Institute of Computer Science Chair of Communication Networks Prof. Dr.-Ing. P. Tran-Gia Dynamic Bandwidth Allocation for Multiple Network Connections: Improving User QoE and Network Usage of YouTube in Mobile Broadband Florian Wamser, Thomas Zinner, Phuoc Tran-Gia University of Würzburg, Germany Jing Zhu Intel Labs, Hillsboro, OR, United States
Competing Applications at a Bottleneck Link YouTube Video Voice over IP Online Office Software Download Internet 2 2
Impact on Application Quality Content unaware networks Fair share with respect to QoS (throughput) Bulk data download performance: good YouTube quality: bad d a t a [ k B / s ] 400 300 200 100 bulk data (downloads) Throughput YouTube Influence on YouTube buffered playtime [s] 15 10 5 Desired progress Actual progress 0 0 20 40 60 time [s] 0 0 10 20 30 40 50 60 time [s] 3 3
Application-Aware Networking Tasks and objectives Integrating application needs in network resource management Add or re-allocate resources on demand Application and Network Monitoring 1. Application and network monitoring Collects information with high correlation to QoE Example: YouTube monitor (YoMo), browsing monitor, etc. 2. Decision entity Evaluates the information and decides about appropriate resource management action Decision Entity 3. Dynamic resource management Enforces resource management actions Example: resource allocation, scheduling, traffic prioritization, access technology selection, Dynamic Resource Management 4 4
Resource Management: Dual Connectivity of Devices More than just one transmission technology is available at current mobile devices Wi-Fi Communications Cellular Communications 5 5
Framework for Intelligent Bandwidth Aggregation Virtual access network (VAN) to aggregate multiple networks into single IP pipe Technical implementation: TCP/IP over UDP tunneling (mobile IP-like approach) Features of Intel s OTT VAN Configurable bandwidth aggregation for multiple networks (TCP) packet reordering (re-sequencing) Missing features Smart algorithms for dynamic offloading Application specific guidelines 6 6
Intel s OTT VAN Testbed: Hardware and Software Client Virtual network device provides tunneling functionality Access technologies Wi-Fi communications (limited to max. 2 Mbps) Mobile communications (limited to 4 Mbps) Server Tunnel endpoint Implements re-sequencing buffer Enforces resource management Application TCP/UDP IP UDP IP (Wi-Fi) UDP IP (3G) Mobile Access Network Wi-Fi Access Network Core Network Application TCP/UDP IP Internet 7 7
Resource Management Algorithms Adjust offload ratio between Wi-Fi and 3G cellular traffic, based on a required throughput Always use Wi-Fi and dynamically add 3G If current throughput < required throughput Increase 3G bandwidth If current throughput > required throughput Decrease 3G bandwidth 8 8
Resource Management Algorithms for Algorithm 1: Static Offloading Based on Video Request Defines required throughput based on requested video quality Detects uplink request by YouTube with DPI Algorithm 2: Dynamic Offloading Based on Buffer Estimation Constant monitoring of the buffer level Adaption of the required throughput based on the buffer level Algorithm 1 Req. throughput Time Algorithm 2 Req. throughput Time Algorithm 3: Burst-wise Offloading Based on Buffer Estimation Make use of the complete bandwidth until the buffer is filled Disables 3G link until the buffer gets low req. throughput = max. Algorithm 3 Time req. throughput = 0 WiFi-only 9 9
Time Series of Algorithm 1 Time series of one video with 1080p resolution Wi-Fi and 3G available Buffered time [s] Throughput [Mbps] 50 40 30 20 Static req. throughput = 6 Mbps Buffer filled 10 Request detected 0 0 50 100 150 200 250 300 Time [s] 6 5 4 Adaption to 3 required throughput No WiFi or 3G activity 2 1 Request detected 0 0 50 100 150 200 250 300 Time [s] 10 10
