1. Saving Energy in Mobile Devices for On-Demand Multimedia Streaming
- A Cross Layer Approach
Authors: Presented By:
Mohammad Asharful Hoque Koushik Nadakuduti
Matti Siekkinen and Jukka K. Nurmen Vivek Chintalapati
Sasu Tarkoma
Mika Aalto

2. Introduction
On-demand multimedia streaming has gained great acceptance among smartphone users.
However, the battery life of smartphones becomes critical when accessing these multimedia services via Wireless networks (e.g., Wi-Fi, 3G and 4G).
Most often, power consumption for using wireless networks is greater than playback power consumption.
In this work, energy spent by mobile devices for TCP-based multimedia streaming is optimized.
Spotify, Netflix, YouTube, Hulu

3. Introduction
It is well known that shaping multimedia traffic into periodic bursts can save more energy for UDP-based traffic over Wi-Fi
Wireless Network Interface (WNI) is kept active only for a short period
Today, HTTP over TCP is by far the most prevalent set of protocols used for streaming
TCP based streaming differs from UDP based streaming in one major way: TCP is a reliable protocol.
Number of packets are collected over period of time and sent together as one burst to client.

4. Energy Efficiency of Wireless Multimedia Streaming over TCP
Understand the energy consumption of wireless network interfaces for TCP-based multimedia streaming.
Some inactivity timers control the energy consumption.
Because of these timers, there is residual time spent by any interface in active state after transmission or reception for the sake of user experience, which leads to some energy spent doing nothing.
This energy is referred to as Tail Energy.

5. Energy Consumption of WiFi WNI
standard includes a Power Saving Mechanism
In commercial mobile phones, slightly modified version is used which is referred to as PSM-Adaptive (PSM-A)
Keeps the interface in idle state for a certain fixed period of time instead directly switching to sleep state after transmission or reception.
Power draw in sleep state is lower than in idle state.
While transmitting and receiving, interface may consume up to 50% more power than in idle state.

6. Energy Consumption of HSPA/3G WNI
Usage of radio resources and power consumption is controlled by the Radio Resource Control (RRC) protocol.
This protocol has 4 different states.
RRC State machine diagram illustrates the inactivity timers that control transition among these states.

7. Fig 1: 3G HSPA State transitions and power consumptions
-Different states, different timer and power consumption.
-Some network providers disable CELL_PCH state (Fast Dormancy)
- Require signalling message exchange for state changes
Fig 1: 3G HSPA State transitions and power consumptions

8. Energy Consumption of LTE/4G WNI
LTE RRC contains 2 states: RRC_IDLE and RRC_CONNECTED.
Includes a discontinuous reception mechanism specifically to be used in RRC_CONNECTED state.
The idea is that after no packets have been received for a long period of time specified by the cDRX inactivity timer, device starts a duty cycle.
Wakes up periodically for DRX on-duration specified amount of time to check for new incoming packets.
If there are no packets, switches to RRC_IDLE state.

9. Fig 2: LTE RRC with connected state DRX
Only two states, less signalling due to state transitions in the network.
Fig 2: LTE RRC with connected state DRX

10. Mobile Multimedia Streaming
Playback Duration, File Size, Formats and Quality are key-properties of multimedia content.
Video length distribution has 3 peaks: 1 min (20%); 3-4 min (16.7 %); 10 min
File sizes observed to follow a smaller distribution. Typically < 30MB
Streaming services need to deal with bandwidth fluctuation and jitter. Therefore, do initial buffering at possible maximum rate. This phase is known as Fast Start. After this phase, different approaches are used by different services.
YouTube throttles the sending rate a little bit higher than the encoding rate.
Fast caching downloads the whole content as fast as possible, which can eventually be very resource wasteful when the whole content is consumed.
Content length, file size, file format and bit rate.
Short music files or live radios.

11. Delivering Multimedia Content in Bursts using TCP
Large fraction of content has the bit rate of a few hundred kilobits per second.
Smartphone users enjoy 2Mbps mobile network speed on average.
WNI will be less utilized if the same amount content can be received in bursts at maximum possible rate instead of continuously receiving at encoding rate.
Typically there is more available bandwidth on the path than encoding rate of the content.

12. Delivering Multimedia Content in Bursts using TCP
There is a caveat. Multimedia players maintain a fixed-size playback buffer. The size depends on the implementation of the player and can be restricted by the operating system of the smartphone
If size of the burst is larger than the available space in playback and TCP receive buffer together, TCP flow control becomes active in the client.
This can happen during constant-quality and adaptive streaming
The activation of TCP flow control leads to increased energy consumption as we will see in the next topics.

