GreenBag: Energy-efficient Bandwidth Aggregation For Real-time Streaming in Heterogeneous Mobile Wireless Networks

Transcription

1 GreenBag: Energy-efficient Bandwidth Aggregation For Real-time Streaming in Heterogeneous Mobile Wireless Networks Duc Hoang Bui, Kilho Lee, Sangeun Oh, Insik Shin Dept. of Computer Science KAIST, South Korea Hyojeong Shin Dept. of ECE Duke University, USA Honguk Woo, Daehyun Ban Software R&D Center Samsung Electronics, South Korea

2 Mobile Multimedia Streaming Popularity Global Mobile Video Traffic Trend 66% 60% 63% 51% 54% 57% [Cisco Visual Networking Index] 2

3 Motivation High quality video streaming requires high bandwidth Streaming ultra-hd 4K videos: 19 Mbps [Youtube] Recording 1080p full-hd videos on smartphones: Mbps 3

4 Motivation High quality video streaming requires high bandwidth Streaming ultra-hd 4K videos: 19 Mbps [Youtube] Recording 1080p full-hd videos on smartphones: Mbps 19 Mbps 4

5 Motivation High quality video streaming requires high bandwidth Streaming ultra-hd 4K videos: 19 Mbps [Youtube] Recording 1080p full-hd videos on smartphones: Mbps LTE service download speeds in US: 4-13 Mbps [TechHive13] 19 Mbps 13 Mbps 5

6 Motivation High quality video streaming requires high bandwidth Streaming ultra-hd 4K videos: 19 Mbps [Youtube] Recording 1080p full-hd videos on smartphones: Mbps LTE service download speeds in US: 4-13 Mbps [TechHive13] 10 Mbps 19 Mbps 13 Mbps 6

7 Motivation High quality video streaming requires high bandwidth Streaming ultra-hd 4K videos: 19 Mbps [Youtube] Recording 1080p full-hd videos on smartphones: Mbps LTE service Hard download to provide speeds QoS in US: over 4-13 a Mbps single [TechHive13] link 10 Mbps 19 Mbps 13 Mbps 7

8 Opportunities Mobile bandwidth aggregation Multiple network interfaces on smartphones: 3G/LTE and WiFi Create one logical link via two physical links LTE provides high bandwidth comparable to WiFi 10 Mbps 23 Mbps 19 Mbps 13 Mbps 8

9 Why Bandwidth Aggregation? 9

10 Why Bandwidth Aggregation? LTE Thrput. WiFi Thrput. LTE Thrput. Interruption WiFi Thrput. Interruption LTE Thrput. WiFi Thrput. Interruption 10

11 Why Bandwidth Aggregation? 11

12 QoS Requirements in Video Streaming A video consists of multiple frames Video file 12

13 QoS Requirements in Video Streaming A video consists of multiple frames Video frame Video file 13

14 QoS Requirements in Video Streaming A video consists of multiple frames Each frame has its own deadline d i Video frame Deadline of frame i th Video file d 1 d 2 d 3 d 4 d 5 14

15 QoS Requirements in Video Streaming A video consists of multiple frames Each frame has its own deadline Video file Frame 5 Frame 4 d i Deadline of frame i th Frame 3 Frame 2 Frame 1 d 1 d 2 d 3 d 4 d 5 Time 15

16 QoS Requirements in Video Streaming A video consists of multiple frames Each frame has its own deadline Video file Frame 5 Frame 4 d i Deadline of frame i th Frame 3 Frame 2 Frame 1 a 3 Need amount of data a 3 for frame 3 Display frame 3 at deadline d 3 d 1 d 2 d 3 d 4 d 5 Time 16

17 QoS Requirements in Video Streaming Hard to obtain deadline of each frame in the encoded video Approximation Video file Frame 5 Frame 4 d i Deadline of frame i th Frame 3 Frame 2 Frame 1 d 1 d 2 d 3 d 4 d 5 Time 17

18 QoS Requirements in Video Streaming Hard to obtain deadline of each frame in the encoded video Approximation: Linear deadlines of bits GangnamStyle.mp4 [YouTube] 18

