Institutionen för datavetenskap
Department of Computer and Information Science
Final thesis
Energy-Efficient Mobile Communication with
Cached Signal Maps
by
Rasmus Holm
LIU-IDA/LITH-EX-G--15/075--SE
2016-02-06
Linköping University Electronic Press
Upphovsrätt
Detta dokument hålls tillgängligt på Internet – eller dess framtida ersättare – från publiceringsdatum under förutsättning att inga extraordinära omständigheter uppstår.
Tillgång till dokumentet innebär tillstånd för var och en att läsa, ladda ner, skriva ut enstaka kopior för enskilt bruk och att använda det oförändrat för icke-kommersiell forskning och för undervisning. Överföring av upphovsrätten vid en senare tidpunkt kan inte upphäva detta tillstånd. All annan användning av dokumentet kräver upphovsmannens medgivande. För att garantera äktheten, säkerheten och tillgängligheten finns lösningar av teknisk och administrativ art.
Upphovsmannens ideella rätt innefattar rätt att bli nämnd som upphovsman i den omfattning som god sed kräver vid användning av dokumentet på ovan be-skrivna sätt samt skydd mot att dokumentet ändras eller presenteras i sådan form eller i sådant sammanhang som är kränkande för upphovsmannens litterära eller konstnärliga anseende eller egenart.
För ytterligare information om Linköping University Electronic Press se för-lagets hemsida http://www.ep.liu.se/.
Copyright
The publishers will keep this document online on the Internet – or its possible replacement – from the date of publication barring exceptional circumstances.
The online availability of the document implies permanent permission for anyone to read, to download, or to print out single copies for his/her own use and to use it unchanged for non-commercial research and educational purpose. Subsequent transfers of copyright cannot revoke this permission. All other uses of the document are conditional upon the consent of the copyright owner. The publisher has taken technical and administrative measures to assure authenticity, security and accessibility.
According to intellectual property law the author has the right to be mentioned when his/her work is accessed as described above and to be protected against infringement.
For additional information about the Linköping University Electronic Press and its procedures for publication and for assurance of document integrity, please refer to its www home page: http://www.ep.liu.se/.
Linköpings universitet Institutionen för datavetenskap
Final thesis
Energy-Efficient Mobile Communication with
Cached Signal Maps
by
Rasmus Holm
LIU-IDA/LITH-EX-G--15/075--SE
2016-02-06
Supervisor: Ekhiotz Jon Vergara Alonso Examiner: Simin Nadjm-Tehrani
Abstract
Data communication over cellular networks is expensive for the mobile de-vice in terms of energy, especially when the received signal strength (RSS) is low. The mobile device needs to amplify its transmission power to com-pensate for noise leading to an increased energy consumption. This thesis focuses on developing a RSS map for the third generation cellular technol-ogy (3G) which can be stored locally at the mobile device, and can be used for avoiding expensive communication in low RSS areas.
The proposed signal map is created by crowdsourced information col-lected from several mobile devices. An application is used to collect data in the mobile device of the user and the application periodically sends the information back to the server which computes the total signal map.
The signal map is composed of three levels of information: RSS infor-mation, data rate tests and estimated energy levels. The energy level cate-gorizes the energy consumption of an area into "High", "Medium" or "Low" based on the RSS, data rate test information and an energy model developed from physical power measurements. The coarse categorization provides an estimation of the energy consumption at each location. It is evaluated by collecting data traces on a smartphone at different locations and comparing the measured energy consumption at each location to the energy level cat-egories of the map.
The RSS prediction is preliminarily evaluated by collecting new data along a path and comparing how well it correlates to the signal map. The evaluation in this thesis shows that with the current collected data there are not enough observations in the map to properly estimate the RSS. However, we believe that with more observations a more accurate evaluation could be done.
