Message-ID: <aa2845b9-5e57-4df2-9c59-814eedd752dc@mailbox.org>
Date:   Wed, 1 Feb 2023 18:35:04 +0000
MIME-Version: 1.0
Subject: Re: [PATCH v5 0/6] amd_pstate: Add guided autonomous mode support
Content-Language: en-US
To:     Wyes Karny <wyes.karny@amd.com>,
        Rafael J Wysocki <rafael@kernel.org>,
        Huang Rui <ray.huang@amd.com>,
        Jonathan Corbet <corbet@lwn.net>,
        Viresh Kumar <viresh.kumar@linaro.org>,
        Mario.Limonciello@amd.com, Perry.Yuan@amd.com
Cc:     linux-doc@vger.kernel.org, linux-kernel@vger.kernel.org,
        linux-pm@vger.kernel.org, Bagas Sanjaya <bagasdotme@gmail.com>,
        santosh.shukla@amd.com, Len Brown <lenb@kernel.org>,
        Robert Moore <robert.moore@intel.com>,
        Borislav Petkov <bp@alien8.de>,
        Ananth Narayan <ananth.narayan@amd.com>, gautham.shenoy@amd.com
References: <20230201065823.189828-1-wyes.karny@amd.com>
From:   Tor Vic <torvic9@mailbox.org>
In-Reply-To: <20230201065823.189828-1-wyes.karny@amd.com>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 7bit
Precedence: bulk