19 QoS Requirements in Video Streaming Hard to obtain deadline of each frame in the encoded video Approximation: Linear deadlines of bits Video file Frame 5 Frame 4 Frame 3 d i Deadline of frame i th Deadlines of bits Frame 2 Frame 1 d 1 d 2 d 3 d 4 d 5 Time 19

20 Downloaded Data (1/3) Video playback decodes data in order Not all downloaded data is in order In-order data Out-of-order data All downloaded data In-order data Out-of-order data File Video playback Downloaded portion Undownloaded portion 20

21 Downloaded Data (2/3) QoS satisfied: In-order data satisfies QoS requirement Downloaded Data File size All downloaded data QoS requirement Out-of-order data In-order data Time 21

22 Downloaded Data (2/3) QoS satisfied: In-order data satisfies QoS requirement Downloaded Data File size All downloaded data QoS requirement Out-of-order data In-order data All deadlines satisfied Time 22

23 Downloaded Data (2/3) QoS satisfied: In-order data satisfies QoS requirement Downloaded Data File size QoS requirement Decoded data All deadlines satisfied Time 23

24 Downloaded Data (3/3) Out-of-order data can cause deadline misses Same all-downloaded-data line, lots of out-of-order data Downloaded Data File size QoS requirement All downloaded data Out-of-order data In-order data Deadline miss Time 24

25 Downloaded Data (3/3) Out-of-order data can cause deadline misses Same all-downloaded-data line, lots of out-of-order data Downloaded Data File size QoS requirement All downloaded data Decoded data Deadline miss Time 25

26 Bandwidth Aggregation Divide the file downloading into the two different interfaces Divide the file into small segments and assign each segment to one of the links for downloading 26

27 Bandwidth Aggregation Divide the file downloading into the two different interfaces Divide the file into small segments and assign each segment to one of the links for downloading File 27

28 Bandwidth Aggregation Divide the file downloading into the two different interfaces Divide the file into small segments and assign each segment to one of the links for downloading File 28

29 Bandwidth Aggregation Divide the file downloading into the two different interfaces Divide the file into small segments and assign each segment to one of the links for downloading File LTE WiFi WiFi LTE LTE WiFi LTE WiFi WiFi WiFi 29

30 Bandwidth Aggregation Divide the file downloading into the two different interfaces Divide the file into small segments and assign each segment to one of the links for downloading Two basic problems Segment size decision Segment channel assignment File LTE WiFi WiFi LTE LTE WiFi LTE WiFi WiFi WiFi 30

31 File Bandwidth Aggregation Divide the file downloading into the two different interfaces Divide the file into small segments and assign each segment to one of the links for downloading Two basic problems Segment size decision Segment channel assignment Major issue: out-of-order delivery Inherent issue when using multiple links Out-of-order segment LTE WiFi WiFi LTE LTE WiFi LTE WiFi WiFi WiFi Downloaded portion Downloading in progress 31

32 Challenge: out-of-order data Out-of-order data can cause playback interruptions Longer initial start time Interruptions during playback 32

33 Challenge: Link heterogeneity Requires load-balancing Optimal case: Two links finish segments at the same time Places Bandwidth heterogeneity in different places 33

34 Challenge: Bandwidth fluctuation Hard to have accurate prediction of network condition Recovery: replan downloading when prediction is wrong LTE bandwidth fluctuation 34

35 System Goals Goal: mobile bandwidth aggregation system for real-time multimedia streaming Easy to deploy No change to the existing Internet infrastructure and servers Energy efficient subject to QoS satisfaction Mobile devices are battery-powered Significant energy usage from the use of multiple network interfaces Yet, the user experience is more important 35

36 Problem Statement Support the QoS requirements of real-time video streaming in the most energy-efficient way Address the two basic problems Segment size decision Segment channel assignment 0 1 ii NN 1 ssssssss ccccccccccccc aaaaaaaaaaaaaaaaaaaa 36

