GPGPU Performance and Power Modeling with DVFS

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 and X.-W. Chu, "GPGPU Performance Estimation with Core and Memory Frequency Scaling," IEEE International Conference on Parallel and Distributed Systems (ICPADS) 2018, Singapore, Dec 2018.[An extended journal version is under review.]

Citation

@inproceedings{Wang2018perf, 
    author={Q. {Wang} and X. {Chu}}, 
    booktitle={2018 IEEE 24th International Conference on Parallel and Distributed Systems (ICPADS)}, 
    title={GPGPU Performance Estimation with Core and Memory Frequency Scaling}, 
    year={2018}, 
    pages={417-424}, 
    month={Dec},
}

Content

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

Usage

Dependencies and prerequisites

• Python 2.7
• CUDA 9.0 or above
• NVIDIA GPU Driver (the latest version is recommended.)
• OS requirement: Windows 7/10, Ubuntu 16.04 or above, CentOS 6/7.
• Using "pip install -r requirements.txt" to install the required python libraries.

Data Collection

To use the performance/power models, one should first collect the needed performance counters and the groundtruth of kernel execution time and average runtime power. There are two configuration files that users should edit. One is called the benchmark-setting file, stored in configs/benchmarks/, which defines how the benchmarks run and repeat, what performance counters to collect and what frequencies are tested. The other is called the kernel-setting file, stored in configs/kernels/, which defines the set of tested GPU applications.

We provide some examples in those two folders. For the benchmark-setting file, the parameters contain:

[profile_control]

• iters: the number of repetitions for GPU kernels
• secs: the execution time that the GPU application should run by repeating the kernel, prior to "iters"
• cuda_device_id: the GPU index under CUDA runtime API
• nvIns_device_id: the GPU index under CUDA Driver, which might be different from "cuda_device_id"
• rest_time: the time interval (s) between two consecutive benchmarks
• power_sample_interval: the power sampling interval (ms)
• metrics: the list of profiling performance counters

[dvfs_control]

• coreF: the list of tested core frequencies
• memF: the list of tested memory frequencies
• powerState: the fixed power state controlled by the GPU driver

The kernel-setting file is composed by a list of elements, each of which defines one GPU application. The format is:

[(application name)]
args = ["(arguments)"]
kernels = ["(GPU kernel name)"]

Then one can run the benchmarks with the following command (The file name arguments have no suffix.)

python dvfs_benchmark.py --benchmark-setting (your benchmark-setting file) \
                         --kernel-setting (your kernel-setting file) \
                         --app-root (the folder where stores your applications)

By default, we provide the executable files of those GPU applications that occurs in our paper, and they are compiled under CUDA 10.0. One can also add their own applications and revise the kernel-setting file accordingly.

After the benchmarks are finished, one should find the logs stored in the folder /logs/(your benchmark-setting file)-(your kernel-setting file)

Data Extraction

Use the following command to extract the performance counters and the average kernel execution time from the logs (The file name arguments have no suffix.)

python gpuPerfExtracter.py --benchmark-setting (your benchmark-setting file) \
                           --kernel-setting (your kernel-setting file) 

Then the corresponding csv file will be generated under csvs/raw/.

Performance Modeling with DVFS

Now all the data are ready in the csv file. Use the following command to estimate the execution time of each kernel with DVFS and compare them with the groundtruth (The file name arguments have no suffix.)

python analytical.py --benchmark-setting (your benchmark-setting file) \
                     --kernel-setting (your kernel-setting file)

Then the program should output some results detailedly in csvs/analytical/.

Contact

Email: [email protected]

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

Welcome any suggestion or concern!