Re: [PATCH v3 0/7] Dynamic Pause Loop Exiting window.

[Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com>]

On 08/21/2014 10:00 PM, Paolo Bonzini wrote:
> Il 21/08/2014 18:08, Radim Krčmář ha scritto:
>> v2 -> v3:
>>   * copy&paste frenzy [v3 4/7] (split modify_ple_window)
>>   * commented update_ple_window_actual_max [v3 4/7]
>>   * renamed shrinker to modifier [v3 4/7]
>>   * removed an extraneous max(new, ple_window) [v3 4/7] (should have been in v2)
>>   * changed tracepoint argument type, printing and macro abstractions [v3 5/7]
>>   * renamed ple_t to ple_int [v3 6/7] (visible in modinfo)
>>   * intelligent updates of ple_window [v3 7/7]
>>
>> ---
>> v1 -> v2:
>>   * squashed [v1 4/9] and [v1 5/9] (clamping)
>>   * dropped [v1 7/9] (CPP abstractions)
>>   * merged core of [v1 9/9] into [v1 4/9] (automatic maximum)
>>   * reworked kernel_param_ops: closer to pure int [v2 6/6]
>>   * introduced ple_window_actual_max & reworked clamping [v2 4/6]
>>   * added seqlock for parameter modifications [v2 6/6]
>>
>> ---
>> PLE does not scale in its current form.  When increasing VCPU count
>> above 150, one can hit soft lockups because of runqueue lock contention.
>> (Which says a lot about performance.)
>>
>> The main reason is that kvm_ple_loop cycles through all VCPUs.
>> Replacing it with a scalable solution would be ideal, but it has already
>> been well optimized for various workloads, so this series tries to
>> alleviate one different major problem while minimizing a chance of
>> regressions: we have too many useless PLE exits.
>>
>> Just increasing PLE window would help some cases, but it still spirals
>> out of control.  By increasing the window after every PLE exit, we can
>> limit the amount of useless ones, so we don't reach the state where CPUs
>> spend 99% of the time waiting for a lock.
>>
>> HP confirmed that this series prevents soft lockups and TSC sync errors
>> on large guests.
>
> Hi,
>
> I'm not sure of the usefulness of patch 6, so I'm going to drop it.
> I'll keep it in my local junkyard branch in case it's going to be useful
> in some scenario I didn't think of.

I think grow knob may be helpful to some extent considering number of 
vcpus can vary from few to hundreds, which in turn helps in fast
convergence of ple_window value in non overcommit scenarios.

I will try to experiment with shrink knob. One argument favouring
shrink knob may be the fact that we rudely reset vmx->ple_window
back to default 4k. Ofcourse danger on the other side is slow 
convergence during overcommit/sudden burst of load.