On 2018-04-17 14:37, Alberto Garcia wrote:
> Hi,
>
> we talked about this the other day, so here are the patches to change
> the default cache sizes in qcow2.
>
> Without this patch:
>
> * refcount-cache-size = l2-cache-size / 4
>
> unless otherwise specified by the user. This is wasteful, the refcount
> cache is accessed sequentially during normal I/O, so there's no point
> in caching more tables. I measured the effect on the refcount cache
> size when populating an empty qcow2 image using random writes, and
> there's no difference between having the minimum or the maximum
> sizes(*).
>
> With this patch:
>
> * refcount-cache-size is always 4 clusters by default (the minimum)
>
> * If "cache-size" is set then l2-cache-size is set to the maximum if
> possible (disk_size * 8 / cluster_size) and the remainder is
> assigned to the refcount cache.
>
> Regards,
>
> Berto
>
> (*) there is, actually: having a very large cache can even make the
> I/O slightly slower, because the larger the cache the longer it
> takes longer to find a cached entry. I only noticed this under
> tmpfs anyway.
>
> Changes:
> v3:
> - Mention that if you use internal snapshots you may want to increase
> the cache size [Max]
>
> v2: https://lists.gnu.org/archive/html/qemu-block/2018-03/msg00822.html
> - s/overriden/overridden/ (in both patches)
>
> v1: https://lists.gnu.org/archive/html/qemu-block/2018-03/msg00709.html
> - Initial release
>
> Alberto Garcia (2):
> qcow2: Give the refcount cache the minimum possible size by default
> docs: Document the new default sizes of the qcow2 caches
>
> block/qcow2.c | 31 +++++++++++++++++++------------
> block/qcow2.h | 4 ----
> docs/qcow2-cache.txt | 33 ++++++++++++++++-----------------
> tests/qemu-iotests/137.out | 2 +-
> 4 files changed, 36 insertions(+), 34 deletions(-)
Thanks, applied to my block-next branch:
https://github.com/XanClic/qemu/commits/block-next
Max