Department of Computer and Information Science
Final thesis
3G Transmission Energy Savings through
Adaptive Traffic Shaping Policies
by
Henning Hall and Christian Luckey
LIU-IDA/LITH-EX-G--14/079—SE
2014-09-02
Linköpings universitet
SE-581 83 Linköping, Sweden
Linköpings universitet
581 83 Linköping
Final thesis
3G Transmission Energy Savings through
Adaptive Traffic Shaping Policies
by
Henning Hall and Christian Luckey
LIU-IDA/LITH-EX-G--14/079—SE
2014-09-02
Supervisor: Simin Nadjm-Tehrani
Examiner: Nahid Shahmehri
through Adaptive Traffic
Shaping Policies
Henning Hall
Christian Luckey
Students in the 5 year Information Technology program complete a semes-ter-long software development project during their sixth semester (third year). The project is completed in mid-sized groups, and the students im-plement a mobile application intended to be used in a multi-actor setting, currently a search and rescue scenario. In parallel they study several top-ics relevant to the technical and ethical considerations in the project. The project culminates by demonstrating a working product and a written re-port documenting the results of the practical development process including requirements elicitation. During the final stage of the semester, students cre-ate small groups and specialise in one topic, resulting in a bachelor thesis. The current report represents the results obtained during this specialization work. Hence, the thesis should be viewed as part of a larger body of work required to pass the semester, including the conditions and requirements for a bachelor thesis.
Abstract
This bachelor thesis will explore how two traffic shaping mechanisms can help preserve battery power while retaining a certain Quality of Service (QoS) in an Android based application developed for crisis management.
The implemented user-space mechanisms will delay all elastic data re-quests in order to reduce the number of times the 3G transmission radio enters high power states. This lowers the QoS but extends the user equip-ment’s battery life.
The thesis will show that a shaping mechanism has the capability to reduce radio energy usage by up to 50% for the given Android application at the cost of added transmission delays by up to 134 seconds for background traffic. The study also presents two policies that help the application adapt to the current battery level and lower the QoS accordingly, namely one that has a lenient savings effect and one that has an aggressive savings effect.
1 Introduction 1 1.1 Motivation . . . 1 1.2 Purpose . . . 1 1.3 Problem statements . . . 1 1.4 Delimitations . . . 2 2 Background 3 2.1 Traffic shaping . . . 3 2.1.1 Burst buffering . . . 3
2.1.2 Traffic shaping mechanisms . . . 4
2.2 Policy . . . 4
2.3 Quality of Service . . . 4
2.4 Test platform & project application . . . 4
3 Theory 6 3.1 The UMTS Radio . . . 6
3.2 EnergyBox . . . 6
4 Solution space 8 4.1 Traffic shaping mechanisms . . . 8
4.2 Three shaping modes . . . 9
4.2.1 No shaping mechanism . . . 9
4.2.2 Wait Until Non-Elastic Traffic . . . 9
4.2.3 Fixed Maximum-Delay Shaper . . . 9
4.3 Policies . . . 9
4.3.1 The lenient saving policy . . . 10
4.3.2 The aggressive saving policy . . . 10
5 Methodology 12 5.1 Hardware setup . . . 12
5.2 Puppet master . . . 12
5.3 Capturing data . . . 12
5.4 Battery discharge time estimation . . . 12
5.4.1 Choice of constants . . . 14
5.4.2 Average transmission delay . . . 14
6 Results 15 6.1 Shaping mechanism tests . . . 15
6.2 Battery discharge time estimates . . . 17
6.2.1 LSP . . . 17
6.2.2 ASP . . . 17
7.2 Results . . . 19
7.2.1 Shaping mechanisms tests . . . 20
7.2.2 Policy estimation . . . 20
7.3 Wider context . . . 20
1
Introduction
This thesis considers the power saving potentials of an Android applica-tion that communicates with the Internet using a 3G (UMTS1) radio. It
examines how energy savings can be made through different traffic shap-ing mechanisms on the user equipment (UE) applied at different times in adaptation to the UE’s current battery level.