On 01.02.23 06:58, Wyes Karny wrote:
>  From ACPI spec[1] below 3 modes for CPPC can be defined:
> 1. Non autonomous: OS scaling governor specifies operating frequency/
>     performance level through `Desired Performance` register and platform
> follows that.
> 2. Guided autonomous: OS scaling governor specifies min and max
>     frequencies/ performance levels through `Minimum Performance` and
> `Maximum Performance` register, and platform can autonomously select an
> operating frequency in this range.
> 3. Fully autonomous: OS only hints (via EPP) to platform for the required
>     energy performance preference for the workload and platform autonomously
> scales the frequency.
> 
> Currently (1) is supported by amd_pstate as passive mode, and (3) is
> implemented by EPP support[2]. This change is to support (2).
> 
> In guided autonomous mode the min_perf is based on the input from the
> scaling governor. For example, in case of schedutil this value depends
> on the current utilization. And max_perf is set to max capacity.
> 
> To activate guided auto mode ``amd_pstate=guided`` command line
> parameter has to be passed in the kernel.
> 
> Below are the results (normalized) of benchmarks with this patch:
> System: Genoa 96C 192T
> Kernel: 6.2.0-rc2 + EPP v12 + patch
> Scaling governor: schedutil
> 
> ================ dbench comparisons ================
> dbench result comparison:
> Here results are throughput (MB/s)
> Clients:   acpi-cpufreq		   amd_pst+passive	   amd_pst+guided
>      1	   1.00 (0.00 pct)	   1.01 (1.00 pct)	   1.02 (2.00 pct)
>      2	   1.07 (0.00 pct)	   1.06 (-0.93 pct)	   1.07 (0.00 pct)
>      4	   1.68 (0.00 pct)	   1.70 (1.19 pct)	   1.72 (2.38 pct)
>      8	   2.61 (0.00 pct)	   2.68 (2.68 pct)	   2.76 (5.74 pct)
>     16	   4.16 (0.00 pct)	   4.24 (1.92 pct)	   4.53 (8.89 pct)
>     32	   5.98 (0.00 pct)	   6.17 (3.17 pct)	   7.30 (22.07 pct)
>     64	   8.67 (0.00 pct)	   8.99 (3.69 pct)	  10.71 (23.52 pct)
>    128	  11.98 (0.00 pct)	  12.52 (4.50 pct)	  14.67 (22.45 pct)
>    256	  15.73 (0.00 pct)	  16.13 (2.54 pct)	  17.81 (13.22 pct)
>    512	  15.77 (0.00 pct)	  16.32 (3.48 pct)	  16.39 (3.93 pct)
> dbench power comparison:
> Clients:   acpi-cpufreq		   amd_pst+passive	   amd_pst+guided
>      1	   1.00 (0.00 pct)	   1.00 (0.00 pct)	   1.04 (4.00 pct)
>      2	   0.99 (0.00 pct)	   0.97 (-2.02 pct)	   1.02 (3.03 pct)
>      4	   0.98 (0.00 pct)	   0.98 (0.00 pct)	   1.02 (4.08 pct)
>      8	   0.98 (0.00 pct)	   0.99 (1.02 pct)	   1.02 (4.08 pct)
>     16	   0.99 (0.00 pct)	   1.00 (1.01 pct)	   1.04 (5.05 pct)
>     32	   1.02 (0.00 pct)	   1.02 (0.00 pct)	   1.07 (4.90 pct)
>     64	   1.05 (0.00 pct)	   1.05 (0.00 pct)	   1.11 (5.71 pct)
>    128	   1.08 (0.00 pct)	   1.08 (0.00 pct)	   1.15 (6.48 pct)
>    256	   1.12 (0.00 pct)	   1.12 (0.00 pct)	   1.20 (7.14 pct)
>    512	   1.18 (0.00 pct)	   1.17 (-0.84 pct)	   1.26 (6.77 pct)
> 
> ================ git-source comparisons ================
> git-source result comparison:
> Here results are throughput (compilations per 1000 sec)
> Threads:   acpi-cpufreq		   amd_pst+passive	   amd_pst+guided
>    192	   1.00 (0.00 pct)	   0.93 (-7.00 pct)	   1.00 (0.00 pct)
> git-source power comparison:
> Threads:   acpi-cpufreq		   amd_pst+passive	   amd_pst+guided
>    192	   1.00 (0.00 pct)	   1.00 (0.00 pct)	   0.96 (-4.00 pct)
> 
> ================ kernbench comparisons ================
> kernbench result comparison:
> Here results are throughput (compilations per 1000 sec)
> Load:	   acpi-cpufreq		   amd_pst+passive	   amd_pst+guided
> 32	   1.00 (0.00 pct)	   1.01 (1.00 pct)	   1.02 (2.00 pct)
> 48	   1.26 (0.00 pct)	   1.28 (1.58 pct)	   1.25 (-0.79 pct)
> 64	   1.39 (0.00 pct)	   1.47 (5.75 pct)	   1.43 (2.87 pct)
> 96	   1.48 (0.00 pct)	   1.50 (1.35 pct)	   1.49 (0.67 pct)
> 128	   1.29 (0.00 pct)	   1.32 (2.32 pct)	   1.33 (3.10 pct)
> 192	   1.17 (0.00 pct)	   1.20 (2.56 pct)	   1.21 (3.41 pct)
> 256	   1.17 (0.00 pct)	   1.18 (0.85 pct)	   1.20 (2.56 pct)
> 384	   1.16 (0.00 pct)	   1.17 (0.86 pct)	   1.21 (4.31 pct)
> kernbench power comparison:
> Clients:   acpi-cpufreq		   amd_pst+passive	   amd_pst+guided
>     32	   1.00 (0.00 pct)	   0.97 (-3.00 pct)	   1.00 (0.00 pct)
>     48	   0.87 (0.00 pct)	   0.81 (-6.89 pct)	   0.88 (1.14 pct)
>     64	   0.81 (0.00 pct)	   0.73 (-9.87 pct)	   0.77 (-4.93 pct)
>     96	   0.75 (0.00 pct)	   0.74 (-1.33 pct)	   0.75 (0.00 pct)
>    128	   0.83 (0.00 pct)	   0.79 (-4.81 pct)	   0.83 (0.00 pct)
>    192	   0.92 (0.00 pct)	   0.88 (-4.34 pct)	   0.92 (0.00 pct)
>    256	   0.92 (0.00 pct)	   0.88 (-4.34 pct)	   0.92 (0.00 pct)
>    384	   0.92 (0.00 pct)	   0.88 (-4.34 pct)	   0.92 (0.00 pct)
> 
> ================ tbench comparisons ================
> tbench result comparison:
> Here results are throughput (MB/s)
> Clients:   acpi-cpufreq		   amd_pst+passive	   amd_pst+guided
>      1	   1.00 (0.00 pct)	   0.70 (-30.00 pct)	   1.37 (37.00 pct)
>      2	   2.64 (0.00 pct)	   1.39 (-47.34 pct)	   2.70 (2.27 pct)
>      4	   4.89 (0.00 pct)	   2.75 (-43.76 pct)	   5.28 (7.97 pct)
>      8	   9.46 (0.00 pct)	   5.42 (-42.70 pct)	  10.22 (8.03 pct)
>     16	  19.05 (0.00 pct)	  10.42 (-45.30 pct)	  19.94 (4.67 pct)
>     32	  37.50 (0.00 pct)	  20.23 (-46.05 pct)	  36.87 (-1.68 pct)
>     64	  61.24 (0.00 pct)	  43.08 (-29.65 pct)	  62.96 (2.80 pct)
>    128	  67.16 (0.00 pct)	  69.08 (2.85 pct)	  67.34 (0.26 pct)
>    256	 154.59 (0.00 pct)	 162.33 (5.00 pct)	 156.78 (1.41 pct)
>    512	 154.02 (0.00 pct)	 156.74 (1.76 pct)	 153.48 (-0.35 pct)
> tbench power comparison:
> Clients:   acpi-cpufreq		   amd_pst+passive	   amd_pst+guided
>      1	   1.00 (0.00 pct)	   0.97 (-3.00 pct)	   1.08 (8.00 pct)
>      2	   1.04 (0.00 pct)	   0.97 (-6.73 pct)	   1.11 (6.73 pct)
>      4	   1.12 (0.00 pct)	   0.99 (-11.60 pct)	   1.18 (5.35 pct)
>      8	   1.25 (0.00 pct)	   1.04 (-16.80 pct)	   1.31 (4.80 pct)
>     16	   1.53 (0.00 pct)	   1.13 (-26.14 pct)	   1.58 (3.26 pct)
>     32	   2.01 (0.00 pct)	   1.36 (-32.33 pct)	   2.03 (0.99 pct)
>     64	   2.58 (0.00 pct)	   2.14 (-17.05 pct)	   2.61 (1.16 pct)
>    128	   2.80 (0.00 pct)	   2.81 (0.35 pct)	   2.81 (0.35 pct)
>    256	   3.39 (0.00 pct)	   3.43 (1.17 pct)	   3.42 (0.88 pct)
>    512	   3.44 (0.00 pct)	   3.44 (0.00 pct)	   3.44 (0.00 pct)
> 
> Note: this series is based on top of EPP v12 [3] series
> 
> Change log:
> 
> v4 -> v5:
> - Rebased on top of EPP v12 series
> - Addressed comments form Mario regarding documentation
> - Picked up RB flags from Mario and Bagas Sanjaya
> 
> v3 -> v4:
> - Fixed active mode low frequency issue reported by Peter Jung and Tor Vic
> - Documentation modification suggested by Bagas Sanjaya
> 
> v2 -> v3:
> - Addressed review comments form Mario.
> - Picked up RB tag from Mario.
> - Rebase on top of EPP v11 [3].
> 
> v1 -> v2:
> - Fix issue with shared mem systems.
> - Rebase on top of EPP series.
> 
> [1]: https://uefi.org/sites/default/files/resources/ACPI_6_3_final_Jan30.pdf
> [2]: https://lore.kernel.org/lkml/20221110175847.3098728-1-Perry.Yuan@amd.com/
> [3]: https://lore.kernel.org/linux-pm/20230131090016.3970625-1-perry.yuan@amd.com/
> 
> Wyes Karny (6):
>    acpi: cppc: Add min and max perf reg writing support
>    acpi: cppc: Add auto select register read/write support
>    cpufreq: amd_pstate: Add guided autonomous mode
>    Documentation: amd_pstate: Move amd_pstate param to alphabetical order
>    cpufreq: amd_pstate: Add guided mode control support via sysfs
>    Documentation: amd_pstate: Update amd_pstate status sysfs for guided
> 
>   .../admin-guide/kernel-parameters.txt         |  41 ++--
>   Documentation/admin-guide/pm/amd-pstate.rst   |  34 +++-
>   drivers/acpi/cppc_acpi.c                      | 121 +++++++++++-
>   drivers/cpufreq/amd-pstate.c                  | 183 +++++++++++++-----
>   include/acpi/cppc_acpi.h                      |  11 ++
>   include/linux/amd-pstate.h                    |   2 +
>   6 files changed, 311 insertions(+), 81 deletions(-)
> 

