Re: [PATCH 4/4] cpufreq: intel_pstate: enable boost for Skylake Xeon

[Francisco Jerez]

Srinivas Pandruvada <srinivas.pandruvada@xxxxxxxxxxxxxxx> writes:

> On Fri, 2018-07-27 at 22:34 -0700, Francisco Jerez wrote:
>> Srinivas Pandruvada <srinivas.pandruvada@xxxxxxxxxxxxxxx> writes:
>> 
>> > Enable HWP boost on Skylake server and workstations.
>> > 
>> 
>> Please revert this series, it led to significant energy usage and
>> graphics performance regressions [1]. 
> Which SKX platform is targeted to graphics?
>

See the bug report, it's a regular desktop SKL.

>>  The reasons are roughly the ones
>> we discussed by e-mail off-list last April: This causes the
>> intel_pstate
>> driver to decrease the EPP to zero 
> No. You didn't check this series. We are not using EPP at all.
> The boost mechanism used here is not boost to max.
>

I see you've changed the mechanism to obtain a latency boost since our
last discussion, but that doesn't address my concerns in any way: This
series causes the intel_pstate driver to clamp the CPU frequency above
the optimal frequency to run the workload at, as a response to the
application waiting on IO frequently, even though that's only a sign of
the application being IO-bound and *not* a sign of it being
latency-sensitive, since the application's IO and CPU work are properly
pipelined.  This leads to a decrease in parallelism due to the active
CPU core using a larger fraction of the package TDP in order to achieve
the same work, leading to a decrease in system performance.

> Thanks,
> Srinivas
>
>> when the workload blocks on IO
>> frequently enough, which for the regressing benchmarks detailed in
>> [1]
>> is a symptom of the workload being heavily IO-bound, which means they
>> won't benefit at all from the EPP boost since they aren't
>> significantly
>> CPU-bound, and they will suffer a decrease in parallelism due to the
>> active CPU core using a larger fraction of the TDP in order to
>> achieve
>> the same work, causing the GPU to have a lower power budget
>> available,
>> leading to a decrease in system performance.
>> 
>> You may want to give a shot to my previous suggestion of using [2] in
>> order to detect whether the system is IO-bound, which you can use as
>> an
>> indicator that the optimization implemented in this series cannot
>> possibly improve performance and can be expected to hurt energy
>> usage.
>> 
>> Thanks.
>> 
>> [1] https://bugs.freedesktop.org/show_bug.cgi?id=107410
>> [2] https://patchwork.kernel.org/patch/10312259/
>> 
>> > Reported-by: Mel Gorman <mgorman@xxxxxxxxxxxxxxxxxxx>
>> > Tested-by: Giovanni Gherdovich <ggherdovich@xxxxxxx>
>> > Signed-off-by: Srinivas Pandruvada <srinivas.pandruvada@xxxxxxxxxxx
>> > .com>
>> > ---
>> >  drivers/cpufreq/intel_pstate.c | 10 ++++++++++
>> >  1 file changed, 10 insertions(+)
>> > 
>> > diff --git a/drivers/cpufreq/intel_pstate.c
>> > b/drivers/cpufreq/intel_pstate.c
>> > index 70bf63bb4e0e..01c8da1f99db 100644
>> > --- a/drivers/cpufreq/intel_pstate.c
>> > +++ b/drivers/cpufreq/intel_pstate.c
>> > @@ -1794,6 +1794,12 @@ static const struct x86_cpu_id
>> > intel_pstate_cpu_ee_disable_ids[] = {
>> >  	{}
>> >  };
>> >  
>> > +static const struct x86_cpu_id intel_pstate_hwp_boost_ids[]
>> > __initconst = {
>> > +	ICPU(INTEL_FAM6_SKYLAKE_X, core_funcs),
>> > +	ICPU(INTEL_FAM6_SKYLAKE_DESKTOP, core_funcs),
>> > +	{}
>> > +};
>> > +
>> >  static int intel_pstate_init_cpu(unsigned int cpunum)
>> >  {
>> >  	struct cpudata *cpu;
>> > @@ -1824,6 +1830,10 @@ static int intel_pstate_init_cpu(unsigned
>> > int cpunum)
>> >  			intel_pstate_disable_ee(cpunum);
>> >  
>> >  		intel_pstate_hwp_enable(cpu);
>> > +
>> > +		id = x86_match_cpu(intel_pstate_hwp_boost_ids);
>> > +		if (id)
>> > +			hwp_boost = true;
>> >  	}
>> >  
>> >  	intel_pstate_get_cpu_pstates(cpu);
>> > -- 
>> > 2.13.6

Attachment: signature.asc
Description: PGP signature