[Devel] Re: css_id numbering

[Ryo Tsuruta]

Hi Kamezawa-san,

> On Thu, 21 May 2009 19:36:06 +0900 (JST)
> Ryo Tsuruta <ryov at valinux.co.jp> wrote:
> 
> > Hi Kamezawa-san and Kosaki-san,
> > 
> > > > I have a suggestion for css_id numbering. How about using the same
> > > > numbering scheme as process ID for css_id instead of idr? It prevents
> > > > the same ID from being reused quickly.
> > > 
> > > process ID is not "short"(2bytes).
> > > 
> > > > blkio-cgroup can benefit from it. If some pages still have an old ID,
> > > > it can somewhat prevents IOs associated with those pages from being
> > > > charged for a new cgroup which are reusing the old ID.
> > > 
> > > plz see memcg's code. no reuse if id is still used.
> > 
> > blkio-cgroup embeds css_id into page_cgroup->flags, and the ID in the
> > flags still remain even after the cgroup which marks up the page is
> > removed.
> > 
> > The current numbering scheme of css_id choose the lowest number from
> > available IDs. When a cgroup is created right after another cgroup is
> > removed (no one referes to the cgroup), the same number will be reused
> > for the new cgroup.
> > 
> > When an IO associated with the page is issued, an IO controller which
> > uses blkio-cgroup charges the new cgroup according to the ID in the
> > page_cgroup->flags, instead of the removed cgroup. If css_id is simply
> > incremented by one like process ID, I think we could avoid such wrong
> > charges in many cases.
> > 
> IDR uses structure lika a radix-tree and it's important to keep the range
> as small as possible.
> And scheme like PID will have reuse issue anyway. 99.9% is not 100%.
> If 1023 of IDs are used under total 1024 IDS, you'll see reuse soon.
> please rewrite css_id by yourself if you have better codes than IDR.
> IIUC, it's keeps footprint small.

Thank you for explaining it. I know this solution is not perfect, but
I think it works well in many cases. I think I'll try to create a
patch if I really need this solution.
 
> > > it seems bogus solusion.
> > > pid rule also can make conflict. I think blkio-cgroup should be fixed.
> > 
> > This solution is reasonable for accuracy, even conflicts occur.
> > I think keeping down the overhead of blkio-cgroup is more important.
> > 
> please use private refcnt for avoiding reuse. It's the only way to prevent
> reuse of ID in theory.

Thanks,
Ryo Tsuruta
_______________________________________________
Containers mailing list
Containers at lists.linux-foundation.org
https://lists.linux-foundation.org/mailman/listinfo/containers