37 Problem Statement Multi-objective optimization problem as a lexicographic optimization Considering QoS satisfaction more important than energy saving Problem #1: minimize playback time for QoS satisfaction Problem #2: minimize energy consumption subject to results in problem #1 0 1 ii NN 1 ssssssss ccccccccccccc aaaaaaaaaaaaaaaaaaaa 37

38 Bandwidth Aggregation Techniques Packet scheduler on network proxies EDPF [TMC06]: uses Earliest Delivery Path First packet scheduler Transport layer MPTCP [IFIPNet11]: extends TCP to support multiple paths Application layer MultiNets [RTAS12]: provides seamless switching between 3G and WiFi 38

39 Bandwidth Aggregation Techniques Packet scheduler on network proxies EDPF [TMC06]: uses Earliest Delivery Path First packet scheduler Transport layer MPTCP [IFIPNet11]: extends TCP to support multiple paths Application layer MultiNets [RTAS12]: provides seamless switching between 3G and WiFi Hard to deploy Require changes to the existing servers or Internet infrastructure Not designed for energy saving and QoS satisfaction simultaneously 39

40 GreenBag: Multi-link Data Streaming Scheme Divide a file into segments, then divide each segment further into two subsegments, and assign each subsegment into one of the two interfaces for downloading 40

41 GreenBag: Multi-link Data Streaming Scheme Divide a file into segments, then divide each segment further into two subsegments, and assign each subsegment into one of the two interfaces for downloading In-order data Out-of-order data 41

42 GreenBag: Multi-link Data Streaming Scheme Divide a file into segments, then divide each segment further into two subsegments, and assign each subsegment into one of the two interfaces for downloading In-order data Out-of-order data 42

43 GreenBag: Multi-link Data Streaming Scheme Divide a file into segments, then divide each segment further into two subsegments, and assign each subsegment into one of the two interfaces for downloading In-order data Out-of-order data 43

44 GreenBag: Multi-link Data Streaming Scheme Divide a file into segments, then divide each segment further into two subsegments, and assign each subsegment into one of the two interfaces for downloading In-order data 44

45 GreenBag Architecture A middleware on smartphones which provides energyefficient multi-link data streaming services Acts as a proxy between video player and server GreenBag Video Player HTTP Engine Buffer Download Engine LTE WiFi Server Download Planner 45

46 GreenBag Architecture A middleware on smartphones which provides energyefficient multi-link data streaming services Acts as a proxy between video player and server GreenBag Video Player HTTP Engine Buffer Download Engine LTE WiFi Server Segment Manager Download Planner Medium Load Balancer Recovery Decision Maker Energy-aware Link Mode Chooser Energy Goodput Model Predictor Video Player State Monitor 46

47 GreenBag Architecture A middleware on smartphones which provides energyefficient multi-link data streaming services Acts as a proxy between video player and server GreenBag Video Player HTTP Engine Buffer Download Engine LTE WiFi Server Segment Manager Download Planner Medium Load Balancer Recovery Decision Maker Energy-aware Link Mode Chooser Energy Goodput Model Predictor Video Player State Monitor 47

48 GreenBag Architecture A middleware on smartphones which provides energyefficient multi-link data streaming services Acts as a proxy between video player and server GreenBag Video Player HTTP Engine Buffer Download Engine LTE WiFi Server Segment Manager Download Planner Medium Load Balancer Recovery Decision Maker Energy-aware Link Mode Chooser Energy Goodput Model Predictor Video Player State Monitor 48

49 LTE Power Management States of LTE power management: ACTIVE, TAIL, and IDLE LTE interface in TAIL state has high power consumption The TAIL state timeout TT tttttttt is long (11.2 seconds) Keep LTE idle long enough until LTE interface demotes to IDLE state Not opportunistically offloading data to WiFi ACTIVE ACTIVE any data traffic IDLE any data traffic TAIL state timeout expires no data traffic TAIL IDLE TAIL TT tttttttt (a) LTE power states (b) LTE power trace 49

