[RFC/RFT][PATCH] cpufreq: intel_pstate: Improve IO performance

[RFC/RFT][PATCH] cpufreq: intel_pstate: Improve IO performance

[Srinivas Pandruvada]

In the current implementation the latency from SCHED_CPUFREQ_IOWAIT is
set to actual P-state adjustment can be upto 10ms. This can be improved
by reacting to SCHED_CPUFREQ_IOWAIT faster in a milli second. With this
trivial change the IO performance improves significantly.

With a simple "grep -r . linux" (Here linux is kernel source folder) with
dropped caches every time on a platform with per core P-states
(Broadwell and Haswell Xeon ), the performance difference is significant.
The user and kernel time improvement is more than 20%.

The same performance difference was not observed on clients and on a
IvyTown server. which don't have per core P-state support.
So the performance gain may not be apparent on all systems.

Signed-off-by: Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com>
---
The idea of this patch is to test if it brings in any significant
improvement on real world use cases.

 drivers/cpufreq/intel_pstate.c | 9 +++++++--
 1 file changed, 7 insertions(+), 2 deletions(-)

diff --git a/drivers/cpufreq/intel_pstate.c b/drivers/cpufreq/intel_pstate.c
index 8c67b77..639979c 100644
--- a/drivers/cpufreq/intel_pstate.c
+++ b/drivers/cpufreq/intel_pstate.c
@@ -38,6 +38,7 @@
 #include <asm/intel-family.h>
 
 #define INTEL_PSTATE_DEFAULT_SAMPLING_INTERVAL	(10 * NSEC_PER_MSEC)
+#define INTEL_PSTATE_IO_WAIT_SAMPLING_INTERVAL	(NSEC_PER_MSEC)
 #define INTEL_PSTATE_HWP_SAMPLING_INTERVAL	(50 * NSEC_PER_MSEC)
 
 #define INTEL_CPUFREQ_TRANSITION_LATENCY	20000
@@ -287,6 +288,7 @@ static struct pstate_funcs pstate_funcs __read_mostly;
 
 static int hwp_active __read_mostly;
 static bool per_cpu_limits __read_mostly;
+static int current_sample_interval = INTEL_PSTATE_DEFAULT_SAMPLING_INTERVAL;
 
 static struct cpufreq_driver *intel_pstate_driver __read_mostly;
 
@@ -1527,15 +1529,18 @@ static void intel_pstate_update_util(struct update_util_data *data, u64 time,
 
 	if (flags & SCHED_CPUFREQ_IOWAIT) {
 		cpu->iowait_boost = int_tofp(1);
+		current_sample_interval = INTEL_PSTATE_IO_WAIT_SAMPLING_INTERVAL;
 	} else if (cpu->iowait_boost) {
 		/* Clear iowait_boost if the CPU may have been idle. */
 		delta_ns = time - cpu->last_update;
-		if (delta_ns > TICK_NSEC)
+		if (delta_ns > TICK_NSEC) {
 			cpu->iowait_boost = 0;
+			current_sample_interval = INTEL_PSTATE_DEFAULT_SAMPLING_INTERVAL;
+		}
 	}
 	cpu->last_update = time;
 	delta_ns = time - cpu->sample.time;
-	if ((s64)delta_ns < INTEL_PSTATE_DEFAULT_SAMPLING_INTERVAL)
+	if ((s64)delta_ns < current_sample_interval)
 		return;
 
 	if (intel_pstate_sample(cpu, time)) {
-- 
2.7.4

Re: [RFC/RFT][PATCH] cpufreq: intel_pstate: Improve IO performance

[Rafael J. Wysocki]

On Thursday, July 27, 2017 11:44:52 PM Srinivas Pandruvada wrote:
> In the current implementation the latency from SCHED_CPUFREQ_IOWAIT is
> set to actual P-state adjustment can be upto 10ms. This can be improved
> by reacting to SCHED_CPUFREQ_IOWAIT faster in a milli second. With this
> trivial change the IO performance improves significantly.
> 
> With a simple "grep -r . linux" (Here linux is kernel source folder) with
> dropped caches every time on a platform with per core P-states
> (Broadwell and Haswell Xeon ), the performance difference is significant.
> The user and kernel time improvement is more than 20%.
> 
> The same performance difference was not observed on clients and on a
> IvyTown server. which don't have per core P-state support.
> So the performance gain may not be apparent on all systems.
> 
> Signed-off-by: Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com>
> ---
> The idea of this patch is to test if it brings in any significant
> improvement on real world use cases.
> 
>  drivers/cpufreq/intel_pstate.c | 9 +++++++--
>  1 file changed, 7 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/cpufreq/intel_pstate.c b/drivers/cpufreq/intel_pstate.c
> index 8c67b77..639979c 100644
> --- a/drivers/cpufreq/intel_pstate.c
> +++ b/drivers/cpufreq/intel_pstate.c
> @@ -38,6 +38,7 @@
>  #include <asm/intel-family.h>
>  
>  #define INTEL_PSTATE_DEFAULT_SAMPLING_INTERVAL	(10 * NSEC_PER_MSEC)
> +#define INTEL_PSTATE_IO_WAIT_SAMPLING_INTERVAL	(NSEC_PER_MSEC)
>  #define INTEL_PSTATE_HWP_SAMPLING_INTERVAL	(50 * NSEC_PER_MSEC)

First offf, can we simply set INTEL_PSTATE_DEFAULT_SAMPLING_INTERVAL to NSEC_PER_MSEC?

I guess it may help quite a bit in the more "interactive" cases overall.

Or would that be too much overhead?

