Re: busted CFS group load balancer?

[Peter Zijlstra <a.p.zijlstra@chello.nl>]

On Mon, 2008-11-17 at 15:19 -0600, Chris Friesen wrote:
> Ken Chen wrote:
> > On Mon, Nov 17, 2008 at 7:37 AM, Chris Friesen wrote:
> >>> It appears that the fair-group load balancer in 2.6.27 does not work
> >>> properly.
> >> There was an issue fixed post 2.6.27 where the load balancer didn't work
> >> properly if there was one task per group per cpu.  You might try
> >> backporting commit 38736f4 and see if that helps.
> > 
> > Tested git commit 38736f4, it doesn't fix the problem I'm seeing.
> > 
> 
> I plugged in the same weights into my test app (groups 1 and 2 instead 
> of ant/bee) and got the results below for a 10-sec run.  The "actual" 
> numbers give the overall average and then the values for each hog 
> separately.  In this case we see that both tasks in group 2 ended up 
> sharing a cpu with one of the tasks from group 1.
> 
>    group      actual(%)      expected(%)  ctx switches   max_latency(ms)
>        1  99.69(99.38/99.99)   99.81       160/262          4/0
>        2   0.31( 0.31/0.31)     0.19       32/33         391/375
> 
> I've only got a 2-way system.  If the results really are that much worse 
> on larger systems, then that's going to cause problems for us as well. 
> I'll see if I can get some time on a bigger machine.

Note that with larger cpu count and/or lower group weight we'll quickly
run into numerical trouble...

I would recommend trying this with the minimum weight in the order of
8-16 times number of cpus on your system.

There is only so much one can do with 10 bit fixed precision math :/