I tested "active" mode on a Zen2, using EPP v12 and this series on top.

Here is the output of 'cpupower frequency-info' run as root under full load:

   analyzing CPU 0:
     driver: amd_pstate_epp
     CPUs which run at the same hardware frequency: 0
     CPUs which need to have their frequency coordinated by software: 0
     maximum transition latency:  Cannot determine or is not supported.
     hardware limits: 550 MHz - 4.67 GHz
     available cpufreq governors: performance powersave
     current policy: frequency should be within 550 MHz and 4.67 GHz.
                   The governor "powersave" may decide which speed to use
                   within this range.
     current CPU frequency: Unable to call hardware
     current CPU frequency: 4.18 GHz (asserted by call to kernel)
     boost state support:
       Supported: yes
       Active: yes
       Boost States: 0
       Total States: 3
       Pstate-P0:  3800MHz
       Pstate-P1:  2800MHz
       Pstate-P2:  2200MHz

Changing EPP mode via sysfs works.

----

Live-switching to "guided" mode via sysfs also seems to work fine:

   analyzing CPU 0:
     driver: amd-pstate
     CPUs which run at the same hardware frequency: 0
     CPUs which need to have their frequency coordinated by software: 0
     maximum transition latency: 20.0 us
     hardware limits: 550 MHz - 4.67 GHz
     available cpufreq governors: ondemand powersave performance schedutil
     current policy: frequency should be within 550 MHz and 4.67 GHz.
                   The governor "ondemand" may decide which speed to use
                   within this range.
     current CPU frequency: Unable to call hardware
     current CPU frequency: 3.71 GHz (asserted by call to kernel)
     boost state support:
       Supported: yes
       Active: yes
       AMD PSTATE Highest Performance: 166. Maximum Frequency: 4.67 GHz.
       AMD PSTATE Nominal Performance: 135. Nominal Frequency: 3.80 GHz.
       AMD PSTATE Lowest Non-linear Performance: 62. Lowest Non-linear 
Frequency: 1.74 GHz.
     AMD PSTATE Lowest Performance: 20. Lowest Frequency: 550 MHz.

----

Seems to work fine so far.
Hope this helps.
I can test more in detail and give you a Tested-by tag if wanted.

Thanks,

Tor Vic