1. DASH2M: Exploring HTTP/2 for Internet Streaming to Mobile Devices
Mengbai Xiao1, Viswanathan swaminathan2, sheng wei2,3, songqing chen1
1george mason university
2adobe research
3University of nebraska-lincoln

2. Streaming to Mobile Devices

3. HTTP Streaming
HTTP streaming is the most prevalent streaming method over Internet
The well-known implementations: Apple HLS, Adobe HDS and MPEG DASH
Qlty#1
Qlty#2
Qlty#3
Qlty#4
Video
HTTP request (Qlty#1)
Video segment (Qlty#1)

4. HTTP/2
HTTP/2 is the new HTTP protocol, featuring couples of advanced techniques
Server-initiated push
Stream termination
Etc.
HTTP/2 is replacing HTTP/1.1 and is expected to eventually fully replace its predecessor
Can HTTP streaming benefit from the protocol evolution?

5. Existing Schemes
Most pioneering studies investigating HTTP/2 for Internet streaming exploit the server-initiated push mechanism
Improve the user experience in HTTP streaming
Live latency [Wei et al. 2014, Huysegems et al. 2015]
Fast start [Cherif et al. 2015]
Not consider the power efficiency
Improve the power efficiency in HTTP streaming [Wei et al. 2015]
Fixed push number
Over-push video segments

6. Push Mechanism in HTTP/2 Streaming
K-push is an effective means to improve HTTP streaming performance
Client
Server
Client
Server
req seg 1
req seg 1
Request overhead elimination
Aggregated network transmissions
Degraded network adaptability
Over-pushed content
seg 1
seg 1
Push cycle
req seg 2 …
seg k
seg 2
req seg k+1 … …
(a) No-Push
(b) K-Push

7. DASH2M Design
DASH2M directs how to operate for the push cycles in a video streaming session
Push number determination per cycle
Playback
Stop watching here
Waste of power

8. Watching Energy per Segment
Watching energy per segment indicates the power efficiency in a push cycle
Achieve the lowest watching energy per segment
Calculate the expectation of the watching energy per segment for every K
K is constrained by current network condition
K=5
1 J
watched
not watched
1 J
2.5 J

9. Bitrate Selection per Segment
Objective: Maximize the cumulative bitrate of all involved segments in a push cycle
Integer linear programming problem
1080p
720p
360p
bandwidth
Playback
K=5

10. Linear Programming Constraints
Continuous playback: the playback buffer should not be consumed up during the time of segment retrievals
Smooth playback: the quality switch should not be abrupt
Maximum buffer size: the playback buffer should not overflow ✔️ ✖️ ✖️

11. Bandwidth Prediction and Push Cycle Termination
DASH2M predicts future bandwidth from two sources
Local measurement: weighted by the time distance
Server support (optional): allocated bandwidth on the server side
Terminate all upcoming transmissions of pushed segments
If the consumed network resources have significantly exceeded the predicted available network resources
Send the RST_STREAM to terminate all active pushing transmissions

12. Linear Programming Evaluation Setup
Experimental parameters
Bandwidth candidates (kbps): 60, 200, 600, 800
Bandwidth changing interval (s): 10
Video quality (kbps): 49, 217, 504, 752
Segment duration (s): 2
Streaming session duration (s): 40
Solver exit time (s): 0.5
m, the quality duration: 2, 3, 4
Maximum K: 5, 10, 15, 20
Average quality distance
The average bitrate level difference from the selected bitrate to the optimal bitrate

13. Linear Programming Performance

14. Rate Adaptation Evaluation Setup
An example video is selected to perform real streaming experiments
Video quality (kbps): 49, 217, 504, 752
Segment duration (s): 2
Compare with x-push schemes and the regular DASH scheme (HTTP/1.1)
x-push stands for the k-push scheme when K is set as x
The x candidates: 2, 5, 10, 20
Bandwidth candidates (kbps): 400, 600, 800, 1600

15. Rate Adaptation Performance
572 kbps
440 kbps
485 kbps
434 kbps
489 kbps
293 kbps

16. Push Termination Performance
647 kbps
561 kbps
586 kbps
512 kbps
644 kbps
357 kbps

17. Power Efficiency Evaluation Setup
Dataset: Apple HLS trace collected at the client side from Vuclip
~ 12 million video sessions: the session duration and the watched duration

18. Power Efficiency Evaluation Setup (Cont.)
The dataset is divided into two halves
One half is used to build the user watching portion distribution in the client
The other half is fed as the input of the simulation
A perl simulator is implemented to calculate the energy wasted by the over-push
The ratio of bandwidth to video quality: 10:1, 10:4, 10:7

19. Power Efficiency DASH2M DASH2M
Ratio of bandwidth to segment bitrate 10:1
Ratio of bandwidth to segment bitrate 10:7

20. Conclusion
We design and implement the DASH2M, which are based on the server-initiated push technique in HTTP/2, to improves the streaming experience in terms of power efficiency and user QoE
DASH2M is able to switch the video quality even inside a push cycle in a gentle manner without buffer starvation and buffer overflow
DASH2M approaches the most power efficient solution no matter how the bandwidth changes

21. Thank you!