AmpereOne CPPC CPUFreq Schedutil vs. Performance Governor Benchmarks

Written by Michael Larabel in Hardware on 27 September 2024 at 11:02 AM EDT. 2 Comments

Similar to the ACPI CPUFreq and AMD/Intel P-State CPU frequency scaling driver and scaling governor benchmarks and power efficiency comparisons I routinely do on Phoronix, when recently having the Supermicro AmpereOne server in the lab with the 192-core A192-32X processor, I carried out some CPPC CPUFreq schedutil vs. performance governor benchmarks for curiosity and reference purposes while looking at the performance and power efficiency.

With all the AmpereOne A192-32X benchmarks I ran for the CPU comparisons while having the Supermicro ARS-211M-NR R13SPD in the lab for a few weeks, they were as usual for server CPUs done with the "performance" governor. But Ubuntu 24.04 LTS ARM and others default to using the "schedutil" governor out-of-the-box for relying on scheduler utilization data. For those curious what schedutil vs. performance difference there is for ARM processors with AmpereOne, I ran some comparison benchmarks.

AmpereOne CPUFreq Driver Governor Performance

These benchmarks are of the AmpereOne A192-32X with the schedutil default and then again when switching to the performance governor. With my earlier AmpereOne / Intel / AMD server processor benchmarks there were done with "performance", so no surprises there for better performance or the like. The CPU package power consumption was also monitored for seeing the impact on power efficiency between these governor options.

Timed Linux Kernel Compilation benchmark with settings of Build: defconfig. cppc_cpufreq performance was the fastest.

Timed LLVM Compilation benchmark with settings of Build System: Ninja. cppc_cpufreq performance was the fastest.

7-Zip Compression benchmark with settings of Test: Compression Rating. cppc_cpufreq schedutil was the fastest.

7-Zip Compression benchmark with settings of Test: Decompression Rating. cppc_cpufreq performance was the fastest.

For many workloads there was no real difference in performance or CPU power consumption between Schedutil and Performance... Good to see Schedutil behaving appropriately.

FFmpeg benchmark with settings of Encoder: libx265, Scenario: Live. cppc_cpufreq performance was the fastest.

FFmpeg benchmark with settings of Encoder: libx265, Scenario: Platform. cppc_cpufreq performance was the fastest.

Like we've seen with AMD and Intel CPUs, in cases like video encoding/transcoding they can be much more sensitive to the CPU frequency scaling governor choices...

x265 benchmark with settings of Video Input: Bosphorus 4K. cppc_cpufreq performance was the fastest.

Switching to the performance governor can be rather impactful for better performance on AmpereOne and similar to the delta seen on AMD and Intel processors too.

TTSIOD 3D Renderer benchmark with settings of Phong Rendering With Soft-Shadow Mapping. cppc_cpufreq performance was the fastest.

The TTSIOD Renderer saw much better performance with the performance governor than the Schedutil default behavior.

RawTherapee benchmark with settings of Total Benchmark Time. cppc_cpufreq performance was the fastest.

GEGL benchmark with settings of Operation: Reflect. cppc_cpufreq performance was the fastest.

Quantum ESPRESSO benchmark with settings of Input: AUSURF112. cppc_cpufreq performance was the fastest.

TensorFlow Lite benchmark with settings of Model: Inception V4. cppc_cpufreq schedutil was the fastest.

In total I ran 70 benchmarks for this comparison and you can see all the individual data here.

It was predominantly with the video encoding benchmarks where the performance governor was offering much better performance.

CPU Power Consumption Monitor benchmark with settings of Phoronix Test Suite System Monitoring.

Using the performance governor rather than schedutil drove up the AmpereOne A192-32X CPU power consumption from a 137 Watt average to 154 Watts. With either governor the idle CPU power consumption of the AmpereOne CPU remained high at nearly 100 Watts. Both peaked just below 400 Watts.

Geometric Mean Of All Test Results benchmark with settings of Result Composite, AmpereOne CPUFreq Driver Governor Performance. cppc_cpufreq performance was the fastest.

Again these results are mainly being put out for reference purposes for those wondering if the "performance" governor is very impactful for AmpereOne / how the default Schedutil governor behaves on the likes of Ubuntu 24.04 LTS for ARM servers, etc. No real surprises and largely similar to what is observed with AMD and Intel processors under Linux with opting for different governors.

2 Comments