Time Series of Algorithm 2 Algorithm 2 dynamically adjusts required throughput according to video playback buffer Buffered time [s] Throughput [Mbps] 50 40 30 20 Buffer filled 10 Request detected 0 Low buffer 0 50 100 150 200 250 300 Time [s] 6 5 4 3 2 1 Request detected Throughput decreases 0 0 50 100 150 200 250 300 Time [s] 11 11
Comparison of Algorithm 1 and 2 Total 3G bandwidth consumed in MB 400 300 200 100 Consumption of 3G bandwidth 720p per 1000 seconds playtime 0 1.6 1.8 2 2.2 2.4 2.6 Average bit-rate in Mbps Total 3G bandwidth consumed in MB 400 300 200 100 Resolution based 1080p 2.5 3 3.5 4 4.5 5 Average bit-rate in Mbps Buffer based 720p 98 MB -64,3% 35 MB 1080p 321 MB -30,5% 223 MB Algorithm 1 (resolution based) Algorithm 2 (buffer based) 12 12
Comparison of Algorithm 1 and 2 Energy consumption per video in Joule 1400 1200 1000 800 600 400 Average amount of consumed energy 720p per 1000 seconds playtime 200 1.6 1.8 2 2.2 2.4 2.6 Average bit-rate in Mbps Energy consumption per video in Joule Resolution based 1600 1400 1200 1000 800 600 400 Buffer based 720p 1180 J -49,9% 591 J 1080p 1404 J -7,5% 1299 J J. Huang et al. A close examination of performance and power characteristics of 4g lte networks 1080p Resolution based Buffer based 200 2.5 3 3.5 4 4.5 5 Average bit-rate in Mbps 13 13
Time Series of Algorithm 3 Algorithm 3 activates 3G link in bursts Buffered time [s] Throughput [Mbps] 50 40 30 20 10 0 Request detected Full Buffer Low Buffer Full Buffer Low Buffer 0 50 100 150 200 250 300 Time [s] 6 5 4 3 Request 2 detected 1 3G link disabled 0 0 50 100 150 200 250 300 Time [s] 14 14
Comparison of Algorithm 2 and 3 Energy consumption per video in Joule 1400 1200 1000 800 600 400 Average amount of consumed energy 720p per 1000 seconds playtime 200 1.6 1.8 2 2.2 2.4 2.6 Average bit-rate in Mbps Dynamic buffer Energy consumption per video in Joule 1600 1400 1200 1000 800 600 400 Burst-wise buffer 720p 591 J 519 J 1080p 1299 J -8,9% 1183 J Dynamic buffer Burst-wise buffer 200 2.5 3 3.5 4 4.5 5 Average bit-rate in Mbps -12,2% 1080p 15 15
Conclusion and Outlook Contribution of the work Assessment and quantification of the benefits of cross-layer resource management on the example of YouTube Analysis of three application-aware algorithms which differ in complexity and impact on user and network Results of the evaluation The application-aware algorithms can enhance the QoE level for end users (if both networks provide enough resources) save costs in terms of energy & Cellular resources Future work Investigations on scalability of our approach and field trials with many users Providing a holistic resource allocation for popular applications with respect to their instaneneous needs 16 16
http://dl.acm.org/authorize?n71341 Florian Wamser, Thomas Zinner, Phuoc Tran-Gia and Jing Zhu Dynamic Bandwidth Allocation for Multiple Network Connections: Improving User QoE and Network Usage of YouTube in Mobile Broadband 17 17
http://dl.acm.org/authorize?n71209 Florian Wamser, Thomas Zinner, Lukas Iffländer, Phuoc Tran-Gia Demonstrating the Prospects of Dynamic Application-Aware Networking in a Home Environment 18 18
http://dl.acm.org/authorize?n71235 Steffen Gebert, David Hock, Thomas Zinner, Phuoc Tran-Gia (University of Würzburg); Marco Hoffmann, Michael Jarschel, Ernst-Dieter Schmidt (Nokia); Ralf-Peter Braun (Deutsche Telekom T-Labs), Christian Banse (Fraunhofer AISEC); Andreas Kopsel (BISDN) Demonstrating the Optimal Placement of Virtualized Cellular Network Functions in Case of Large Crowd Events 19 19