Re: [PATCH] mm: Free per cpu pages async to shorten program exit time

[Michal Hocko <mhocko@suse.com>]

On Mon 11-10-21 11:40:12, David Hildenbrand wrote:
> On 11.10.21 11:28, Michal Hocko wrote:
> > On Fri 08-10-21 10:17:50, David Hildenbrand wrote:
> > > On 08.10.21 08:39, ultrachin@163.com wrote:
> > > > From: chen xiaoguang <xiaoggchen@tencent.com>
> > > > 
> > > > The exit time is long when program allocated big memory and
> > > > the most time consuming part is free memory which takes 99.9%
> > > > of the total exit time. By using async free we can save 25% of
> > > > exit time.
> > > > 
> > > > Signed-off-by: chen xiaoguang <xiaoggchen@tencent.com>
> > > > Signed-off-by: zeng jingxiang <linuszeng@tencent.com>
> > > > Signed-off-by: lu yihui <yihuilu@tencent.com>
> > > 
> > > I recently discussed with Claudio if it would be possible to tear down the
> > > process MM deferred, because for some use cases (secure/encrypted
> > > virtualization, very large mmaps) tearing down the page tables is already
> > > the much more expensive operation.
> > > 
> > > There is mmdrop_async(), and I wondered if one could reuse that concept when
> > > tearing down a process -- I didn't look into feasibility, however, so it's
> > > just some very rough idea.
> > 
> > This is not a new problem. Large process tear down can take ages. The
> > primary road block has been accounting. This lot of work has to be
> > accounted to the proper domain (e.g. cpu cgroup).
> 
> In general, yes. For some setups where admins don't care about that
> accounting (e.g., enabled via some magic toggle for large VMs), I guess this
> accounting isn't the major roadblock, correct?

Right, I would be careful about magic toggles though. Besides there are
ways to achive this in the userspace. We used to have a request to help
paralleling process exit from a DB vendor and Vlastimil has come up with
a clone(CLONE_VM) and madvise(DONT_NEED) from several threads as a
"workaround". This would work properly from the accounting POV.
Admittedly a bit of an involved approach though.
-- 
Michal Hocko
SUSE Labs