GPOEO

GPOEO is a micro-intrusive GPU online energy optimization framework for iterative applications. We also implement ODPP [1] as a comparison.

[1] P. Zou, L. Ang, K. Barker, and R. Ge, “Indicator-directed dynamic power management for iterative workloads on gpu-accelerated systems,” in 2020 20th IEEE/ACM International Symposium on Cluster, Cloud and Internet Computing (CCGRID). IEEE, 2020, pp. 559-568.

1. ./EPOpt contains source code of the GPOEO and ODPP [1].

2. ./PerformanceMeasurement (PerfMeasure) is a NVIDIA GPU measurer for energy/power/utilities/clocks

Make GPOEO

Modify pathes of headers and libraries in ./EPOpt/makefile . cd ./EPOpt && mkdir ./build && cp makefile ./build cd ./build && make

Make PerfMeasure

Modify pathes of headers and libraries in ./PerformanceMeasurement/makefile . cd ./PerformanceMeasurement && mkdir ./build && cp makefile ./build cd ./build && make

Use GPOEO in python applications

GPOEO only has two APIs:

Begin(GPUID4CUDA, GPUID4NVML, RunMode, MeasureOutDir, ModelDir, TestPrefix)
End()

GPUID4CUDA: GPU ID used in CUDA environment.

GPUID4NVML: GPU ID queried with nvidia-smi and used to initialize CUPTI.

RunMode: "WORK" (run energy saving online); "MEASURE" (measure hardware performance counter metrics and other data for training multi-objective prediction models).

MeasureOutDir: measurement output file path.

ModelDir: the path of multi-objective prediction models.

TestPrefix: prefix name of one run.

The two APIs should be inserted at the beginning and end of the main python file respectively. As shown below:

from PyEPOpt import EPOpt

if __name__=="__main__":
    EPOpt.Begin(GPUID4CUDA, GPUID4NVML, RunMode, MeasureOutDir, ModelDir, TestPrefix)

    .....

    EPOpt.End()

Use ODPP [1] in python applications

ODPP can be implemented as a daemon. However, for the convenience of comparing GPOEO and ODPP, we also implement ODPP into the same form: two APIs.

ODPPBegin(GPUID4CUDA, GPUID4NVML, RunMode, MeasureOutDir, ModelDir, TestPrefix)
ODPPEnd()

GPUID4CUDA: GPU ID used in CUDA environment.

GPUID4NVML: GPU ID queried with nvidia-smi and used to initialize CUPTI.

RunMode: "ODPP" (run ODPP online).

MeasureOutDir: not used.

ModelDir: the path of ODPP models.

TestPrefix: prefix name of one run.

The two APIs should be inserted at the beginning and end of the main python file respectively. As shown below:

from ODPP import ODPPBegin, ODPPEnd

if __name__=="__main__":
    ODPPBegin(GPUID4CUDA, GPUID4NVML, RunMode, MeasureOutDir, ModelDir, TestPrefix)

    .....

    ODPPEnd()

Citation

If you use GPOEO in your research, please cite us as follows:

Wang F, Zhang W, Lai S, et al. Dynamic GPU energy optimization for machine learning training workloads[J]. IEEE Transactions on Parallel and Distributed Systems, 2021, 33(11): 2943-2954. https://github.com/ruixueqingyang/GPOEO

BibTex:

@article{GPOEO,
  title={Dynamic GPU energy optimization for machine learning training workloads},
  author={Wang, Farui and Zhang, Weizhe and Lai, Shichao and Hao, Meng and Wang, Zheng},
  journal={IEEE Transactions on Parallel and Distributed Systems},
  volume={33},
  number={11},
  pages={2943--2954},
  year={2021},
  publisher={IEEE}
}