Re: [Qemu-devel] [PATCH v3 0/8] eliminate data write in bdrv_write_zeroes on Linux in raw-posix.c

[Peter Lieven <pl@kamp.de>]

On 30.12.2014 12:07, Denis V. Lunev wrote:
> On 30/12/14 13:55, Peter Lieven wrote:
>> Am 30.12.2014 um 10:20 schrieb Denis V. Lunev:
>>> These patches eliminate data writes completely on Linux if fallocate
>>> FALLOC_FL_ZERO_RANGE or FALLOC_FL_PUNCH_HOLE are  supported on
>>> underlying filesystem.
>>>
>>> I have performed several tests with non-aligned fallocate calls and
>>> in all cases (with non-aligned fallocates) Linux performs fine, i.e.
>>> areas are zeroed correctly. Checks were made on
>>>     Linux 3.16.0-28-generic #38-Ubuntu SMP
>>>
>>> This should seriously increase performance in some special cases.
>> Could you give a hint what that special cases are? It would help
>> to evaluate and test the performance difference.
>>
>> Thanks,
>> Peter
>
> - 15% in Parallels Image expansion, see my side patchset

I will have a look.

> - writing zeroes to raw image with BLOCKDEV_DETECT_ZEROES_OPTIONS_ON set
>    (actually I have kludged raw-posix.c to have this flag always set
>    to perform independent testing)

Is there a valid reason to write chunks of >16MB blocksize from a guest? Or is this
just users performing pseudo benchmarks with dd?

Peter