Contents
List of Figures ix
List of Tables xi
Acronyms xiii
1 Introduction 1
1.1 Background and Motivation . . . 1
1.2 Purpose . . . 2
1.3 Problem Definition . . . 2
1.4 Methodology . . . 2
1.4.1 Developing the Signal Map . . . 3
1.4.2 Energy Estimation In the Signal Map . . . 3
1.4.3 Evaluating the Signal Map and Energy Map . . . 3
1.5 Related Works . . . 3
1.6 Structure of the Report . . . 4
2 Background 5 2.1 Energy Consumption In 3G . . . 5
2.1.1 3G States and State Transitions . . . 5
2.1.2 Received Signal Strength and Data Rate . . . 7
2.2 Measurement Tools . . . 8
3 Data Collection 10 3.1 System Overview . . . 10
3.2 Data Application . . . 10
3.2.1 Logging Mechanism . . . 11
3.2.2 Extending Network Monitor . . . 11
3.3 Collected Type of Data . . . 12
3.4 Collected Data . . . 13
4 Energy Estimation for 3G 16 4.1 Provided Data Set . . . 16
CONTENTS
4.3 Power Consumption of DCH . . . 18 4.4 Power Models for 3G . . . 21
5 Signal Map 25
5.1 Data Representation . . . 25 5.2 Location Accuracy . . . 27 5.3 Categorizing by Energy Consumption . . . 29
6 Evaluation 31
6.1 RSS Estimation . . . 31 6.2 Evaluation of the Energy Categorization . . . 34
7 Closing 36
7.1 Summary and Conclusions . . . 36 7.2 Future Work . . . 37 Bibliography 39 Appendices 42 A Applications Used 43 B Configuration Files 45 C Energy Samples 50 D Regression Lines 55
List of Figures
2.1 The different states and transitions that are available for user devices in 3G. . . 6 2.2 The power trace and marked RRC states for a ping sending. 6
3.1 System overview. . . 10 3.2 Overview of Network Monitor data collection process. . . 11 3.3 Number of data points when there is no WiFi connected. . . 14 3.4 Heatmap mapping logged data points. . . 15
4.1 Energy consumption of pings with different RSS. . . 18 4.2 Box plot of the power consumption in DCH resulted from
the ping experiments. The median value is red and edges of the inner box represents the 25th and 75th percentiles. The upper and lower edges are the 95 % confidence interval. . . . 19 4.3 Average DCH power during 20 seconds of unlimited data rate
download and 20 seconds upload. . . 20 4.4 Summary of three different downlink data rates per RSS, 100
kbps, 1 Mbps, and unlimited. . . 21 4.5 2nd degree exponential regression of small pings. The red
error bars per RSS shows the standard deviation of the power readings, the cross is the mean and the black dots are power readings. . . 22 4.6 Plot of residuals for the regression of the DCH tail. . . 23 4.7 Exponential second degree regression of unlimited uplink. . . 24
5.1 A small subset of data points in the signal map for 3G when no WiFi is connected. . . 26 5.2 Mean STD of all the squares with observations when using
different dimensions. . . 27 5.3 Total number of squares in small data set for different
dimen-sions. . . 27 5.4 CDF of the accuracy for the small data set. . . 28 5.5 weighted average RSS of the small data set. . . 29
LIST OF FIGURES
5.6 Box plot with power readings from all the experiments under analysis in section 4.3. . . 30
6.1 Collected new data along the small path mapped with GPS. 32 6.2 Observations per square in the signal map along the small
path. . . 33 6.3 Observations per square for the new data along the small
path. . . 33 6.4 RSS per square in the signal map VS the newly collected data. 33
C.1 Different download rates at a single location with signal strength -50 dBm. . . 51 C.2 Different download rates at a single location with signal strength
-70 dBm. . . 52 C.3 Different download rates at a single location with signal strength
-90 dBm. . . 53 C.4 Different download rates at a single location with signal strength
-110 dBm. . . 54
D.1 Exponential second degree regression of empty pings. The red error bars per RSS shows the STD of the power readings and the cross is the mean. . . 55 D.2 Plot of residuals for the power model of the DCH tail. . . 56 D.3 Exponential second degree regression of unlimited downlink. 56 D.4 Plot of residuals for the unlimited downlink. . . 57 D.5 Exponential second degree regression of unlimited uplink. . . 57 D.6 Plot of residuals for the unlimited uplink. . . 58 D.7 Exponential second degree regression of 1 Mbps downlink. . 59 D.8 Plot of residuals for 1 Mbps downlink. . . 60 D.9 Exponential second degree regression of 100 Kbps downlink. 60 D.10 Plot of residuals for 100 Kbps downlink. . . 61