13. Power Modeling and Analysis of Bursty Streaming
Firstly, we consider the case when the fixed-size burst can be entirely absorbed by a streaming client.
Results show that power draw either stays the same or decreases when T is increased as long as player buffer and TCP buffer together can hold the whole burst.

14. Power Modeling and Analysis of Bursty Streaming
Now, In the case where a burst is larger than the available buffer space at the client.
Portion of the burst that exceeds the buffer size can be transmitted at an average encoding rate.
Results show that power consumption is a non decreasing function of burst interval, T
The lowest energy consumption can be achieved when the burst size exactly matches the available buffer space at the receiver

15. 5a and 5b: average power consumption vs burst interval and buffer space.
2 observations:
Average power consumption decreases much more sharply with WiFi than with LTE when burst interval is increased. Reason is difference in the length of inactivity timers. Takes longer burst cycle with LTE to amortize the tail energy.
Setting the burst size too large leads to a significant increase in power consumption. Penalty is more striking in LTE. Hence it is not a good idea to just blindly choose a “large enough” burst size. Instead a smarter mechanism is used.
Fig: Average power consumption with different burst intervals and amount of buffer space.

16. EStreamer
Multimedia delivery system that can be integrated with an audio/video streaming service.
Works with both constant quality and rate adaptive streaming
Turns an input multimedia stream into bursts and delivers these to the client
Determines the optimal burst size for the client and uses that when constructing the bursts
Cross layer system consists of two components
Traffic Profiler at the Transport layer
Traffic Shaper at the Application layer

17. EStreamer - Traffic Profiler
Capture TCP ACK packets coming from the streaming clients and log the arrival time of the ACKs
Collect the sequence numbers and advertised receive window size from the ACK packets
ACKs arrival times are used to calculate the duration of a burst and the receive window size is used to estimate the total buffer at the streaming client
Window size zero indicates the client’s total playback buffer B is full
Sequence number is used to determine whether a streaming client received all the packets of a burst

18. EStreamer - Traffic Shaper
Find the encoding rate of the content by parsing the stream header
Keep track of the amount of data sent to the client during the Fast Start to calculate the maximum burst cycle, Tmax
Tmax is the maximum duration that the player can play without distorted playback unless more content is received.
Tmax = L / rs where L is the number of bytes and rs is the encoding rate
Decide a burst cycle T according to the player buffer status received from the profiler
Continues buffering the incoming traffic and sends a burst of size T x rs to the client when T expires
Can also determine the end-to-end bandwidth for a burst, when it receives the corresponding burst duration from the profiler

19. Finding the optimal burst interval Topt
Burst size depends only on T as rs is constant for a given stream quality
TCP client sends ZWAs to the sender to halt data transmission when the client’s TCP receive buffer becomes full.
Profiler uses ZWAs to find the optimal burst size BSopt and reports to the Shaper
Shaper finds Topt as BSopt / rs

20. Finding optimal burst size
During Fast Start - If the profiler finds that the client’s total buffer is filled even before the Fast Start is over, BSopt = Sent Bytes until the first zero window advertisement is received by the profiler
After Fast Start -
Uses binary search to speed up calculation of optimal burst size
Calculate Tmax during Tfs
Let T = Tmax / 2 seconds
If there are no ZWAs, T is increased
Either EStreamer reaches Tmax without experiencing any ZWA or
Observes ZWA and BSopt = Sent bytes for a T which is less than Tmax

21. Dealing with bandwidth fluctuation
Playback can be interrupted when using CBR streaming applications even without EStreamer if the bandwidth drops below the encoding rate
This situation occurs when traffic shaping is in one of the following three state
BSopt has not been found yet
BSopt has been found and limited by the client’s buffer space
BSopt is limited by Tmax
EStreamer preserves one of these occurred states by setting Told = T
After which it continuously sends content to the client while measuring the bandwidth in order to detect when the situation improves.
It also continuously updates Tmax based on the sent content and the encoding rate as long as the low bandwidth situation continues
When the bandwidth situation improves, it resets the old state by setting T = Told
Traffic shaping is resumed to find the optimal burst size

22. Rate-adaptive streaming for bandwidth fluctuations
EStreamer applies traffic shaping and rate adaption together
The stream is encoded into n different qualities, each of which have a different encoding rate rs = 1,2, 3… n.
Initially, EStreamer choses an encoding rate which is less than or equal to 1Mbps for each of the n qualities
Tmax = Topt = BSopt / rs
EStreamer allows upgrading the quality only when the current bandwidth is at least twice the encoding rate of the higher quality.
If client does not send any ZWA during the Fast Start phase, the content sent in that phase marks the largest possible burst size that can be used
EStreamer sets this burst size as Bopt and computes Tmax = Topt = BSopt / rs
Initially, it sets Topt to be the optimal for any other lower quality stream with inferior encoding rate
When switching to a higher quality stream, EStreamer uses the previously found BSopt to determine an equivalent T and continues Traffic Shaping using binary search.
If ZWAs are observed, then EStreamer sets the optimal burst size for all qualities.

