CPUPowerWatcher provides a framework to automate CPU monitoring and time series storage while running workloads. The main purpose of this framework is to generate and collect datasets for the training and testing phases of power consumption models. It is recommended to use the 'stress-system' workload for the training phase, as it allows the tuning of different stress parameters such as the stress pattern, time under stress or the type of operations executed on the CPU, among others.
The system must meet the following requirements:
| Requirement type | Details | Mandatory |
|---|---|---|
| Operating system💻 | Bash shell support | ✅ |
| Virtualization engine ⚙️ | Docker or Apptainer (formerly Singularity) installed | ✅ |
| Virtualization engine ⚙️ | Ability to execute the virtualization engine with sudo privileges |
✅ |
| Filesystem access 📋 | /proc/stat must be available (CPU and system statistics) |
✅ |
| Filesystem access 📋 | /sys/devices/system/cpu must be available (CPU topology and information) |
✅ |
| Filesystem access 📋 | /sys/class/thermal available (for temperature monitoring) |
❌ |
| MSR access 🗂️ | /dev/cpu/[cpu_id]/msr available (to read CPU voltage MSR) |
❌ |
Clone this repository along with the required submodules:
git clone --recurse-submodules https://github.com/TomeMD/CPUPowerWatcher.gitOr update submodules if you have already cloned the repository:
git submodule update --init --recursiveThen, you can run CPUPowerWatcher just using run.sh.
| Metric | Description |
|---|---|
| 🔋 CPU power | Real-time CPU power consumption measured with RAPL_PKG |
| 📊 CPU usage | User, system and iowait CPU load |
| ⏱️ CPU frequency | Current operating frequency per core |
| 💤 CPU C-States | CPU sleep and idle state information |
| ⚡️ CPU voltage | Core voltage readings from CPU MSRs (vCore) |
| Workload | Description |
|---|---|
| NAS Parallel Benchmarks (NPB) | Kernels IS, FT, CG, MG, BT, BT I/O, scaling cores 1, 2, 4, ... to maximum |
| Fio | I/O benchmark scaling cores 1, 2, 4, ... to maximum |
| SMusket | DNA error correction using Spark Standalone scaling cores 1, 2, 4, ... to maximum |
| stress-system | Run stress-ng using different sets of cores (e.g., 1st cores 1,3,5, then cores 2,4,7,8) |
| Sysbench | General system benchmark (deprecated) |
| Geekbench | Cross-platform benchmark suite (deprecated) |
For more detail use the --help option:
~$ ./run.sh --help
Usage: run.sh [OPTIONS]
-v, --os-virt Technology for OS-level virtualization. [Default: apptainer]
docker
apptainer
-i, --influxdb-host InfluxDB host to send metrics. [Default: montoxo.des.udc.es]
-b, --influxdb-bucket InfluxDB bucket to store metrics. [Default: public]
-o, --output <dir> Directory (absolute path) to store log files. [Default: ./log]
-w, --workload Workload to stress the system with. [Default: stress-system]
npb Run NPB kenerls.
sysbench Run Sysbench kernels.
geekbench Run Geekbench kenerls.
fio Run fio to make random reads/writes over specified target with
different numbers of threads.
--fio-target Directory to make random reads/writes. [Default: /tmp/fio]
spark Run Spark-based DNA error correction algorithm (SMusket) using
Spark Standalone.
--spark-data-dir Directory to store Spark temporary files and Spark Smusket input.
Input must be a FASTQ file named "input.fastq".
stress-system Run stress tests using stress-system tool and different core
distributions. In addition, the tool can follow different patterns
to gradually use all cores from core distribution. Options:
--stressors Comma-separated list of stressors to run with stress-system.
[Default: cpu]
--core-distributions Comma-separated list of core distributions, which represent the order
followed by the tool when selecting cores to achieve a desired CPU load level.
The supported core distributions depend on the CPU topology.
[Default: all]
--stress-pattern Pattern followed to get groups of cores from the core distribution.
Patterns: stairs-up, stairs-down, zigzag, uniform or udrt [Default: stairs-up]
--stress-time Time (in seconds) under stress for each group of cores.
[Default: 120]
--idle-time Time (in seconds) to keep the CPU idle between one group of
cores and the next one. [Default: 20]
--stress-load-types Comma-separated list of types of load to stress the CPU.
Used together with CPU stressor. [Default: all]
--stress-extra-options Comma-separated list of other stress-ng options specified
in key=value format.
--base Get base measurements before tests to have idle consumption and overhead metrics.
--add-io-noise Run fio to make random reads/writes over specified target while running the specified
workload. Use --fio-target to specify target directory. This option is not compatible with
fio tests.
--custom-tests <file> Use custom tests file to create custom lists of cores to stress.
[Default: ./bin/test/stress-tests-custom.sh]
-h, --help Show this help and exit