1.1
Motivation
Smartphone usage is increasing rapidly around the world, especially in the anglosphere and in Scandinavia where the majority of the population owned one in 2013 [1]. Hence, reducing the power consumption of these devices has an impact on the overall power consumption in these countries. Fur-thermore, an estimated average 50% of an applications energy consumption on the UE is caused by the UMTS radio [2].
A recent study from the University of Michigan showed that not only is a lot of energy consumed by the radio, but much of it is perhaps unnecessarily wasted. The popular American music streaming service Pandora’s mobile application while only standing for 3.6% of the UE’s total traffic could be held responsible for 64.1% of the energy consumed by the radio. Specifically it found that periodic data which only stood for 0.2% of total data sent accounted for 46% of the total energy consumption of the radio [3].
As such authors find it motivated to explore what countermeasures can be put in place to reduce the energy waste. Not only because having to charge your phone is inconvenient but also because of the environmental argument.
1.2
Purpose
The purpose of this thesis project is to explore the potential of Quality of Service (QoS) degradation through traffic shaping as a tool for saving energy on UE’s with an UMTS radio. Further, it explores whether using a battery dependent policy when applying such mechanisms also has these benefits.
1.3
Problem statements
The following issues have been chosen as the focus of the study:
1. Through traffic shaping, to what a degree must the QoS be decreased in order to achieve battery discharge time improvements on the UE and how much can it be decreased while retaining acceptable func-tionality?
2. Will a policy applying different shaping mechanisms depending on the UE’s battery level also have an impact on the UE’s battery discharge time?
1.4
Delimitations
This thesis project only examines the characteristics of a single Android ap-plication. This limitation was put in place due to the strict time limitations of the project. This does mean that the results are in the strictest sense only valid for this one Android application, but it was a limitation put in place by necessity.
The project only examines the energy consumption of the UMTS radio. It might very well be that the extra CPU activity caused by the shaping mechanisms causes the CPU to enter a higher energy state which then cancels all the power savings made on the radio.
There will be no measurements on how large a part the UMTS radio’s energy consumption is of the total energy consumption of the UE. Only different traffic shaping mechanisms and their associated radio energy usage will be measured and compared.
The application is only looked at as a whole. No parts of the application will be tested in isolation from the others and as such the battery drain by the radio caused by different application components will not be determined. Tests will be carried out using a virtual x86 Android machine on a x86 laptop with a UMTS radio. Rooting an actual Android device and performing the tests on it was unfortunately not an available option.
2
Background
It has come to the attention of many researchers that when it comes to power consumption:
“Little attention, however, has been paid by OS designers to the efficiency of the interactions between cellular networks and mobile applications.” [4]
And when it comes to adaptation:
“Simulations demonstrate that the adaption has a positive im-pact on the battery life time, increasing it by 14%, without de-teriorating the network-wide performance...”[5]
It is against this background that finding viable methods for reducing bat-tery drain caused by the radio and applying those methods when needed seems like a good target for research.
2.1
Traffic shaping
Traffic shaping is a term denoting the improvement of the network behaviour through the act of controlling when packets are sent. Often used to smooth out bursts shaping will in this thesis rather be used to create bursts. Shaping is applied on uplink traffic and can be used to reduce the energy consumption [6].
2.1.1 Burst buffering
Burst buffering as a form of traffic shaping is the act of delaying the trans-mission of one or more packets until some later point when a whole horde of packets has been gathered at which point they’re all transmitted at the same time.
Figure 2: Uplink data transfer with burst buffering, all three packets are sent directly after one another.
2.1.2 Traffic shaping mechanisms
A traffic shaping mechanism will in this thesis refer to a mechanism that determines when transmissions are sent to the underlying Android runtime. The shaping mechanism queues both user and application initiated requests and decides when these are passed on.
2.2
Policy
A policy is for this project defined as an algorithm which applies different traffic shaping mechanisms depending on the battery level of the device.
Figure 3: An instance of application of a policy consisting of three shaping mechanisms.
2.3
Quality of Service
QoS in a networking context is a term that denotes the considered applica-tion’s requirements, normally encompassing the count of dropped packets, throughput, latency, jitter and count of packets delivered out of order.
