Re: [Qemu-block] Performance impact of the qcow2 overlap checks

[Max Reitz]

On 19.01.2017 15:18, Alberto Garcia wrote:
> On Wed 18 Jan 2017 04:30:49 PM CET, Max Reitz wrote:
> 
>>> In other words: the vast majority of the entries in the refcount
>>> table are probably not going to be used ever, but we're still
>>> checking them for each write request. One user reported a >200%
>>> performance increase on a fast SSD when using overlap-check=constant.
>>>
>>> I think this is at least worth documenting a bit better (unless
>>> there's existing documentation that I have missed), but my main
>>> question is: does it make sense to try to optimize these checks, or
>>> is it better to simply tell the user to disable them in these
>>> scenarios?
>>
>> I have optimized these checks. See:
>>
>> http://lists.nongnu.org/archive/html/qemu-block/2015-08/msg00020.html
>>
>> Feel free to review. If you want, I can rebase the series.
>>
>> So far nobody has seriously done so because it wasn't considered
>> important enough, and the patches are not exactly trivial.
> 
> I see. I wonder if it's worth complicating the code so much more instead
> of simply disabling the tests.
> 
> For the 'refcount-block' check, which is the one that has an immediate
> impact in all kinds of images, I was wondering if we could simply use
> the image size to determine how many entries need to be checked. This
> would keep things much faster unless the image grows abnormally bigger.

Have you tested that it helps? I would assume that checking one (mostly)
unused cluster of the refcount table does not take longer than checking
a single refcount block. Any entry that's 0 will just be skipped.

Max

signature.asc
Description: OpenPGP digital signature