[PATCH 0/4] CGroup: Support for named and empty hierarchies
kamezawa.hiroyu at jp.fujitsu.com
Wed Jul 22 22:45:20 PDT 2009
On Wed, 22 Jul 2009 22:37:53 -0700
Paul Menage <menage at google.com> wrote:
> On Wed, Jul 22, 2009 at 10:13 PM, KAMEZAWA
> Hiroyuki<kamezawa.hiroyu at jp.fujitsu.com> wrote:
> > Thank you. I'd like to look into procps package to allow
> > #ps -lf -cgroup=/xxxx/yyy
> > or
> > #top -cgroup=/xxx/yyy
> Wouldn't it be less invasive to change procps so that doing
> ps <pid1> <pid2> ...
> only actually looks at /proc/<pid1>, /proc/<pid2>, etc, rather than
> looking at all of them and then throwing away any that aren't in the
> supplied list?
> Then you could do what you want with ps $(cat /dev/cgroup/xxxx/yyy/tasks)
Good point. When you try strace
#strace ps -ef --pid XXXX
you'll see ps scan all pids. The same behavior for pkill, pgrep etc...
AFAIK, this is a behavior from Linux 2.4 ages.
By this, ps -ef --pid=XXX can take 1ms if 2000 procs are on system.
# I think this is because ps uses an unified filter routine.
So, I have to rewrite filter routine or add cgroup support.
Abyway, adding a new toggle switch to "top", "pgrep" is worth to do, I think.
More information about the Containers