1. Optical Networks & Smart Grid Lab
Nonlinear Auto-Regressive Neural Network Model for Forecasting Hi-Def H.265 Video Traffic Over Ethernet Passive Optical Networks
Collin J. Daly, David L. Moore, Rami J. Haddad
Department of Electrical Engineering
Georgia Southern University

2. Outline Motivation Discussion of Dataset Proposed Prediction Models
Simulation Parameters
Discussion of Results
Conclusion

3. Motivation Netflix, Hulu, YouTube
80% of all data will be video, 2 zettabytes
5 million years of video each month in 2019
Why Compress?
4K video contains 3840 x 2160 pixels per frame = 8,294,400 pixels
8 bits per pixel for color (low estimate) = 66,355,200 bits/frame
30 frames per second = 1,990,656,000 = 1.99Gb/s
Ethernet IP networks are expanding, but not as quickly as utilization.

4. H.264 vs H.265
H.264 is current standard, adopted 2003 while H.265 adopted 2013.
H.265 compresses video to a smaller size than what H.264 would.
The same size file yields higher resolution in H.265.
However, the higher compression ratio results in high variability in the video stream bitrate, which is detrimental to networks (Bandwidth Allocation).

5. EPON Basics Ethernet Passive Optical Network
Fiber link from ISP to end users
No active amplification
Lower implementation costs
Regulated network

6. EPON Architecture

7. Proposed Prediction Model
Nonlinear Auto-Regressive (NAR) Neural Network

8. Dataset – Video Traces Arizona State University Video Trace LibraryP
Arizona State University Video Trace Library
Video traces vs regular video files
All videos use a G24B7 group of pictures (GOP) format
Quantization levels: 20, 25, 30, 35, 40, 45
Higher compression lowers quality

9. Video Traces Video Traces Utilized: Harry Potter Finding Neverland
Speed
Lake House

10. Finding Neverland
Video Delay

11. Finding Neverland
Video Delay

12. Finding Neverland
Data Delay

13. Finding Neverland
Data Delay

14. EPON Simulation Results

15. EPON Simulation Results

16. Conclusion
Data delay increases while overall video delay decreases when predicted grants are issued
Prediction is harmful at low quantization levels in an EPON environment
Prediction is most beneficial when the network is nearing traffic saturation

17. Questions