Re: [PATCH 0/2] RFC: readd fair sleepers for server systems

[Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com>]

On 05/23/2012 01:49 PM, Peter Zijlstra wrote:
> On Wed, 2012-05-23 at 13:32 +0200, Christian Ehrhardt wrote:
>>> Why is this, is this some weird interaction with your hypervisor?
>>
>> It is not completely analyzed, as soon as debugging goes out of Linux it
>> can be kind of complex even internally.
>
> Is there significant steal time in these workloads? If so, does it help
> if you implement
> CONFIG_PARAVIRT_TIME_ACCOUNTING/paravirt_steal_rq_enabled for s390?
> (although I guess we'd better loose the paravirt part of the name then).

Interesting, yeah there is enough steal time - not in all, but in most 
cases we had in conflict with fair sleepers so far.
We don't have any code for CONFIG_PARAVIRT and its childs yet, so I need 
to look further into it.

> This 'feature' subtracts steal time from the task-clock so that the
> scheduler doesn't consider a task to be running when the vcpu wasn't
> running as well.
>
> Not doing that (current situation) could result in over-active
> preemption because we think a task ran significantly longer than it
> actually did. Same for sleeper fairness, we might think a task slept
> very long (and give a bigger boost) when in fact it didn't.

Great - sounds like a good thing to check, I'll definitely try this out.
This week we are changing our automation environment, so give me a few 
days for numbers on that.

-- 

Grüsse / regards, Christian Ehrhardt
IBM Linux Technology Center, System z Linux Performance