Skip to content

πŸ”‹πŸŽ― Thread-level, NUMA-aware energy attribution model for multi-tenancy

License

Notifications You must be signed in to change notification settings

HongyuHe/energat

Folders and files

NameName
Last commit message
Last commit date

Latest commit

Β 

History

2 Commits
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 

Repository files navigation

EnergAt πŸ”‹πŸŽ―

License: MIT version pypi Maintenance PRs Welcome build

EnergAt is a prototype implementation of the thread-level, NUMA-aware energy attribution model for multi-tenancy, as proposed in our paper. It offers precise tracking of the energy consumption of your software application, even when it is running alongside other jobs from different users.

Install

To use EnergAt, the following requirements have to be met:

  • Python 3.10
  • Linux with root permission
  • Intel RAPL with sysfs power capping interface

Note There are several ways of accessing RAPL meters on Linux (e.g., via model-specific registers or perf), but the current implementation only supports the powercap interface. Contributions are more than welcome!

Please run the following commands:

# * Configure the host system.
$ ./scripts/setup/init_system.sh
# * Use root permission for invoking energat with `sudo` later.
$ sudo python -m pip install energat

Usage

Warning Please always first estimate the static power of the machine by running the following command.

$ sudo energat -basepower

Python API

import psutil, time
from energat.tracer import EnergyTracer

def xyz():
    tmp = 20_000
    for _ in range(100): 
        time.sleep(0.1)
        tmp = tmp * 0.314

with EnergyTracer(psutil.Process().pid, output='xyz_energy') as tracer:
    xyz()

# * The traces of `xyz()` will be saved to ./xyz_energy.csv at this point.

Command line interface

First, check the system setup by running:

$ sudo energat -check

EnergAt @ Jul 03 10:30:27] INFO     | Socket count:        2
EnergAt @ Jul 03 10:30:27] INFO     | Host CPU times:      [334541.14, 334861.87]
EnergAt @ Jul 03 10:30:27] INFO     | Total NUMA memories: [32094.24, 32211.02]
EnergAt @ Jul 03 10:30:27] INFO     | RAPL domain ranges:  [262143.32885, 65712.999613]
EnergAt @ Jul 03 10:30:27] INFO     | System check passed!

Then, you can attach EnergAt to a running application for which you want to trace the energy consumption by providing its PID:

$ sudo energat -pid <PID>

Other command-line options include:

$ sudo energat [FLAGS]

Commands:
  --pid PID                PID of the target application
                           (default: -1)
  --name NAME              Name of the target application
  --check                  Check hardware support
                           (default: False)
  --basepower              Estimate static power
                           (default: False)

Configurations:
  --output OUTPUT          Output directory
                           (default: ./data/results)
  --basefile BASEFILE      File recording the baseline power
                           (default: ./data/baseline_power.json)
  --base_period BASE_PERIOD
                           Sampling period in seconds for baseline power estimation
                           (default: 2)
  --rapl_period RAPL_PERIOD
                           Sampling period in seconds for RAPL power meters
                           (default: 0.01)
  --interval INTERVAL      Interval in seconds between two power estimations
                           (default: 1)
  --gamma GAMMA            Non-linear scaling factor for CPU power
                           (default: 0.3)
  --delta DELTA            Non-linear scaling factor for DRAM power
                           (default: 0.2)
  --logging LOGGING        Logging interval in seconds (with `loglvl=debug` only)
                           (default: 1)
  --loglvl LOGLVL          Logging level (info/debug)
                           (default: debug)

Once the target application finishes, EnergAt will save the energy traces to the -output directory and exits. You can also stop the tracing by Ctrl+C, and EnergAt will still save your result before exiting.

Development

EnergAt has been heavily tested on a few dual- and single-socket machines on CloudLab.

To set up the development environment, please run the following commands.

# * Run the setup script.
./scripts/setup/init_cloudlab.sh
# * Install python dependencies.
pip install -r requirements.txt
# * Configure git hooks.
pip install pre-commit
pre-commit install
pre-commit autoupdate

Runtime Configuration

There are several ways by which you can change the configurations of EnergAt:

  1. Directly change the default parameters in configs/default.py.
  2. Provide your own config file in the configs/ directory and specify it in the command line, e.g.,:
$ sudo energat -config=./configs/your_config.py
  1. Overwrite the default parameters through command-line flags, e.g.,:
$ sudo energat -config.OUTPUT_DIR='./out' -config.DELTA_MEM=0.1

Contributing

Pull requests (PRs) are most welcome! Please follow the PR template: .github/pull_request_template.md.

Citation

If you find this tool useful, please cite our paper:

@inproceedings{hotcarbon2023energat,
  author = {Hè, Hongyu and Friedman, Michal and Rekatsinas, Theodoros},
  title = {EnergAt: Fine-Grained Energy Attribution for Multi-Tenancy},
  booktitle = {2nd Workshop on Sustainable Computer Systems (HotCarbon '23)},
  year = {2023},
  month = {July},
  day = {9},
  location = {Boston, MA, USA},
  publisher = {ACM},
  address = {New York, NY, USA},
  pages = {8},
  doi = {10.1145/3604930.3605716},
  url = {https://doi.org/10.1145/3604930.3605716}
}