All Projects → kdrag0n → Freqbench

kdrag0n / Freqbench

Licence: mit
Comprehensive CPU frequency performance/power benchmark

Programming Languages

python
139335 projects - #7 most used programming language

Projects that are alternatives of or similar to Freqbench

c2clat
A tool to measure CPU core to core latency
Stars: ✭ 37 (-43.08%)
Mutual labels:  benchmark, cpu
Pcm
Processor Counter Monitor
Stars: ✭ 1,240 (+1807.69%)
Mutual labels:  cpu, energy
Cpu X
CPU-X is a Free software that gathers information on CPU, motherboard and more
Stars: ✭ 676 (+940%)
Mutual labels:  cpu, benchmark
Eemeter
An open source python package for implementing and developing standard methods for calculating normalized metered energy consumption and avoided energy use.
Stars: ✭ 134 (+106.15%)
Mutual labels:  efficiency, energy
cult
CPU Ultimate Latency Test.
Stars: ✭ 67 (+3.08%)
Mutual labels:  benchmark, cpu
Sympact
🔥 Stupid Simple CPU/MEM "Profiler" for your JS code.
Stars: ✭ 439 (+575.38%)
Mutual labels:  cpu, benchmark
Go Benchmark App
Application for HTTP benchmarking via different rules and configs
Stars: ✭ 21 (-67.69%)
Mutual labels:  efficiency, benchmark
Dbbench
🏋️ dbbench is a simple database benchmarking tool which supports several databases and own scripts
Stars: ✭ 52 (-20%)
Mutual labels:  benchmark
Opentaps seas
opentaps Smart Energy Applications Suite
Stars: ✭ 55 (-15.38%)
Mutual labels:  energy
Darkriscv
opensouce RISC-V cpu core implemented in Verilog from scratch in one night!
Stars: ✭ 1,062 (+1533.85%)
Mutual labels:  cpu
Awesome Semantic Segmentation
🤘 awesome-semantic-segmentation
Stars: ✭ 8,831 (+13486.15%)
Mutual labels:  benchmark
Docker Backuppc
Docker container with BackupPC version 4.x/3.x based on Alpine distribution.
Stars: ✭ 53 (-18.46%)
Mutual labels:  alpine
Rainbarf
it's like Rainmeter, but for CLI!
Stars: ✭ 1,087 (+1572.31%)
Mutual labels:  cpu
Raspberrypi tempmon
Raspberry pi CPU temperature monitor with many functions such as logging, GPIO output, graphing, email, alarm, notifications and stress testing. Python 3.
Stars: ✭ 52 (-20%)
Mutual labels:  cpu
Rb
A thread-safe fixed-size circular buffer written in safe Rust.
Stars: ✭ 59 (-9.23%)
Mutual labels:  benchmark
Jsbench Me
jsbench.me - JavaScript performance benchmarking playground
Stars: ✭ 50 (-23.08%)
Mutual labels:  benchmark
Scalajs Benchmark
Benchmarks: write in Scala or JS, run in your browser. Live demo:
Stars: ✭ 63 (-3.08%)
Mutual labels:  benchmark
Functional Components Benchmark
Directly calling functional components instead of mounting them is faster.
Stars: ✭ 59 (-9.23%)
Mutual labels:  benchmark
Socat
Run socat command in alpine container
Stars: ✭ 53 (-18.46%)
Mutual labels:  alpine
Pytest Django Queries
Generate performance reports from your django database performance tests.
Stars: ✭ 54 (-16.92%)
Mutual labels:  benchmark

freqbench

Power usage in mW per frequency per cluster for Qualcomm Snapdragon 835, 855, and 765G

freqbench is a comprehensive CPU benchmark that benchmarks each CPU frequency step on each frequency scaling domain (e.g. ARM DynamIQ/big.LITTLE cluster). It is based on a minimal Alpine Linux userspace with the EEMBC CoreMark workload and a Python benchmark coordinator.

