Skip to content

Latest commit

 

History

History

PCS-2024

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
metadata
metadata
 
 
metrics
metrics
 
 
model
model
 
 
rapl-tools
rapl-tools
 
 
src
src
 
 
tek_power
tek_power
 
 
.DS_Store
.DS_Store
 
 
Overview.png
Overview.png
 
 
README.md
README.md
 
 
requirements.txt
requirements.txt
 
 

README.md

Measuring Energy, Quality, and Rate in Video Compression

This repository contains data and code from the following paper: A. Katsenou, X. Wang, D. Schien, and D. Bull, "Comparative Study of Hardware and Software Power Measurements in Video Compression", accepted in PCS 2024.

The paper presents a comprehensive study of power measurement techniques in video compression, comparing the use of hardware and software power meters. An experimental methodology to ensure reliability of measurements is introduced. Key findings demonstrate the high correlation of hardware and software based energy measurements for two video codecs across different spatial and temporal resolutions at a lower computational overhead.

Experiment Pipeline

Overview

The figure shows the basic experimental pipeline.

Preprocessing

  • Convert MKV to YUV
  • Read Bitrate for Raw Video
  • Downsample 4K to 1080p and 720p
  • Hardware Power Measurement:
    • 2 duplicate times for 4K videos (this can be adapted, 2 is the usual number)
    • 12 duplicate times for 1080p and 720p videos (this can be adapted, 12 is the usual number)
  • Software Power Measurement:
    • No duplications required for 4K and 1080p videos due to the high sampling rate
    • 2 duplicate times for 720p videos (this can be adapted, 2 is the usual number)
  • Read Bitrate for Raw Duplicate Video

Encoding

  • Encode at different QP levels: [10, 20, 30, 40, 50]
  • Hardware Power Measurement: Repeat X times (these reps have been tested for conformance with a t-test, 10 is the usual number)
  • Software Power Experiment: No repeat time (because of the high sampling rate)
  • Read Bitrate for Encoded Video
  • Calculate Encoding Energy Consumption

Decoding

  • Hardware Power Measurement: Repeat X times (these reps have been tested for conformance with a t-test, 10 is the usual number)
  • Software Power Experiment: Different repetitions based on resolution
  • Calculate Decoding Energy Consumption

Note: During decoding no data are stored locally.

Measuring Quality

  • Upsample Decoded Downsampled Videos to 4K (native spatial resolution)
  • Calculate PSNR
  • Calculate VMAF

Energy Consumption

  • Measure Time and Get Power Log
  • Calculate Energy for Each Coding Job (Encoding/Decoding)

Note: In the case of hardware power measurements, the alignment of the timestamps of the encoding/decoding process with the power logging is important.

Main Workflow

  • run_codec_metrics.py: Main workflow

4K Video Processing

Hardware Power Measurement

  • run_codec_metrics_downsample_4k_repeat_duplicate.py: Downsample and duplicate (repeat time with t-distribution)
  • run_codec_metrics_nodownsample_4k_repeat_duplicate.py: No downsample and duplicate (repeat time with t-distribution)

Software Power Measurement

  • run_codec_metrics_downsample_4k_duplicate_rapl.py: Downsample and duplicate
  • run_codec_metrics_nodownsample_4k_duplicate_rapl.py: No downsample and duplicate

Deployment

  • Deploying rapl-tools on Ubuntu

    1. Introduction:

      • Tools for monitoring CPU power with Intel's RAPL interface.
      • Original GitHub Repository: kentcz/rapl-tools.

    2. Building the Tool:

      • The tool is built with Gnu Make. G++ version 4.7 or higher is required: 
        make

    3. Setup on Ubuntu:

      • Install and load the msr module:

        sudo apt-get install msr-tools
sudo modprobe msr

      • Set permissions:

        sudo chmod o+rw /dev/cpu/0/msr

Requirement

# Create and activate a Conda environment
conda create -n videoenc pip python=3.10
conda activate videoenc

# Install dependencies
pip install -r requirements.txt

# Deactivate the Conda environment
conda deactivate

4K Test Sequences:

  • Download some original videos from YouTube-UGC Dataset [Yilin Wang, 2019]: UGC Dataset Google Cloud Storage

    1. Find files with gsutil ls gs://ugc-dataset/original_videos/
    2. Download desired files. For example: 
      gsutil -m cp -r gs://ugc-dataset/original_videos/TelevisionClip/1080P/TelevisionClip_1080P-68c6.mkv .
      Make sure to recreate folder structure /
    3. Keep the 'src/logs' folder

The video sequences used in this paper are:

  • Gaming_2160P-2dc4.mkv
  • HDR_2160P-06ae.mkv
  • Sports_2160P-0455.mkv
  • Vlog_2160P-030a.mkv
  • Animation_2160P-41dc,mkv
  • Gaming_2160P-2cb1.mkv
  • Gaming_2160P-3002.mkv
  • Gaming_2160P-348d.mkv
  • Gaming_2160P-34a1.mkv
  • Sports_2160P-086d.mkv
  • Sports_2160P-1261.mkv
  • Sports_2160P-1733.mkv
  • Sports_2160P-1b70.mkv