Because the mechanisms and policies are implemented on the the UE the only factor of the QoS that can be influenced is latency through addition of transmission delays within the traffic shaping mechanisms. This is why added transmission delays the only measure of decrease in QoS that will be treated in the rest of this thesis.
2.4
Test platform & project application
The platform available to the authors to test the premise of the thesis was an issue tracking system consisting of an Android application, an Undertow
web server and a PostgreSQL database [7] [8]. The Android application will from now on be referred to as the project app and all components together as the project system.
The purpose of the application itself is to provide means of communica-tion for police work, to easily allow for informacommunica-tion regarding current and past investigations to be distributed. But also to provide the command central the status position of each unit.
In this application it is possible to do different requests which are gener-ating three basic classes of traffic: non-user initiated elastic (background), user initiated elastic and user initiated non-elastic traffic. Examples of such is displayed in figure 4.
Background User initiated Elastic Position updates,
automatic case sync
Issue upload on save
Non-elastic Issue sync on button push
Figure 4: Some project application functions (possible requests) divided into categories.
3
Theory
Figure 5: The power states and pa-rameters of a UMTS radio [9]. A number of great studies laid
the ground upon which this thesis stands. Their most relevant infor-mation will be reiterated here.
3.1
The UMTS Radio
The UE radio in a UMTS network is using a protocol called Radio Re-source Control (RRC). RRC has three different power states, Pag-ing Channel (PCH), Forward Ac-cess Channel (FACH) and Dedi-cated Channel (DCH), in order of increasing power consumption [10]. It is generally desirable to stay in a lower power state as much as pos-sible and make the time between stepping up as long as possible from an energy consumption standpoint. But since data cannot be sent in PCH the steps into FACH and DCH are a necessity from a QoS stand-point.
A step from PCH to DCH is triggered when the UMTS radio buffer occupancy C exceeds B1 and lasts until all dedicated channels have been
released followed by a radio specific inactivity timer T1, alternatively that
a low activity timer has run for a time Td. The radio then steps down into
FACH where it remains until another inactivity timer T2 runs out and it
steps down into PCH; unless C exceeds B2 in which case a step up into
DCH is triggered. A step from PCH to FACH is triggered when there is buffered data but the amount is less than B1. Note that not all carriers and
radios implement all these features.
Therefore, it is desirable to send as much data as possible once the radio enters the DCH state. Furthermore, there is often a lot of unused time spent in DCH which could be filled with data transmissions as shown in the traffic backfilling study by Lagar-Cavilla et al. [4]. It is also sometimes possible to make the radio only step up into FACH if the amount of data given is low enough.
3.2
EnergyBox
EnergyBox is a piece of software developed at Link¨oping University by Ekhiotz Jon Vergara. It takes a Packet Capture (PCAP) generated by for example tcpdump and outputs the amount of energy that would have
been spent transmitting the captured data over the UMTS radio. Fed with the correct parameters for a specific carrier it is capable of estimating when and for how long the radio stays in which power state.
4
Solution space
The implementation of a policy to select shaping mechanism under differ-ent battery conditions could be done in the user space of the UE, in the application itself. The reasoning behind such an implementation being that it allows for fine grained control over which data is or isn’t sent at any specific time. The drawback of this approach is that the policy has to be implemented for every application running on the phone. And the imple-mentations have to be coordinated in some way in order to achieve full effect.
Another implementation could be in the kernel space in order for the policy to apply to every application running on the UE with the downside that the mechanism must be much coarser, for example it cannot know exactly how the transfer was initiated. On the other hand the implementa-tion would not only be faster, and therefore more energy efficient in terms of CPU utilization, but also universal for all running applications.
A third alternative could be to implement some sort of middleware in-tercepting packets as they are sent from the applications to the kernel. The amount of control available to the shaping mechanism would be the same as it is at the kernel level, but the implementation would not necessarily be as effective as using the already existing in-kernel facilities for traffic shaping. Due to the limited amount of time available to the authors the simpler of the three approaches, the in-app implementation, was chosen. The results may therefore be of limited use but will still point towards what application developers can hope to achieve in terms of battery savings.
