Energy-aware Non-Preemptive Task Scheduling with Deadline Constraint in DVFS-enabled Heterogeneous Clusters

This repository contains the code for modeling/benchmarking NVIDIA GPU performance and power with dynamic voltage and frequency scaling. The relevant papers are as follows:

• Q. Wang, X. Mei, H. Liu, Y.-W. Leung, Z. Li and X. Chu, "Energy-aware Non-Preemptive Task Scheduling with Deadline Constraint in DVFS-enabled Heterogeneous Clusters," IEEE Transactions on Parallel and Distributed Systems, accepted. (preprint)

Citation

@article{mei2021energy,
  title={Energy-aware Task Scheduling with Deadline Constraint in DVFS-enabled Heterogeneous Clusters},
  author={Wang, Qiang and Mei, Xinxin and Liu, Hai and Leung, Yiu-Wing and Li, Zongpeng and Chu, Xiaowen},
  journal={arXiv preprint arXiv:2104.00486},
  year={2021}
}

Content

1. Introduction
2. Usage
3. Results
4. Contacts

Usage

Dependencies and prerequisites

• Python 3.6+
• CUDA 10.0 or above
• NVIDIA GPU Driver (the latest version is recommended.)
• OS requirement: Windows 7/10/11, Ubuntu 16.04 or above, CentOS 6/7.
• Using "pip install -r requirements.txt" to install the required python libraries.

Tips: If you are using Windows, we recommend you to use WSL to install a linux sub-system (e.g., Ubuntu), and run the scripts.

Performance Modeling with DVFS

The GPU configurations are as follows. We collect the performance and power data of two GPU devices.

GPU	Base Core Frequency (MHz)	Base Memory Frequency (MHz)
GTX 2070 SUPER	1880	6300
GTX 1080 Ti	1800	5000

1. Firstly please set the GPU variables in cluster.py, including:

GPU_NAME = 'gtx2070s'
CORE_BASE = 1880
MEM_BASE = 6300

The benchmarking results are stored in ./csvs, one csv file per GPU.

2. Fit the parameters of performance and power for each benchmark.

python model.py

Then the fitting results will be printed out and "apps.pkl" which stores the results will be generated.

Simulation

1. Generate the task set.

python gen_task.py ${type} ${util} ${mode}

${type} can be "offline" or "online". ${util} is a float number to define the average GPU utilization. ${mode} can be "synthetic" or "real". For example, to generate an offline task set with U=1.0, in which the tasks are from the real benchmarking data, one can run

python gen_task.py offline 1.0 real

The task set file will be generated in the folder ./job_configs.

2. Run the simulation for a task set.

python main.py ${type}-${util}-{mode} ${algo}-${dvfs}-${theta} ${gpn}

${type} can be "offline" or "online". ${util} is a float number to define the average GPU utilization. ${mode} can be "synthetic" or "real". ${algo} indicates the algorithm chosen to schedule the tasks. ${dvfs} can be "on" or "off", indicating that if the dvfs optimization is activated. ${gpn} means gpus per node, and is an integer. For example, one can run

python main.py offline-0.4-synthetic edf+spt-on-1.0 8

More example scripts can be found in ./scripts.

Real Experiments

Contact

Email: [email protected]

Personal Website: https://blackjack2015.github.io

Welcome any suggestion or concern!