Re: [Qemu-block] [PATCH] qcow2: Optimize L2 table cache size based on image and cluster sizes

[Kevin Wolf]

Am 05.10.2016 um 17:13 hat Max Reitz geschrieben:
> On 05.10.2016 16:57, Alberto Garcia wrote:
> > On Tue 04 Oct 2016 05:51:26 PM CEST, Max Reitz wrote:
> > 
> >>> At least giving users a way to skip the math would be an improvement.
> >>> Would you be okay with an explicitly-set option like
> >>> l2_cache_size=auto or =max that optimizes for performance at the
> >>> expense of memory?
> >>
> >> That (cache-size=max) is actually something Stefan Hajnoczi has
> >> proposed at KVM Forum 2015.
> >>
> >> I agree that implementing the formula in qemu's qcow2 driver itself
> >> makes sense and will help users; however, I do think it is appropriate
> >> to expect the user to pass cache-size=max if they need it.
> > 
> > Frank Myhr's suggestion (in bugzilla) is that we allow specifying a % of
> > the disk size, so
> > 
> > l2-cache-size=100%  (that would be cache-size=max)
> > l2-cache-size=80%
> > ...
> > 
> > https://bugzilla.redhat.com/show_bug.cgi?id=1377735#c3
> > 
> > Either one looks good to me. They would change however the data type of
> > 'l2-cache-size' from int to string in BlockdevOptionsQcow2, can we
> > actually do that?
> 
> I think we can do that with an alternate data type which accepts both an
> integer and a string. If we only had "max", we'd probably want to make
> it an enumeration instead of a free-form string, though. But with a %
> suffix, we'd probably need a string.
> 
> For me, either works fine.

I'm not sure if we want to support such syntax in QMP. It sounds like a
typical convenience option for human users.

> Apart from that, I have to say I think it would be a bit more useful if
> one would specify the area covered by the metadata caches as an absolute
> number instead of a relative one (I guess it's generally easier to know
> what area your applications will perform random accesses on than the
> relative size, but maybe that's just me).
> 
> Supporting that with cache-size is difficult, though, because how are
> you going to decide between whether the user is specifying the actual
> size of the cache or the area it covers?
> 
> So maybe it would make more sense to introduce a whole new set of
> options called {,l2-,refcount-}cache-coverage or something. These
> options would then accept both an absolute and a relative number.

If we want easy, make it easy: cache-coverage and apply it to both.

Kevin

pgp5WuQZzznEs.pgp
Description: PGP signature