4.1
Traffic shaping mechanisms
One of the potential power saving mechanisms that could be implemented is something along the lines of Cross-Layer Burst Buffering (CLBB) [11]. When policy dictates that energy must be saved the shaper delays the trans-mission of elastic, time-independent data and sends it at some later time in bursts as mentioned in chapter 2.1.1. In our case, however, the notion of ”cross layer” is not applicable since we only use application level parameters. In addition to the submission of data, the requests for data from a remote server could in a similar manner be rounded up into bursts and sent at the same time. The shaping mechanism is only shaping uplink traffic, but it still affects incoming traffic.
Inelastic traffic such as a Voice over IP (VoIP) call should in no way be delayed by the mechanisms. However, while a call is ongoing there is no harm in sending a lot of background traffic since the radio is already in a high power state, as long as the connection does not get completely saturated by the background traffic.
Another power saving mechanism could completely stop background traffic until the point at which some critical traffic is queued to be sent, such as a VoIP call. This would be a much harsher shaping mechanism, perhaps only applied when the battery level is extremely low.
4.2
Three shaping modes
In the following subsections the three chosen shaping modes will be de-scribed. In section 4.3 a number of different policies will be outlined con-taining these shaping modes.
The Linux kernel, which the Android runtime runs atop, uses the Queu-ing Discipline (QDISC) pfifo fast. ”pfifo fast” is a packet based first in first out (FIFO) discipline with some limited scheduling. This QDISC is always active independently of what happens in user space [12].
4.2.1 No shaping mechanism
This first mode implements no shaping at the application level and will be referred to as the non-shaping mode. Any request sent through it is immediately handed off to the Android runtime in. The Android runtime in turn hands off the packets to the Linux kernel. The ordering of the requests is never changed in any of the mechanisms applied in this work. 4.2.2 Wait Until Non-Elastic Traffic
The second mode implements the Wait Until Non-elastic Traffic (WUNET) mechanism. All transmissions of elastic data are delayed until any non-elastic data is queued for transmission. If no more non-non-elastic data is queued after a certain point, all data from that point on will stay in the queue indefinitely.
4.2.3 Fixed Maximum-Delay Shaper
The third mode implements a traffic shaping mechanism called the Fixed Maximum-Delay Shaper (FMDS). It implements all the features of WUNET and on top of that a timer which guarantees that no packet will be delayed longer than a given maximum delay Dm. Once the timer fires every request
queued so far is sent off. Non-elastic data is sent without delay and also triggers any queued requests to be sent.
In this thesis the maximum delays of 5, 10, 20, 30, 60 and 90 seconds are investigated. A FMDS mechanism with a maximum delay of 60 seconds will be called FMDS 60 and one with a maximum delay of 30s will be called FMDS 30 etc.
4.3
Policies
In this study two different policies will be formulated, evaluated and com-pared to using no policy at all. Note that the policies are only examples produced for this project. Requirements for transmission delay or minimum battery discharge time must be defined before choosing policy in a real world implementation.
4.3.1 The lenient saving policy
The first policy, shown in figure 6, is called the lenient saving policy (LSP). It uses a mix of FMDS 20, FMDS 60 and the non-shaping mode. FMDS 20 is active when the battery is less than half empty but has more than a fifth left. FMDS 60 is activated when less than a fifth but more than 1% of the battery charge remains. Finally the non-shaping mode will be used again if the battery only has 1% or less left of its energy.
The reason for going back to the non-shaping mode at 1% is to give the application a chance of sending it’s data before the phone shuts down; like a final scream for help.
Figure 6: LSP, Etotstands for the total energy stored in battery.
4.3.2 The aggressive saving policy
The second policy is called the aggressive saving policy (ASP). ASP aims at being lower energy consuming policy than LSP. FMDS 90 is active when less than 90% but more than 50% of the battery energy remains. With less than half the battery energy left WUNET is enabled until when only 1% remains. For the last percentage the non-shaping mode is active, just like in LSP. ASP is illustrated in figure 7
5
Methodology
In short terms the method used consists of the following. 1. Implement the mechanisms from section 4.3.
2. Produce a repeatable test procedure, run it with every mechanism, capture the network traces generated and run these through Energy-Box.