E.1 Small data set with energy levels "High"=3, "Medium"=2, and "Low"=1. . . 62
List of Tables
3.1 Statistics of the collected data set. . . 14
6.1 Comparison of squares with different STD and observations. The x and y refer to the index of the square in the signal map (see figure E.1). . . 32 6.2 Measured and estimated values for the selected points. . . . 35
Acronyms
3G Third generation of mobile telecommunications technology. v, ix, 2, 3,
5, 6, 8, 11–14, 16, 26, 36, 43
AP Access Point. 12
BS Base Station. 5, 7, 8, 17, 21
CDF Cumulative Distribution Function. ix, 28
DCH Dedicated Channel. ix, 8, 16–19, 21, 22, 29, 34–36 FACH Forward Access Channel. 7, 16–18
FTP File Transfer Protocol. 11
GPS Global Positioning System. 3, 4, 11, 12, 27, 28, 32, 37 GSM Global System for Mobile Communications. 12 IDLE IDLE State. 16, 17
IM Instant Message. 1
LTE Long Term Evolution. 11–14, 37 OS Operating System. 1
PPMCC Pearson product-moment correlation coefficient. 31–33, 37 QoS Quality of Service. 20
RLC Radio Link Control. 7
Acronyms
RNC Radio Network Controller. 5–8
RRC Radio Resource Control. 5–8, 13, 17, 34
RSS Received Signal Strength. v, ix, x, 1–4, 7–9, 11–14, 16–25, 27–31,
33–38, 50, 55
SOM Self-Organizing Map. 26, 37
STD Standard Deviation. ix–xi, 23–27, 31, 32, 55 TPC Transmit Power Control. 7, 8
UE User Equipment. 4–8, 12, 13, 16–18, 20, 21, 25, 27, 29, 31, 34, 37 UMTS Universal Mobile Telecommunication System. 5
UTRAN UMTS Terrestrial Radio Access Network. 5 WiFi Wireless local area network. ix, 8, 11–14, 26–28, 37, 43
1.
Introduction
This report describes the work performed in the context of a Bachelor the-sis project at the Real-time Systems Laboratory within the Department of Computer and Information Science at Linköping University. The project was performed by the author in order to complete a Bachelor degree in Computer Engineering.
This chapter introduces the project to the reader. First the context and goals of the thesis are described as well as the planned approach. The chapter also presents related scientific work, the structure of the report and the intended audience of the thesis.
1.1
Background and Motivation
In the last decade the development of mobile devices has been remarkable, with an increasing amount of mobile applications and data traffic. The increasing amount of data traffic from various applications is expensive in terms of energy, especially over cellular networks. The management tech-niques employed cause the cellular network interface on the mobile device to be active for some time even after it has stopped sending data. Additionally, energy consumption is even more expensive under low RSS [1, 2].
The applicability of energy-efficient methods for mobile applications dif-fer depending on the service provided as well as the expected quality of experience of the user. For example an instant messaging (IM) application has different requirements compared to a mail application. The latter might only need to synchronize within a 30 minutes interval, while a IM should be instant. This difference of delay tolerance is something that the device must consider when evaluating whether or not to transmit.
We propose to employ a signal map (the mapping of expected RSS based on location) and energy categorization (the mapping of expected energy consumption based on location) in the mobile device to gain knowledge of the high consuming locations for mobile communication. We believe that by being aware of the potential RSS in a given area the applications and operating system (OS) can adapt and save energy. Especially since the RSS can change substantially from one place to another.
CHAPTER 1. INTRODUCTION
1.2
Purpose
The vision in this thesis is to estimate the approximate energy consumption of different areas in a map. For this purpose we develop a signal map and energy categorization for 3G. These can be stored locally in a device and be used to avoid energy expensive communication under poor RSS areas in order to reduce the energy consumption.