50 Energy-aware Link-mode Chooser Choose the link mode which consumes the least energy to download the remaining portion of the file without incurring any QoS violation (playback interruption) Three link modes: Dual-link, LTE-only, and WiFi-only 50

51 Power Energy-aware Link-mode Chooser Choose the link mode which consumes the least energy to download the remaining portion of the file without incurring any QoS violation (playback interruption) Three link modes: Dual-link, LTE-only, and WiFi-only Save energy by avoiding irregular data transfer over LTE tt AA tt BB tt CC LTE TAIL Energy LTE TAIL Energy Power LTE TAIL Energy EE AA EE BB EE CC Time Total data transfer time: tt AA + tt BB = tt CC Total energy consumption: EE AA + EE BB > EE CC Time 51

52 Evaluation Evaluation in real-world networks Evaluation in emulated environment 52

53 Real-world Networks Experiment Setup (1/2) Compare 4 configurations (GB-E, GB-P, LTE-only, and WiFi-only) in 3 scenarios (Stationary #1, Stationary #2, and Mobile) in terms of 2 performance metrics Two performance metrics Playback time Energy consumption Four configurations GB-E: GreenBag-Energy mode, supports QoS while saving energy GB-P: GreenBag-Performance mode, maximizes throughput without saving energy LTE-only (WiFi-only): video streaming using only LTE (WiFi) 53

54 Real-world Networks Experiment Setup (2/2) Compare the 4 configurations (GB-E, GB-P, LTE-only, and WiFi-only) in 3 scenarios (Stationary #1, Stationary #2, and Mobile) Three scenarios Stationary #1: Bandwidth of LTE is always greater than of WiFi Stationary #2: Bandwidth of LTE is always smaller than of WiFi Mobile: WiFi bandwidth fluctuates fast and has a drop due to user mobility Stationary #1 Stationary #2 Mobile Average LTE bandwidth (Mbps) Average WiFi bandwidth (Mbps) Total average bandwidth (Mbps) Experimented video has 6.1Mbps average bit rate, and is 117 seconds long 54

55 Performance in Real-world Environments GreenBag minimizes video interruption time GreenBag provided no interruption while LTE-only and WiFionly had sec interruption during video playback 55

56 Performance in Real-world Environments GB-E not only minimizes video interruption time but also consumes 14-25% less energy than GB-P 56

57 Playback Time Minimization Demo 57

58 Energy Consumption Minimization Demo 58

59 Evaluation in Emulated Environment Evaluates GreenBag in various environments Automated experimentation Experiment setup Emulated network Emulated player Evaluate four aspects of GreenBag Segment size overhead Effectiveness of adaptive load balancing Effectiveness of recovery mechanism Effectiveness of energy-aware link mode switching 59

60 60 Effectiveness of Recovery Mechanism (1/2) Recovery mechanism avoids the long outof-order data caused by bandwidth drops E.g. WiFi bandwidth can drop when the smartphone moves out of WiFi coverage

61 Effectiveness of Recovery Mechanism (2/2) Recovery mechanism keeps playback time minimum even in the presence of multiple of bandwidth drops 61

62 Effectiveness of Energy-aware Link-mode Switching GB-E provides the same playback time with GB-P, while saving 2%-40% of energy consumption of GB-P VVVVVVVVVVVVVVVVVVVVVVVV TTTTTTTTTTTTTTTTTTTTTTTTTTT 62

63 Conclusion 1. Formulate bandwidth aggregation for real-time video streaming as a lexicographic optimization 2. Design a multi-link data streaming middleware to support real-time delivery in the most energy-efficient way 3. Implement a prototype of GreenBag on Android-based mobile devices equipped with LTE and WiFi interfaces To the best of our knowledge, this is one of the first LTEenabled prototypes which demonstrates the effectiveness of bandwidth aggregation for energy-efficient real-time delivery on mobile phones 63