3. Examine the results from EnergyBox and the logging facilities added to the project app to evaluate the effect of the mechanisms.
4. Evaluate the policies using a mathematical estimation method.
5.1
Hardware setup
The UMTS radio used in the trials of this thesis was an Ericsson HSPA F5521gw inside a Lenovo T420s. Unable to find the specific parameters for the carrier used the radio parameters mentioned in chapter 3.1 fed to EnergyBox were set to a ad-hoc: T1 = 4.12 [s], T2 = 5.67 [s], B1 = 1000
[B], B2= 294 [B]. These parameters were the defaults of EnergyBox.
5.2
Puppet master
In order to evaluate the policy and individual mechanisms a repeatable 10 minute test was set up and executed. A puppet master module was programmed for the project application written to emulate user behaviour in a repeatable manner. The puppet master simply navigates the application in a user-like way for any given amount of times, 10 minutes specifically in this work.
5.3
Capturing data
The traffic emanating from the project application was captured with the use of tcpdump on the virtual device with a filter selecting only requests to and from the application [13]. The output of tcpdump being a PCAP was to EnergyBox which allowed the energy consumption of the UMTS radio to be estimated [14] [15].
5.4
Battery discharge time estimation
To evaluate different policies and compare them to each other, the total bat-tery discharge times and average transmission delays has to be estimated for each of the policies. This is done through a calculation using the following constants.
Etot: Total energy stored in the battery [Wh].
n : Total number of shaping mechanisms in the policy.
Po: The power consumption of the UE components other than the radio [W].
We assume that P0 is constant in order to simplify our estimations. In
addition, the following variables are identified as an input to an estimation model for the battery discharge time.
Di: Average transmission delay for shaping mechanism i [s]
Dp: Average transmission delay for the policy [s].
Ei: Battery energy available to the UE while mechanism i is active [Wh].
Pi: Average radio power while shaping mechanism i is active [W].
Ti: Time spent with shaping mechanism i active [s].
Tb: Total UE battery discharge time [s].
The energy consumed by UE with only one shaping mechanism (P1) can
be described as:
Etot= Tb· (Po+ P1)
This can be rewritten as the following expression for the total battery dis-charge time:
Tb=
Etot
Po+ P1
A policy consisting of two shaping mechanisms with powers P1and P2which
are enabled while the UE consumes E1and E2can be described as:
Tb = E1 Po+ P1 + E2 Po+ P2 where Etot= E1+ E2
Written on a general form for a policy consisting of n shaping mechanisms the calculation will look like:
Tb = n X i=1 Ei Po+ Pi where Etot= n X i=1 Ei (1)
It should be duly noted that these calculation only estimates the battery discharge time running a given type of trace and in order to get a real value for a given trace it is necessary to measure Po on a real device.
5.4.1 Choice of constants
The constant Etot of course depends on the battery used. In this thesis a
battery from a Samsung Galaxy S4 with a capacity of 9.88Wh was used as an example.
According to the study by Pathak et al. [2] about 50% of the energy consumed by a regular smartphone application is consumed by the radio. Therefore Po is assumed to be 50% of the total power when no shaping is
used which is equivalent to the radio power during non-shaping mode. 5.4.2 Average transmission delay
Given the time spent with each shaping mechanism (Ti) resulting in the total
discharge time (Tb), the total average transmission delay for the policy can
be calculated. The average delay experienced for a given trace with two shaping mechanisms will be calculated by:
Dp=
D1· T1+ D2· T2
Tb
Di is the average delay experienced by packets in a given trace while
run-ning with the mechanism that uses the average power Pi. The average
delay experienced for the same trace using a policy consisting of n shaping mechanisms can therefore be expressed with the following formula:
Dp= n X i=1 Di· Ti Tb (2)
6
Results
The puppet master, tcpdump and EnergyBox generate test results for all the previously mentioned shaping mechanisms which will be presented in this chapter. The estimations for the total battery discharge time and average request delay under the given policies, assuming that kind of trace is run during the whole life time, will also be presented.