Results include:

  • Performance (CoreMark scores)
  • Performance efficiency (CoreMarks per MHz)
  • Power usage (in milliwatts)
  • Energy usage (in millijoules and joules)
  • Energy efficiency (ULPMark-CM scores: iterations per second per millijoule of energy used)
  • Baseline power usage
  • Time elapsed
  • CPU frequency scaling stats during the benchmark (for validation)
  • Diagnostic data (logs, kernel version, kernel command line, interrupts, processes)
  • Raw power samples in machine-readable JSON format (for postprocessing)

Why?

A benchmark like this can be useful for many reasons:

  • Creating energy models for EAS (Energy-Aware Scheduling)
  • Correcting inaccurate EAS energy models
  • Analyzing performance and power trends
  • Comparing efficiency across SoC and CPU generations
  • Improving performance and battery life of mobile devices by utilizing the race-to-idle phenomenon with efficient frequencies

Usage

It is possible to use freqbench with a stock kernel, but a custom kernel is highly recommended for accuracy. Stock OEM kernels are almost always missing features that the benchmark coordinator relies on for maximum accuracy. Custom kernel results are eligible for high accuracy classification, while stock kernel results are limited to low accuracy. Use a stock kernel at your own risk.

Custom kernel (recommended)

Set the following kernel config options:

CONFIG_NO_HZ_FULL=y
CONFIG_CPU_FREQ_TIMES=n  # may not exist
CONFIG_CPU_FREQ_GOV_USERSPACE=y
CONFIG_HZ_100=y

Example commit: kirin_defconfig: Configure for freqbench

If you have any commits that prevent userspace from controlling CPU affinities and utilization, frequencies, or anything of the sort, revert them for the benchmark to work properly. Here are some common examples of such commits in downstream kernels and their corresponding reverts:

Example freqbench kernel adaptations:

Compile and flash your new kernel. Note that Android will not work properly on this kernel, so make sure you take a backup of your old boot image to restore later.

If necessary, adjust the config parameters in config.sh. Most modern devices will not need any changes. Run pack-zip.sh and flash freqbench-installer.zip.

Unplug the device immediately, before the device starts booting. Do not try to wait for it to finish booting. Leaving the device plugged in will invalidate all power results.

Finally, wait until the device reboots itself. Do not touch the device, any of its buttons, or plug/unplug it during the test. It will be frozen on the bootloader splash screen; do not assume that it is broken. The benchmark is expected to take a long time; 1 hour is reasonable for a slower CPU.

Once the benchmark is done, retrieve the results from /cache/freqbench if your device has a cache partition, or /persist/freqbench otherwise (newer devices with A/B partitions don't have a cache partition).

If you are able to retrieve results, please consider contributing your results! It's very helpful for me to see how well freqbench is working, and enables anyone to analyze results across different SoCs that they don't have.

If you have any problems, check the troubleshooting section before opening an issue.

Manual boot image creation

Manually creating a new boot image with the kernel and ramdisk is only for advanced users. Use the AnyKernel3 installer unless you have good reason to do this.

Additional kernel config options:

CONFIG_CMDLINE="rcu_nocbs=0-7 isolcpus=1-7 nohz_full=1-7 loglevel=0 printk.devkmsg=on"
CONFIG_CMDLINE_EXTEND=y

If you don't have 8 CPU cores, adjust 0-7 to 0-<core count - 1> and 1-7 to 1-<core count - 1> where appropriate. Single-core CPUs are not supported. Be careful when adjusting the CPU sets as rcu_nocbs starts at CPU 0 while all other parameters start at CPU 1.

Create a boot image with your modified kernel and the freqbench ramdisk:

For boot image v0/v1 devices:

cd boot-v1
./unpack.sh path/to/normal/boot.img
./pack.sh
# New boot image will be created as new.img

For boot image v3 devices:

# Extract values from boot.img and update pack-img.sh accordingly
./pack-img.sh
# New boot image will be created as bench.img

After that, boot the modified image with fastboot boot if your device supports it, or flash it to the boot/recovery partition and boot that manually.