>  #define INTEL_CPUFREQ_TRANSITION_LATENCY	20000
> @@ -287,6 +288,7 @@ static struct pstate_funcs pstate_funcs __read_mostly;
>  
>  static int hwp_active __read_mostly;
>  static bool per_cpu_limits __read_mostly;
> +static int current_sample_interval = INTEL_PSTATE_DEFAULT_SAMPLING_INTERVAL;
>  
>  static struct cpufreq_driver *intel_pstate_driver __read_mostly;
>  
> @@ -1527,15 +1529,18 @@ static void intel_pstate_update_util(struct update_util_data *data, u64 time,
>  
>  	if (flags & SCHED_CPUFREQ_IOWAIT) {
>  		cpu->iowait_boost = int_tofp(1);
> +		current_sample_interval = INTEL_PSTATE_IO_WAIT_SAMPLING_INTERVAL;
>  	} else if (cpu->iowait_boost) {
>  		/* Clear iowait_boost if the CPU may have been idle. */
>  		delta_ns = time - cpu->last_update;
> -		if (delta_ns > TICK_NSEC)
> +		if (delta_ns > TICK_NSEC) {
>  			cpu->iowait_boost = 0;
> +			current_sample_interval = INTEL_PSTATE_DEFAULT_SAMPLING_INTERVAL;

Second, if reducing INTEL_PSTATE_DEFAULT_SAMPLING_INTERVAL is not viable, why
does the sample interval have to be reduced for all CPUs if SCHED_CPUFREQ_IOWAIT
is set for one of them and not just for the CPU receiving that flag?

> +		}
>  	}
>  	cpu->last_update = time;
>  	delta_ns = time - cpu->sample.time;
> -	if ((s64)delta_ns < INTEL_PSTATE_DEFAULT_SAMPLING_INTERVAL)
> +	if ((s64)delta_ns < current_sample_interval)
>  		return;
>  
>  	if (intel_pstate_sample(cpu, time)) {
> 

Re: [RFC/RFT][PATCH] cpufreq: intel_pstate: Improve IO performance

[Srinivas Pandruvada]

On Mon, 2017-07-31 at 14:21 +0200, Rafael J. Wysocki wrote:
> 

[...]

> >  #define INTEL_PSTATE_DEFAULT_SAMPLING_INTERVAL	(10 *
> > NSEC_PER_MSEC)
> > +#define INTEL_PSTATE_IO_WAIT_SAMPLING_INTERVAL	(NSEC_PER_MS
> > EC)
> >  #define INTEL_PSTATE_HWP_SAMPLING_INTERVAL	(50 *
> > NSEC_PER_MSEC)
> First offf, can we simply set INTEL_PSTATE_DEFAULT_SAMPLING_INTERVAL
> to NSEC_PER_MSEC?
> 
> I guess it may help quite a bit in the more "interactive" cases
> overall.
> 
> Or would that be too much overhead?
It will be too much overhead for clients.
 
> 
> > 
> > 
[...]

> > @@ -1527,15 +1529,18 @@ static void intel_pstate_update_util(struct
> > update_util_data *data, u64 time,
> >  
> >  	if (flags & SCHED_CPUFREQ_IOWAIT) {
> >  		cpu->iowait_boost = int_tofp(1);
> > +		current_sample_interval =
> > INTEL_PSTATE_IO_WAIT_SAMPLING_INTERVAL;
> >  	} else if (cpu->iowait_boost) {
> >  		/* Clear iowait_boost if the CPU may have been
> > idle. */
> >  		delta_ns = time - cpu->last_update;
> > -		if (delta_ns > TICK_NSEC)
> > +		if (delta_ns > TICK_NSEC) {
> >  			cpu->iowait_boost = 0;
> > +			current_sample_interval =
> > INTEL_PSTATE_DEFAULT_SAMPLING_INTERVAL;
> Second, if reducing INTEL_PSTATE_DEFAULT_SAMPLING_INTERVAL is not
> viable, why
> does the sample interval have to be reduced for all CPUs if
> SCHED_CPUFREQ_IOWAIT
> is set for one of them and not just for the CPU receiving that flag?
> 
Correct, the read data may be passed to some other thread for
processing.

Replacing this patch with one below improves simple grep user time and
system time by 40% on Haswell servers.

diff --git a/drivers/cpufreq/intel_pstate.c
b/drivers/cpufreq/intel_pstate.c
index 48a98f11a84e..dce3e324a9da 100644
--- a/drivers/cpufreq/intel_pstate.c
+++ b/drivers/cpufreq/intel_pstate.c
@@ -1751,6 +1751,16 @@ static void intel_pstate_update_util(struct
update_util_data *data, u64 time,
 
        if (flags & SCHED_CPUFREQ_IOWAIT) {
                cpu->iowait_boost = int_tofp(1);
+               /*
+                * The last time the busy was 100% so P-state was max
anyway
+                * so avoid overhead of computation.
+                */
+               if (fp_toint(cpu->sample.busy_scaled) == 100) {
+                       cpu->last_update = time;
+                       return;
+               }
+                goto set_pstate;
+
        } else if (cpu->iowait_boost) {
                /* Clear iowait_boost if the CPU may have been idle. */
                delta_ns = time - cpu->last_update;
@@ -1762,6 +1772,7 @@ static void intel_pstate_update_util(struct
update_util_data *data, u64 time,
        if ((s64)delta_ns < INTEL_PSTATE_DEFAULT_SAMPLING_INTERVAL)
                return;
 
+set_pstate:
        if (intel_pstate_sample(cpu, time)) {
                int target_pstate;