6.1
Shaping mechanism tests
After running eight ten minute tests, one for each shaping mechanism, the transmission delay was obtained for each queued request along with the average transmission delay for the whole mechanism from the logging facil-ities of the project application. This data is together with energy saving presented in table 1 and visualized in figure 8. Each ten minute test con-tained 33 requests. Shaping mechanism Energy consum-ption (mWh) Energy savings Power (W) Average trans-mission delay (s) WUNET 36 51% 0.21 134 FMDS 90 48 35% 0.29 48 FMDS 60 51 29% 0.31 32 FMDS 30 55 24% 0.33 21 FMDS 20 64 11% 0.39 14 FMDS 10 73 0% 0.44 8 FMDS 5 76 -5% 0.46 4 No shaping 73 0% 0.44 0 Table 1: Energy savings are compared to the energy consumption of the trace with no shaping.
Figure 8: Average transmission delay and energy consumption for each shaping mechanism. The standard deviation of the delay is shown as bars. As expected, the longer max delay parameter the higher the added lantency is, and the lower the energy consumption is.
(a) WUNET (b) FMDS 30
(c) FMDS 5 (d) Non-shaping mode
A few example graphs for individual traces, depicting RRC states, is displayed in figure 9. As expected, the DCH state was reached fewer times with a longer max delay.
6.2
Battery discharge time estimates
After finding the energy consumption for each shaping mechanism, it is possible to estimate the policies described in section 4.3. This is done by calculating Tb and Dp by using the formulas from section 5.4.
6.2.1 LSP
Pi and Di in table 2 are obtained from the shaping mechanism tests in
section 6.1. Ei is the energy share given to the specific shaping mechanism.
i Mechanism name Ei Pi Di
1 Non-shaping 5.04 0.44 0 2 FMDS 20 2.96 0.39 14 3 FMDS 60 1.88 0.31 32 Table 2: Experimentally obtained values for LSP
Equation 1 in section 5.4 with the inputs from table 2 can be used for calculation of the total battery discharge time:
Tb= 5.04 0.44 + 0.44+ 2.96 0.44 + 0.39+ 1.88 0.44 + 0.31 = 11.8h
The average transmission delay for the policy is also estimated with the formula from section 5.4.2:
Dp= 0 · 5.72 11.8 + 14 · 3.57 11.8 + 32 · 2.51 11.8 = 11s 6.2.2 ASP
The total battery discharge time and average transmission delay for ASP is calculated following similar steps as for LSP:
i Mechanism name Ei Pi Di
1 Non-shaping 1.08 0.44 0 2 FMDS 90 3.95 0.29 48 3 WUNET 4.84 0.21 134 Table 3: Experimentally obtained values ASP Tb= 1.08 0.44 + 0.44+ 3.95 0.44 + 0.29+ 4.84 0.44 + 0.21 = 14.1h
Dp= 0 · 1.23 14.1 + 48 · 5.41 14.1 + 134 · 7.45 14.1 = 89s 6.2.3 No shaping
To make any conclusions about the LSP and ASP policies it is necessary to estimate the total battery discharge time when no shaping is applied for the entire time.
i Mechanism name Ei Pi Di
1 Non-shaping 9.88 0.44 0 Table 4: Variables when no new policy is applied.
Tb=
9.88
0.44 + 0.44 = 11.2h Dp= 0s
Figure 10: The battery discharge time compared between the policies. Figure 10 shows that ASP succeeded with its aggressive energy saving and extended the battery discharge time considerably more than LSP.
7
Discussion
As it stands the results from this study is only valid for one single appli-cation running alone on an Android device. It is probable that multiple applications running the optimizations used in this thesis would still end up sending data at different times, if not coordinated by some sort of central facility, resulting in a much smaller gain than those presented in this thesis. With those limitations in mind, the most interesting of the results was that the tests consequently showed that with a too low timeout constant the energy consumption of the radio rose instead of declining.
7.1
Methodology
Probably the most important change to the study that would have been made if the resources were available would have been to implement the shaping mechanisms in kernel space like in the study by Vergara et al. [6]. As mentioned in chapter 1.4 the results from such a study would have been much more useful if that had been the case.
