On Mon, Oct 25, 2010 at 3:12 AM, Peter Zijlstra <peterz@infradead.org> wrote:
quoted text
> On Fri, 2010-10-22 at 16:03 -0700, Venkatesh Pallipadi wrote:
>> I started making small changes to the code, but none of the change helped much.
>> I think the problem with the current code is that, even though idle CPUs
>> update load, the fold only happens when one of the CPU is busy
>> and we end up taking its load into global load.
>>
>> So, I tried to simplify things and doing the updates directly from idle loop.
>> This is only a test patch, and eventually we need to hook it off somewhere
>> else, instead of idle loop and also this is expected work only as x86_64
>> right now.
>>
>> Peter: Do you think something like this will work? loadavg went
>> quite on two of my test systems after this change (4 cpu and 24 cpu).
>
> Not really, CPUs can stay idle for _very_ long times (!x86 cpus that
> don't have crappy timers like HPET which roll around every 2-4 seconds).
>
> But all CPUs staying idle for a long time is exactly the scenario you
> fix before using the decay_load_misses() stuff, except that is for the
> load-balancer per-cpu load numbers not the global cpu load avg. Won't a
> similar approach work here?
>

Yes. Thought about that. One problem there is that works with nohz_idle_balance,
which will not be called if all the CPUs are idle for example.
As this is once in 5 secs, probably doing nr_running() and
nr_uninterruptible() should
be OK even on huge systems. But, that was the original code here, except that
it was inside xtime_lock.

Thanks,
Venki
--
unsubscribe notice
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/