This tool builds a model to predict CPU power consumption from different CPU variables (Utilization, Frequency,...) using InfluxDB time series.
Before using this tool you must configure the InfluxDB server from which the metrics will be exported, as well as the timestamps of the time series you want to obtain from that server.
To modify your InfluxDB server simply modify the following code variables from cpu_power_seer.influxdb.influxdb_env.py:
INFLUXDB_URL = "influxdb-server:port"
INFLUXDB_TOKEN = "your-token"
INFLUXDB_ORG = "your-org"It is assumed that this server stores Glances and RAPL metrics in a proper format.
Note: To store this metrics properly see CPUPowerWatcher.
First of all, you can (and should) create a virtual environment by running:
python -m venv my_venv
source my_venv/bin/activate
To install and use this tool run:
pip install .
If you only want to install the project dependencies, run:
pip install -r requirements.txt
Finally, when you no longer want to use CPUPowerSeer, you can deactivate your virtual environment by running:
deactivate
usage: powerseer [-h] [-v] --vars VARS -t TRAIN_TIMESTAMPS [-a ACTUAL_TIMESTAMPS_LIST] [-p PREDICTION_METHOD] [-b BUCKET] [-o OUTPUT] [-n NAME]
CPU Power Modeling from Time Series.
options:
-h, --help show this help message and exit
-v, --verbose Increase output verbosity
--vars VARS Comma-separated list of variables to use in the regression model. Commonly known as predictor variables.
Supported values: user_load, system_load, wait_load, freq, sumfreq.
-t TRAIN_TIMESTAMPS, --train-timestamps TRAIN_TIMESTAMPS
File storing time series timestamps from train data in proper format. Check README.md to see timestamps proper format.
-a ACTUAL_TIMESTAMPS_LIST, --actual-timestamps-list ACTUAL_TIMESTAMPS_LIST
Comma-separated list of files storing time series timestamps from actual values of predictor variables and power to test
the model (in same format as train timestamps). If any file is specified train data will be split into train and test data.
-p PREDICTION_METHOD, --prediction-method PREDICTION_METHOD
Method used to predict CPU power consumption. By default is a polynomial regression. Supported methods:
polynomial Polynomial Regression with specified variables
freqwointeractionterms Custom Regression using user_load, system_load and freq or sumfreq
perceptron Multilayer Perceptron
svr Support Vector Regression
-b BUCKET, --bucket BUCKET
InfluxDB Bucket to retrieve data from. By default is 'public'.
-o OUTPUT, --output OUTPUT
Directory to save time series plots and results. By default is './out'.
-n NAME, --name NAME Name of the model. It is useful to generate models from different sets of experiments in an orderly manner. By default is 'General'Timestamps files must be stored in the following format:
<EXP-NAME> <TYPE-OF-EXPERIMENT> (CORES = <CORES>) START: <START-DATE>
<EXP-NAME> <TYPE-OF-EXPERIMENT> (CORES = <CORES>) STOP: <STOP-DATE>With the following meaning:
EXP-NAME: User desired name.TYPE-OF-EXPERIMENT: The type of experiment run during that period. It can take 3 values: STRESS-TEST if it's a stress test, IDLE if it's a period in which the CPU is idle and other if it's a period in which test data was obtained.CORES: Comma-separated list of cores used in the experiment.START-DATEandSTOP-DATE: Timestamp of the beginning or end of the experiment in UTC format%Y-%m-%d %H:%M:%S%z.
Example:
Group_P STRESS-TEST (CORES = 0,1) START: 2023-04-21 09:33:53+0000
Group_P STRESS-TEST (CORES = 0,1) STOP: 2023-04-21 09:35:54+0000
Group_P STRESS-TEST (CORES = 0,1,2,3) START: 2023-04-21 09:36:24+0000
Group_P STRESS-TEST (CORES = 0,1,2,3) STOP: 2023-04-21 09:38:24+0000
Group_P STRESS-TEST (CORES = 0,1,2,3,4,5) START: 2023-04-21 09:38:54+0000
Group_P STRESS-TEST (CORES = 0,1,2,3,4,5) STOP: 2023-04-21 09:40:54+0000
Group_P STRESS-TEST (CORES = 0,1,2,3,4,5,6,7) START: 2023-04-21 09:41:24+0000
Group_P STRESS-TEST (CORES = 0,1,2,3,4,5,6,7) STOP: 2023-04-21 09:43:25+0000
Group_P STRESS-TEST (CORES = 0,1,2,3,4,5,6,7,8,9) START: 2023-04-21 09:43:55+0000
Group_P STRESS-TEST (CORES = 0,1,2,3,4,5,6,7,8,9) STOP: 2023-04-21 09:45:56+0000
Group_P STRESS-TEST (CORES = 0,1,2,3,4,5,6,7,8,9,10,11) START: 2023-04-21 09:46:26+0000
Group_P STRESS-TEST (CORES = 0,1,2,3,4,5,6,7,8,9,10,11) STOP: 2023-04-21 09:48:26+0000
Group_P STRESS-TEST (CORES = 0,1,2,3,4,5,6,7,8,9,10,11,12,13) START: 2023-04-21 09:48:56+0000
Group_P STRESS-TEST (CORES = 0,1,2,3,4,5,6,7,8,9,10,11,12,13) STOP: 2023-04-21 09:50:57+0000Note: To obtain timestamps files in proper format see CPUPowerWatcher.
Output will be stored in the specified directory (-o option) or './out' by default. In the output directory you will find 2 subdirectories, train and test, which contains train time series and test/predictions time series along with their results, respectively. The directory tree will have the following appearance:
out
|
├─── train
| ├─── <MODEL-NAME>-temperature-data.png Temperature train time series
| └─── <MODEL-NAME>-train-data.png Model variables train time series
|
|
└─── test
├─── <MODEL-NAME>-summary.out Summary of the results obtained with all benchmarks.
└─── <BENCHMARK>
├─── <MODEL-NAME>-results.out Benchmark results
├─── img
| ├─── <MODEL-NAME>-results.png Expected VS Predicted Points plot
| └─── <MODEL-NAME>-predictions.png Predicted time series
|
├─── <THREADS[0]>
| ├─── <MODEL-NAME>-results.png Benchmark results with <THREADS[0]> threads
| └─── img
| ├─── <MODEL-NAME>-results.png Expected VS Predicted Points plot with <THREADS[0]> threads
| └─── <MODEL-NAME>-predictions.png Predicted time series with <THREADS[0]> threads
├─── ...
|
└─── <THREADS[n]>There will be one subdirectory in benchmark directory for each number of threads used with this benchmark.
Note: Don't forget to specify the cores in the timestamps file because CPUPowerSeer will infer the number of threads/cores used from these files.