On Mon, 2011-09-19 at 15:38 -0300, Glauber Costa wrote:On 09/19/2011 03:35 PM, Peter Zijlstra wrote:On Mon, 2011-09-19 at 13:30 -0300, Glauber Costa wrote:For cpuusage, I am not sure this optimization is a valid one
I was talking about cpuusage, cpuacct_charge() is called for every
ctxsw/tick.
I am not touching it right now.
See hunk #4 of this particular patch:
@@ -9302,7 +9308,7 @@ static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
That's cpuusage muck ;-)
But even for cpuacct tick stuff, wouldn't you need to sum all your child
cgroups to update the current cgroup? and that up the whole tree?
Of course I would. But as I said, it does not need to be done every
tick, in case it poses such a cacheline mayhem as you fear.
Since we'll only really need those values when someone reads it - which
is a far less frequent operation than the tick resolution - and when a
cgroup is destroyed - even less frequent operation - it should work well.
Quite possible, yes. Although if you create a cgroup with 100 subgroups
and poll very frequently.. it all depends on the avg use case etc.. and
since I don't use any of this stuff someone needs to tell me about how
the trade-offs work out in practice.
So explicit changelogs with numbers and agreements from multiple users
go a long way to make me feel good ;-)