64 THANK YOU QUESTIONS AND ANSWERS cps.kaist.ac.kr/greenbag

65

66 Our System Easy to deploy Works on the application layer of the receiver, requiring no changes to the existing servers and Internet infrastructure Contrast to transport layer solutions (MPTCP and MPRTP) and scheduling network proxy solutions (EDPF scheduler and PRISM) Uses the standard HTTP which is prevalent and benefits from HTTP content caching on the Internet Conserves energy in bandwidth aggregation subject to QoS constraints Saves energy basing on LTE power management Other bandwidth aggregation solutions (OPERETTA and MultiNets) paid little attention to energy consumption in cellular networks and QoS satisfaction simultaneously Evaluated in real-world networks LTE-enabled prototype demonstrates its effectiveness for realtime video streaming on real-world networks 66

67 Next Segment Planning Procedure GreenBag plans the next segment to download further when remaining time for a link to finish its subsegment is smaller than a threshold IF NOT Perform recovery THEN Decide the next segment size Choose the most energy-efficient link mode Decide the subsegment sizes of the two links in the next segment Segment Manager Download Planner Medium Load Balancer Recovery Decision Maker Energy-aware Link Mode Chooser Energy Goodput Model Predictor Video Player State Monitor 67

68 Components of Download Planner (1/5) Segment Manager Determine the next segment size Uses a fixed segment size in typical cases Minimizes waiting time between consecutive segments by using HTTP pipelining Segment Manager Download Planner Medium Load Balancer Recovery Decision Maker Energy-aware Link Mode Chooser Energy Goodput Model Predictor Video Player State Monitor 68

69 Components of Download Planner (2/5) Medium Load Balancer Determine the subsegment sizes and links assignment in the next segment Basing on unfinished portion size, next segment size, and estimation of goodput (detailed formulas in the paper) Reduce out-of-order delivery Both interfaces would finish at the same time The faster link would download the earlier subsegment Require good estimation of current goodput Segment Manager Download Planner Medium Load Balancer Recovery Decision Maker Energy-aware Link Mode Chooser Energy Goodput Model Predictor Video Player State Monitor 69

70 Components of Download Planner (3/5) Recovery Decision Maker Handle the case when goodput estimation is wrong Replan the downloading again When GreenBag made poor decisions on load balancing due to fast fluctuation of network bandwidth, such as a drop of WiFi bandwidth Segment Manager Download Planner Medium Load Balancer Recovery Decision Maker Energy-aware Link Mode Chooser Energy Goodput Model Predictor Video Player State Monitor 70

71 Appendix Energy Models of LTE and WiFi Transmission energy used to download xx Mbits with bandwidth yy Mbps is formulated as EE tttt xx, yy = αα yy + ββ xx yy (mmmm) αα yy + ββ : power consumption; xx : time for downloading data yy LTE WiFi αα ββ Energy consumption in TAIL state of LTE: EE tttttttt = PP tttttttt tt Power consumption in TAIL state: PP tttttttt = mmmm tt is the duration in which LTE is in TAIL state We ignore promotion, and data upload energy values, because they are too small 71

72 Video Player Model The video player is modeled as a queue in front of the video decoder Arrived data: AA(tt) Decoded data: DD(tt) Buffered data: XX(tt) XX tt = AA tt DD(tt) Video playback starts when data in the buffer is sufficient: XX tt BB B is called lower threshold Video player pauses when buffered data XX tt decreases to zero Data Source A(t) D(t) L k-th bit A(t) Data size Video Buffer X(t) B X(t) B B D(t) Video Decoder t 0 t 1 t 2 t 3 t 4 Time 72

73 Segment Size Overhead Optimal segment sizes are independent of bandwidth heterogeneity and video bit rates Small segment size imposes more overhead in requesting every new segment Large segment size increases out-of-order data delivery VVVVVVVVVVVVVVVVVVVVVVVV TTTTTTTTTTTTTTTTTTTTTTTTTTT (a) Overhead in different link heterogeneity (b) Overhead in different video bit rates 73

74 Effectiveness of Adaptive Load Balancing Bandwidth heterogeneity has little effect on the performance of the adaptive load balancing scheme, while affecting the fixed load balancing (1:1) significantly (LTE:WiFi) 74