在 2021/6/19 0:00, Peter Xu 写道:
On Fri, Jun 18, 2021 at 11:32:01PM +0800, huang...@chinatelecom.cn wrote:
From: Hyman Huang(黄勇) <huang...@chinatelecom.cn>
v8
I have one last comment in patch 2, the rest looks good to me, thanks.
I think the next step can be that we add another command to configure dirty
rate limit for each of the vcpu. It can be the 3rd bit in the
global_dirty_tracking and it should require dirty ring too to be there, but I
haven't thought deeper.. Then after that, we may teach migration code some
algorithm to auto-setup these per-vcpu throttling, it'll grow into some kind of
per-vcpu auto-converge.
Do you think above would be something worth trying out? And.. Dave/Paolo/Juan?
from my point, absolutely yes, since we have introduced the dirty ring,
the most valuable sicenario i think is to improve migration or at least
provide another auto-converge implementation to management apps, it
seems that there is no reason not trying this.
Btw, this series cannot apply cleanly to master here. Maybe you need to pull
and rebase?
Ok, i'll rebase and post another version later, may be commit "KVM: Fix
dirty ring mmap incorrect size due to renaming accident" should be
merged before this, so that dirty rate calculation can work rightly.
