't see lseek() in the strace output. But in the other hand I
> see that bdrv_get_geometry() uses bdrv_getlength() which is
> implemented using lseek() in block/raw-posix.c...
> May be I'am mistaken about lseek(), but I get 9% slower version if
> disable caching.

"disable caching"?

Stefan