Application-header based traffic engineering for ABR video streaming*

Presented By: Divya Bhat

*Presented as a demo at LCN 2018: Application-based QoS support with P4 and OpenFlow: A demonstration using Chameleon -

Divyashri Bhat, Jason Anderson, Paul Ruth, Michael Zink and Kate Keahey

1

Video Streaming• Killer Application

• Over 71% of downstream Internet

traffic at peak hours in North America

is video.

• Predicted to increase to 80% by 2020

• Live, Video-on-Demand and User

Generated

• Challenges

• High bandwidth

• Low delay

• Various encoding formats – Real Time

• and many more…

2

Adaptive Bit-Rate (ABR) Video Streaming

• Single Video split into multiple segments in various qualities• Sequentially request video segments

• On-demand and Live – Dynamic Adaptive Streaming over HTTP (DASH)

• Download Buffer – stores segments in queue for playback3

ABR Streaming – QoE Metrics• Average Quality Bitrate

• Segment download rate

• Number of Quality Changes• Quality gaps

• Magnitude of Quality Changes

• Rebuffering Ratio• Segment transfer time

4

Image courtesy: https://www.tipard.com/

ABR Segment retransmissions using HTTP/2

• [1] Wang, Cong, Divyashri Bhat, Amr Rizk, and Michael Zink. "Design and analysis of QoE-aware quality adaptation for DASH: a spectrum-based approach." ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM) 13, no. 3s (2017): 45

• [2] Divyashri Bhat, Rajvardhan Deshmukh, and Michael Zink. 2018. Improving QoE of ABR Streaming Sessions through QUIC Retransmissions. In 2018 ACM Multimedia Conference on Multimedia Conference (MM '18). ACM, New York, NY, USA, 1616-1624.

originaltransmission

Quality level

151buffer in # of

segments2 3

1

2

3

4

4

5 6 7 8 9 10 11 12 13 14

5

6

7

8

1

re-transmission

2

3 4

Number indicates priority of segments for retransmission

Original gap(1 segment length).Need to retransmit

segment 3 (8) at quality level 6 to closegap.

Original gap(2 segment lengths).

Need to retransmit segments 13 & 14 at quality level 4 to close gap

0 50 100 150 200 250 3000

5

10

15

time [s]

rate

[Mbp

s] o

r buf

fer s

ize

[seg

]

seg. DL ratequality bitratebuffer fillcross trafficDASH traffic

5

Analysis of Quality Gaps - Akamai

[2] Divyashri Bhat, Rajvardhan Deshmukh, and Michael Zink. 2018. Improving QoE of ABR Streaming Sessions through QUIC Retransmissions. In 2018 ACM Multimedia Conference on Multimedia Conference (MM '18). ACM, New York, NY, USA, 1616-1624.

• Akamai - world’s largest CDN provider

• Delivers 15%-30% of global Internet traffic

• 3-day trace in June 2014• 36.19% of sessions (~5 million)

have at least one quality gap• ~10% sessions experience quality

changes higher than 40 per session

6

Network Protocols

7

Software-Defined Networks

8

HTTP/2 COMMON FRAME HEADER

Programmable Protocol Independent Packet Processors (P4)OpenFlow

Image Courtesy:https://queue.acm.org/detail.cfm?id=2555617https://opencord.org/https://www.opennetworking.org/

How can SDN be used to improve QoE of ABR Streaming?

9

System Architecture

Control Monitoring

EDGE

EDGE CORE

EDGE

ORCHESTRATION AND

VISUALIZATION

SOUTHBOUND

NORTHBOUND

Data Plane

Control Plane

Stats Reply

Stats Request

Install Flow

Request Connection

Send Network Information

Request Network Information

Content Server

Switch

Control Monitoring

Cellular

Home

10

Q-in-Q VLAN tagging

11

Ethernet II Frame (untagged)

Destination MAC

(6 bytes)

Source MAC

(6 bytes)

EtherType (2 bytes) Payload

Destination MAC

(6 bytes)

Source MAC

(6 bytes)EtherType (2 bytes) Payload

Destination MAC

(6 bytes)

Source MAC

(6 bytes)

EtherType (2 bytes) Payload

802.1Q Header(4 bytes)

TPID 0x8100

VLAN Priority

CFI VLAN ID

VLAN tagged (802.1Q)

Q-in-Q VLAN tagged (802.1ad)

TPID 0x8100

VLAN Priority CFI VLAN ID

802.1Q Header(4 bytes)TPID

0x8100VLAN Priority

CFI VLAN ID

802.1Q Header(4 bytes)

C-TAG

S-TAG

Experimental Setup (BYOC)

Compute and Storage• All instances – bare-metal machines with

Ubuntu 16.04• Client – Python hyper library for HTTP/2• Server – Apache2• SDN Controller – RYU• Cross Traffic – Iperf3

FAST PATH

12

Networking• Edge: P4 – Software switch BMV2• Core: Corsa switches at TACC and UC• Circuit1&2 – AL2S by Internet2 - 10Gigabit• 802.1Q tag – HTTP/2 Stream ID

Fast Path for ABR Retransmissions• Baseline – Single path (Circuit1)

for original and retransmitted segment

• Performance Improvement over Baseline

• Loss drastically affects TCP in the form of retransmissions

• Fast path for retransmission can provide up to ~50% reduction in download time and ~70% improvement in average download rate

13

Conclusion• Prototype of P4 at the edge and OpenFlow at the core (BYOC) – Two paths

(Fast and Slow) between Austin, TX and Chicago, IL to provide better QoE

• Use of Fast Path for retransmissions significantly reduces segment transfer timeand simultaneously increases download rate

• ABR Streaming clients can retransmit in higher quality over Fast Path toincrease QoE

• https://www.chameleoncloud.org/blog/2018/09/20/application-based-qos-support-p4-and-openflow/

14

• For popular videos, retransmissions can be made once and reused by caching (CDNs)

• Extended to other applications such as GridFTP, QUIC (Quick UDP Interconnections), etc.

Future Scope

Repeatability with Jupyter• JupyterLab - “JupyterLab is an interactive development

environment for working with notebooks, code, and data.”• IPython kernels for executing notebooks• Orchestrate and Visualize from single vantage point• Chameleon and Jupyter

1. Launch Docker container with JupyterLab in a Chameleon instance

2. Create Notebook3. Run Experiment4. Visualize5. Repeat and Share

• Demo: 6:00pm today15