One potential part of the project which was scoped out is that of mea-suring the power consumption of not only the radio but also the other components. If such measurements had been conducted one could have factored in the additional energy consumption caused by the shaping mech-anisms on the CPU. These measurements could have been done using tools similar to that of EnergyBox but for processors; software estimating the power consumption of a CPU from a trace of it’s machine instructions.
The mathematical model used to estimate total battery discharge time of different policies is very simplified and its inputs are collected for a given repetitive interaction pattern.
The method used in this study could be improved if it had been possible to measure Po instead of only assuming it as given. Given more time Po
could have been measured by attaching a Watt-metering power supply to the battery contact of a real device and deducting the estimated power of the radio estimated by EnergyBox.
Another possible improvement would be to implement a mechanism that sends data directly if it is small enough that it would not trigger the radio to step up into DCH but instead stay in FACH state.
The 10 minutes test trace could be made more realistic by making real user tests to find out how the application is used in real situations. The difficulty in doing so for this study is that the project application used for tests has no real users.
7.2
Results
Almost all the results met the expectations but there are a few exceptions worth mentioning.
7.2.1 Shaping mechanisms tests
Every shaping mechanism, except FMDS with a guaranteed transmission delay less than 10 seconds, decreased the energy consumption compared to having no shaping as seen in figure 9. Because UMTS radio spent less time in DCH the energy consumption decreased with, for example, 35% for FMDS 90 and 51% for WUNET.
All these shaping mechanisms that saved energy also induced some amount of delay. The more energy that could be saved, the more delay was induced. The large average transmission delay of WUNET is explained by its lack of timer and few user-initiated requests in the test trace.
Figure 8 uncovers the irregularity of WUNET seen in the large stan-dard deviation. The obvious conclusion being that WUNET is the most unreliable of the mechanisms.
When FMDS 5 or FMDS 10 was applied the application consumed about the same amount of energy as no shaping did. And since both FMDS 10 and FMDS 5 results in longer average transmission delays they are not worth implementing.
7.2.2 Policy estimation
The result of the policy estimations met the authors´expectations. Using ASP makes a 3.1 hour improvement on the battery discharge time and will therefore be a good option for those circumstances where the UE battery needs to last as long as possible. Note that the average delay is 89 seconds which is affecting the QoS negatively. If some external entity has a latency requirement that will not tolerate for example an average 90s delay, choosing this policy would be unwise.
LSP increased battery discharge time by 0.6 hours which is not much compared to ASP. On the other hand LSP guarantees a maximum delay of 60 seconds and the average delay was estimated to 11s. Because of this, LSP can be good if you want to extend the battery discharge time a bit but still ensure the delay is low, and never higher than 60 seconds.
The two policies are good examples of how setting up a policy consisting of different shaping mechanisms for different battery conditions may have positive impact on the battery discharge time.
One source of error identified in the discharge time estimation is the point where the policy is switching shaping mechanism. In a real imple-mentation the remaining queued requests would be handled by the next shaping mechanism. This is not considered in our estimation since all shap-ing mechanisms tests are startshap-ing with empty request queues.
7.3
Wider context
As mentioned in chapter 1.1 the energy consumption of smartphones and their radios around the world is non-negligible. Serious effort should be put into lower the consumption and the results of this study shows that
relatively simple constructs can help reduce the energy consumption of con-sumer devices all over the world.
8
Conclusion
To finish off this thesis the authors find it fitting to revisit the problem statements from chapter 1.3.
1. How can QoS be decreased to improve the battery discharge time while retaining acceptable functionality?
Our tests indicate that the delays induced by the traffic shaping mech-anisms must be longer than the average spacing between requests. Otherwise no more than one request will be queued for each ”burst” on average and as such no gains in battery discharge time will be had. And just like Calder and Marina [16] found in their study; batching reoccurring transmissions does indeed save a lot of power. In this specific case it saved from 11% to 35% for a 10 minute test sequence depending on the length of fixed timeout or even as much as 51% if all transmission were postponed until user initiated transfers occurred. 2. Will a policy depending on the UE’s battery level have an impact on
the UE’s battery discharge time?
As is obvious in figure10 the battery life of the UE can be increased by the use of a battery dependant policy.