Results

After the benchmark finishes, results can be found in /cache/freqbench, /persist/freqbench, or /mnt/vendor/persist/freqbench, in that order of preference. The first path that exists on your device will be used. Human-readable results, raw machine-readable JSON data, and diagnostic information are included for analysis.

If you got this far, please consider contributing your results to help freqbench evolve and gather data about different SoCs:

Contributing results

If you run the benchmark on a SoC that is not already included, please contribute your results! It's very helpful for me to see how well freqbench is working, and enables anyone to analyze results across different SoCs that they don't have.

Contributing your results is easy:

  1. Fork this repository
  2. Add your entire results folder (not just one file from it) to results/socname/main, replacing socname with the model name of your SoC in lowercase
  3. Open a pull request

If you don't know your SoC's model name, search the name of your SoC (e.g. Snapdragon 855) and find the part number from the SoC manufacturer. You can also get it from your kernel source code and/or device.txt or cpuinfo.txt in the freqbench results. If you are still unsure, feel free to open an issue or guess the name.

Example names:

  • sm8150
  • sm8150ac
  • sm7250ab
  • exynos8895
  • mt6889

Identifiable information such as the device serial number is automatically redacted by freqbench, so it should not be a problem.

Don't worry about getting something wrong; I would much rather have results submitted with mistakes than nothing at all.

Post-processing

Several post-processing scripts, all written in Python and some using matplotlib, are available:

Legacy energy model

Create a legacy EAS energy model for use with older kernels.

Optional argument after path to results: key_type/value_type

Key types:

  • Frequency (default) - looks like 652800 or 2323200
  • Capacity - looks like 139 or 1024

You must use the correct key type for your kernel. When in doubt, refer to your original energy model and check which one the numbers look more like.

In general, downstream Qualcomm kernels will use the following key types depending on version:

  • 3.18: capacity
  • 4.4: capacity
  • 4.9: frequency
  • 4.14: frequency
  • 4.19: N/A (uses simplified energy model instead)

Modifying your kernel to switch from one to the other is left as an exercise for the reader.

Value types:

  • Power (default)
  • Energy (experimental, not recommended)

Do not change the value type unless you know what you're doing. The energy type only exists for testing purposes; do not expect it to work properly.

Once you have a full energy model generated, pick out the parts you need and incorporate them into your SoC device tree. In general, kernels 4.19 need capacity-dmips-mhz, while older kernels need efficiency when it comes to the contents of the CPU sections.

If you have an existing energy model that you want to use for idle and cluster costs, add it as an argument.

Example usage: ./legacy_em.py results.json cap/power old_model.dtsi

Simplified energy model

Create a simplified EAS energy model for use with newer kernels.

Because voltages defined by the CPU frequency scaling driver cannot easily be accessed from userspace, you will need to provide them. Each frequency step for each cluster have its voltage specified as an argument: cpu#.khz=microvolts

For Qualcomm SoCs on the msm-4.19 kernel, voltages can be obtained by booting the kernel (with or without freqbench doesn't matter, as long as you can get kernel logs) with this commit and searching for lines containing volt= in the kernel log.

For msm-4.9 and msm-4.14, the process is the same but with this commit and searching for open_loop_voltage instead.

Example usage: ./simplified_em.py results.json 1.300000=580000 1.576000=580000 1.614400=580000 1.864000=644000 1.1075200=708000 1.1363200=788000 1.1516800=860000 1.1651200=888000 1.1804800=968000 6.652800=624000 6.940800=672000 6.1152000=704000 6.1478400=752000 6.1728000=820000 6.1900800=864000 6.2092800=916000 6.2208000=948000 7.806400=564000 7.1094400=624000 7.1401600=696000 7.1766400=776000 7.1996800=836000 7.2188800=888000 7.2304000=916000 7.2400000=940000

Efficient frequencies (experimental)

Derive a list of efficient frequencies for each cluster and create a new results.json with only those frequencies included.