23. Implementation
EStreamer is deployed in an HTTP proxy server and placed it in the cloud
The smartphones proxy settings were configured to use EStreamer
EStreamer does not shape traffic during the Fast Start phase and forwards 20-45s playback data depending on the service
For rate-adaptive streaming, EStreamer has been integrated with a streaming server and developed a separate client
Communication works based on HTTP over TCP
Client makes initial get request
EStreamer selects the initial quality and responds by sending the stream initialization header of the corresponding quality
Also mentions the duration of the video, bit rate, time range and other parameters in the response header
Client makes the request for actual content in time range
EStreamer responds by sending the next chunk
Client only specifies the start of the range, EStreamer determines an appropriate burst rate and if the stream quality should be switched or not
Whenever a client’s playback buffer goes down to only 5 seconds, it sends a new request specifying the next missing content as the beginning of the range
If EStreamer decides to switch stream quality, it specifies the new bit rate in the response header
Response mismatch can occur when there are ZWAs from the client, as EStreamer might send less content than the time range specified in the header because EStreamer aborts sending the remaining content of the determined burst size.
Client request the remaining content and EStreamer responds with code 204 and correction in the response header

24. Performance Evaluation
Used Internet Radio, YouTube, and DailyMotion in four different smartphones via EStreamer
Two sets of measurement with the following purpose
Study the effect of timer settings on the energy consumption of smartphones
Study the effect of our energy-aware traffic shaping on radio network signaling

25. Burst size tuning and energy consumption of constant-quality streaming
Three possible cases of EStreamer finding the optimal burst size
During the Fast Start phase and determines Topt
Determines Topt using the binary search ending up with a value that is less than Tmax
Finds Topt at Tmax

26. Observations Phones which use Fast Dormancy can save more energy
Energy savings in audio streaming are greater than video
Largest energy savings can be achieved when streaming via Wi-Fi (65%), followed by LTE (55%) and then HSPA (38%)

27. Impact of Background Traffic
applications, free applications with embedded advertisement code might interleave with the video content bursts and increase the overall energy consumption
Browser establishes several TCP connections to YouTube servers, EStreamer shapes only the traffic connection that carries the video content

28. Impact of Bandwidth fluctuation
Power consumption increases when bandwidth decreases
If the bandwidth declines below the lowest available encoding rate, user may experience distorted playback.
EStreamer can react accordingly for both constant bit rate and rate adaptive streaming for improving battery life of a smartphone

29. Impact of Cellular Network Configuration
Shorter timers do not generally help to save energy in the absence of EStreamer
With EStreamer, slight decrease in energy consumption can be observed with more aggressive timer values than default configuration.
Energy savings for audio streaming are higher than video because of quicker transition to lower power state after receiving a burst.

30. Impact of traffic shaping on radio network signaling
Energy saving achieved by reshaping traffic into bursts come with a price when using cellular network access
Savings are achieved by increasing the number of RRC state changes during the streaming session.
Each state change generates some signaling traffic
As the number of signaling messages required for a specific state change is known, the total number of signaling messages was computed
Overall, the signaling traffic grew more compared to the YouTube experiments as EStreamer selected shorter burst intervals in the audio streaming case because of the shorter Tmax

31. Observations
Signaling load in the network increases if a network does not support the CELL_PCH state because of frequent RRC reconnections
Usage of legacy Fast Dormancy by the smartphones reduce energy consumption but increases the signaling load very rapidly as the mobile device frequently closes and re-establishes the RRC connection
Usage of very short timers reduces energy consumption but can increase the signaling load in the network because of the frequent state transitions but the effect is less pronounced compared to the previous cases

32. Conclusions
Modeled the energy consumption of TCP based bursty streaming traffic
Designed and implemented an energy efficient multimedia delivery system called EStreamer
EStreamer strives to maximize energy savings without compromising the quality of the streaming service
Energy consumption differs quite a lot between the devices, which is explained by different hardware components used in the devices.

33. Questions?
Thank you!