It might be sound to implement a policy but it depends on how one defines sound. While it does make the battery of the UE last longer it does also make for a device that behaves erratically. It is up to the product owner whether or not the improved battery discharge time makes up for the worse user experience.
It can be concluded that traffic shaping even as coarse as what was performed in this project can and does have a positive impact on the battery discharge time of the UE.
References
[1] Inc. Google. Google mobile insights. http://think.withgoogle. com/mobileplanet/en/graph/?country=au&country=ca&country= dk&country=ie&country=nz&country=no&country=se&country= uk&country=us&category=DETAILS&topic=Q00&stat=Q00_1&wave= 2013&age=all&gender=all&chart_type=&active=gender. Accessed May 13 2014.
[2] A. Pathak, Y.C. Hu, and M. Zhang. Where is the energy spent inside my app?: fine grained energy accounting on smartphones with eprof. In Proceedings of the 7th ACM European conference on Computer Sys-tems. ACM, 2012.
[3] F. Qian, Z. Wang, A. Gerber, Z. Mao, S. Sen, and O. Spatscheck. Profiling resource usage for mobile applications: a cross-layer approach. In Proceedings of the 9th international conference on Mobile systems, applications, and services, pages 321–334. ACM, 2011.
[4] H.A. Lagar-Cavilla, K. Joshi, A. Varshavsky, J. Bickford, and D. Parra. Traffic backfilling: subsidizing lunch for delay-tolerant applications in UMTS networks. ACM SIGOPS Operating Systems Review, 45(3):77– 81, 2012.
[5] Massimiliano Raciti, Jordi Cucurull, and Simin Nadjm-Tehrani. Energy-based adaptation in simulations of survivability of ad hoc com-munication. In Wireless Days (WD), 2011 IFIP. IEEE, 2011.
[6] E.J. Vergara, J. Sanjuan, and S. Nadjm-Tehrani. Kernel level energy-efficient 3G background traffic shaper for Android smart-phones. In Wireless Communications and Mobile Computing Confer-ence (IWCMC), pages 443–449. IEEE, 2013.
[7] H. Hall, C. Luckey, S. Nilsson, J. Eriksson, W. Daniels-son, E. K¨arnsund, and J. B¨acklund. Ainappen, source code repository. https://github.com/Rovanion/AinAppen/tree/ kandidatarbete, 2014.
[8] H. Hall, C. Luckey, S. Nilsson, E. K¨arnsund, and J. B¨acklund. Ainappen, source code repository. https://github.com/Rovanion/ AinWebServer, 2014.
[9] Ekhiotz Jon Vergara and Simin Nadjm-Tehrani. Energybox: A trace-driven tool for data transmission energy consumption studies. In En-ergy Efficiency in Large Scale Distributed Systems, Lecture Notes in Computer Science, pages 19–34. Springer, 2013.
[10] H. Holma and A. Toskala. Front Matter. John Wiley & Sons, Ltd, 2005.
[11] E.J. Vergara and S. Nadjm-Tehrani. Energy-aware cross-layer burst buffering for wireless communication. In Future Energy Systems: Where Energy, Computing and Communication Meet (e-Energy), 2012 Third International Conference. IEEE, 2012.
[12] Bert Hubert. Manpage of tc. linux.die.net/man/8/tc, 2014. Ac-cessed May 16th 2014.
[13] S. McCanne V. Jacobson, C. Leres. Tcpdump. http://www.tcpdump. org/manpages/tcpdump.1.html, 2014. Accessed May 14 2014. [14] E.J. Vergara, S. Nadjm-Tehrani, and M. Prihodko. Energybox:
Dis-closing the wireless transmission energy cost for mobile devices. In Sustainable Computing: Informatics and Systems. Elsevier, 2014. [15] C. Leres V. Jacobson and S. McCanne. Manpage of pcap. www.
tcpdump.org/manpages/pcap.3pcap.html, 2014. Accesed May 16th 2014.
[16] M. Calder and M.K. Marina. Batch scheduling of recurrent applica-tions for energy savings on mobile phones. In Sensor Mesh and Ad Hoc Communications and Networks (SECON), 2010 7th Annual IEEE Communications Society Conference. IEEE, 2010.