Note that this script is experimental and may not produce optimal results. Manual tuning of the resulting frequency tables is recommended.

Example usage: ./efficient_freqs.py results.json eff_results.json

Filter frequencies

Create a new results.json with only the specified frequencies included.

Example usage: ./filter_freqs.py results.json filtered_results.json 1.1516800 1.1804800 6.1478400 6.1728000 6.2208000 7.1766400 7.2188800 7.2304000 7.2400000

Cross-CPU cluster graph

Performance (iter/s) across 835, 855, and 765G

Graph a value for each cluster across different SoCs/CPUs.

Arguments:

  • Add a SoC: SoC-1:soc1/results.json
  • Specify the value to graph: load/value (load is idle/active)
  • Set a flag: +flagname (soccolor, minscl)

Example usage: ./cross_cpu_cluster_graph.py 835:results/p2/main/results.json 855:results/zf6/main/results.json 855+:results/rog2/main/results.json 765G:results/p5/new-final/results.json active/power_mean +soccolor +minscl

Unified cluster graph

Performance (iter/s) across 765G little, big, and prime clusters

Graph a value for each cluster within the same SoC/CPU.

Example usage: ./unified_cluster_graph.py results.json coremark_score

Unified cluster column

Extract a value for each cluster within the same SoC/CPU and write the results into a CSV file.

Example usage: ./unified_cluster_csv.py results.json coremark_score cm_scores.csv

Troubleshooting

Kernel panics on boot

If your kernel panics on boot, disable CONFIG_CPU_FREQ_STAT. If that causes the kernel to fail to compile, cherry-pick cpufreq: Fix build with stats disabled.

Results vary too much

Check kernel.log, pre- and post-bench interrupts, running processes, and cpufreq stats from the results directory to diagnose the issue.

It's still running after an hour

If you have a slow CPU with a lot of frequency steps, this is not entirely unreasonable.

I want to debug it while it's running

freqbench offers interactive debugging via SSH over virtual USB Ethernet; the device acts as a USB Ethernet adapter and exposes an SSH server on the internal network. This feature can be enabled with the USB_DEBUG option in config.sh. It is disabled by default to avoid unnecessary USB setup that may influence benchmark results, so keeping it enabled for a final benchmark run is not recommended.

CONFIG_USB_CONFIGFS_RNDIS must be enabled for this feature to work. If your kernel does not have or use configfs for USB configuration, it will not work regardless of whether you have the RNDIS function enabled.

Once it's enabled, connect your device to a computer over USB. You should see something like this in your kernel logs if you are running Linux:

[7064379.627645] usb 7-3: new high-speed USB device number 114 using xhci_hcd
[7064379.772208] usb 7-3: New USB device found, idVendor=0b05, idProduct=4daf, bcdDevice= 4.14
[7064379.772210] usb 7-3: New USB device strings: Mfr=1, Product=2, SerialNumber=3
[7064379.772211] usb 7-3: Product: Alpine GNU/Linux
[7064379.772211] usb 7-3: Manufacturer: Linux
[7064379.772212] usb 7-3: SerialNumber: ssh -o StrictHostKeyChecking=no -o UserKnownHostsFile=/dev/null [email protected]
[7064379.818904] rndis_host 7-3:1.0 usb0: register 'rndis_host' at usb-0000:47:00.1-3, RNDIS device, da:34:ab:99:c5:81
[7064379.870018] rndis_host 7-3:1.0 enp71s0f1u3: renamed from usb0

Run the SSH command in the serial number field to open a shell to the device. The password is empty, so just press enter when asked to provide a password.

Note that the project description data, including the texts, logos, images, and/or trademarks, for each open source project belongs to its rightful owner. If you wish to add or remove any projects, please contact us at [email protected].