[pve-devel] qemu 6 : does proxmox backup handle parallel async chunck backup like the new backup code from qemu 6.0 ?

[aderumier at odiso.com]

Le mercredi 02 juin 2021 à 08:58 +0200, Thomas Lamprecht a écrit :
> Hi,
> 
> On 02.06.21 08:39, aderumier at odiso.com wrote:
> > I was looking for qemu 6.0 new features,
> > and it seem that they have implement parallel async chunks backup
> > (and
> > I think for other block operations, through a new block-copy
> > feature)
> > 
> > https://github.com/qemu/qemu/commit/71eed4cebed487a4f3c9f97aba83c611bbe22f8d
> > 
> > https://github.com/qemu/qemu/commit/de4641b46b020c5b332175f80e8bfe3d352888e8#diff-b33323044f2699244c126c6eae6c4083c3c99a16f4840030ac13238b1f569dc0
> > 
> > https://github.com/qemu/qemu/commit/26be9d62dd5f5268b814da24fd8e8b5c5b999ebe
> > 
> > Is it alrealy implemented or on the roadmap for the proxmox backup
> > code
> > ?
> > (To be honest, I didn't follow proxmox backup patches since a long
> > time, so I really don't known how much they are sharing with qemu
> > backup code)
> 
> the parallel stuff works with our implementation, Stefan tried it and
> with their
> default the backup is much faster, so fast that the guest cannot do
> anything during
> that anymore :D So, IIRC, Stefan turned the default number of
> parallel operations down
> for a better balance of backup speed and guest impact.
> 
ok thanks ! (I'll still don't use pbs, I'll try to test with my rbd
cluster soon to compare speed vs rbd export/import)

> FYI, we have basic packages for bullseye as base ready here, and
> possible sync to
> pvetest in the next days/weeks - in those repos there's already QEMU
> 6.0.
> 
yes, I see that.I'll try to do test soon

> Something else we can use with Bullseye/PVE 7 is io_uring instead of
> AIO, looks like
> a much saner (and actually useful) design and there should be some,
> at least slight,
> improvements due to that too.
> 
> https://kernel.dk/io_uring.pdf
> 
yes, If I remember Wolfgang had seen some bug with kernel 5.4, but I
think it's should works fine with 5.11.
I'll try it too soon :)


> cheers,
> Thomas
>