[pve-devel] [PATCH pve-storage 0/2] move qemu_img_create to common helpers and enable preallocation on backed images

[Fiona Ebner]

Am 22.05.25 um 15:53 schrieb Alexandre Derumier via pve-devel:
> This is part of my work on qcow2 external snapshot, but could improve current qcow2 linked clone
> 
> This patch serie move qemu_img_create to common helpers,
> and enable preallocation on backed_image to increase performance
> 
> This require l2_extended=on on the backed image
>
> I don't have enabled it for base image, as I think that Fabian see performance regression some month ago.
> but I don't see performance difference in my bench. (can you could test on your side again ?)
> 
> It could help to reduce qcow2 overhead on disk,
> and allow to keep more metadatas in memory for bigger image, as qemu default memory l2_cache_size=1MB)
> https://www.ibm.com/products/tutorials/how-to-tune-qemu-l2-cache-size-and-qcow2-cluster-size
> Maybe more test with bigger image (>1TB) could be done too to see if it's help
> 
> I have done some tests with suballocated cluster and base image without
> backing_file, indeed, I'm seeing a small performance degradation on big
> 1TB image.
> 
> with a 30GB image, I'm around 22000 iops 4k randwrite/randread  (with
> or without l2_extended=on)
> 
> with a 1TB image, the result is different
> 
> 
> fio –filename=/dev/sdb –direct=1 –rw=randwrite –bs=4k –iodepth=32
> –ioengine=libaio –name=test
> 
> default l2-cache-size (32MB) , extended_l2=off, cluster_size=64k : 2700 iops
> default l2-cache-size (32MB) , extended_l2=on, cluster_size=128k: 1500 iops

It was not Fabian but me, who reported the regression regarding read
performance and performance for initial allocation back then:
https://lore.proxmox.com/pve-devel/d5e11d01-f54e-4dd9-b1c0-a02077a0c65f@proxmox.com/
The space usage on the underlying storage is greatly improved however.

> I have also play with qemu l2-cache-size option of drive (default value
> is 32MB, and it's not enough for a 1TB image to keep all metadatas in
> memory)
> https://github.com/qemu/qemu/commit/80668d0fb735f0839a46278a7d42116089b82816
> 
> 
> l2-cache-size=8MB , extended_l2=off, cluster_size=64k: 2900 iops
> l2-cache-size=64MB , extended_l2=off, cluster_size=64k: 5100 iops
> l2-cache-size=128MB , extended_l2=off, cluster_size=64k : 22000 iops
> 
> l2-cache-size=8MB , extended_l2=on, cluster_size=128k: 2000 iops
> l2-cache-size=64MB , extended_l2=on, cluster_size=128k: 4500 iops
> l2-cache-size=128MB , extended_l2=on, cluster_size=128k: 22000 iops
> 
> 
> So no difference in needed memory, with or with extended_l2.
> 
> but the l2-cache-size tuning is really something we should add in
> another patch I think ,for general performance with qcow2.

If we want to enable extended_l2=on, cluster_size=128k by default for
all new qcow2 image, I think we should do it together with an increased
l2-cache-size then. But yes, should be its own patch. The above results
sound promising, but we'll need to test a bigger variety of workloads.
If we don't find settings that improve most workloads, we can still make
it configurable.