The creation of the signal map require at least a data set with RSS read-ings and their corresponding location. To estimate the energy consumption at each location of the map, there is a need to create a model that maps the energy consumption from the RSS and data rate data. The mapping of the estimated energy and location is called energy categorization in this thesis. Since existing RSS databases (e.g., OpenSignal [3]) only provide limited access to their collected data for the end-user, we employ our own data collection application. Collecting as a single user limits the amount of the points collected, and thus we enable collaborative data collection with the developed application. We chose to only consider 3G due to restrictions in time for this bachelor thesis. However, this work can easily be extended to other technologies since we collect data for all of them.
1.3
Problem Definition
In order to develop the signal map and energy categorization 4 different goals are identified:
1. Collect data and from the collected data set create a received signal strength map for 3G that can be stored locally. We also need to model the signal map in a way that is easy for device to interpret. The signal map should provide location, RSS and data rate characteristics.
2. Develop an energy model using data rate and RSS as input to estimate the energy consumption of the different areas in the signal map.
3. Evaluate the signal map by collecting new data and comparing the observations to the ones in the signal map.
4. Evaluate the energy categorization by collecting new data and com-paring the observations to the ones in the energy categorization.
1.4
Methodology
This section provides an overview of the methodology used during the thesis, divided according to the items in the problem formulation.
1.5. RELATED WORKS
1.4.1
Developing the Signal Map
The first step is to collect all the data for the signal map with connection type (ensures that 3G is in use), location data (using GPS), RSS and measured data rate. For this an application is developed to collect data from different devices. GPS data, RSS and data rate are collected on multiple devices using the developed application. The devices which run the application creates the total collected data set by collaboration.
The data rates are measured in the application by uploading and down-loading files of known size to a server. It makes the calculation easy since it will just be a matter of knowing the transmission time. RSS and network type are exposed directly to the device through the network.
The location data is used for mapping the collected data points. As a first attempt, the collected data points are mapped onto geographical squares of a determined size where all the points are grouped. A weighted average method is proposed in order to provide representative values per square out of the collected data.
1.4.2
Energy Estimation In the Signal Map
In order to provide energy estimation based on the geographical position of the mobile device, we develop an energy model from physical measurements. Using the physical measurements we analyze the impact of RSS and data rate on power consumption and model this using regression analysis.
For categorization we use a simple model to divide the areas in the map into categories of "High", "Medium" or "Low" energy consumption. We consider the RSS of the signal map and the model created from the regression analysis for dividing areas into categories in the energy categorization.
1.4.3
Evaluating the Signal Map and Energy Map
Evaluating the signal map is analogous to evaluating the prediction of the RSS in the area. Thus, collecting a trace consisting of RSS and data rate and comparing it to the developed signal map can provide the accuracy of the map.
In order to evaluate the energy estimation, RSS and packet traces are collected in mobile devices in three specific point of the signal map, ("High", "Medium" and "Low") and the resulting energy consumption is compared.
1.5
Related Works
Energy consumption in mobile devices has been of particular interest in recent years. It has been shown that the signal strength in wireless networks has an significant impact on the energy consumption [2, 1]. Ding et al. [1] further investigates the impact of different transmission types (e.g., data
CHAPTER 1. INTRODUCTION
size, round-trip time) and demonstrate a system-call-driven approach for optimizing transmission. In this thesis we do not implement any algorithm for deciding when to transmit but our work could extend their work by enabling RSS prediction.
Schulman et al. [2] show that energy per bit increases drastically when transmitting over cellular networks and the RSS is low. In order to save energy the mobile device can schedule data transmissions when the RSS is high. By implementing an algorithm (Bartendr) that calculates the en-ergy cost based on RSS and location, applications can reduce their enen-ergy footprint. This thesis extends the work by utilizing a locally cached signal map created by many users. By modifying Bartendr to use a local signal map there would be no need to constantly fetching the signal map data. The computation would be lighter since most computation is done when creating the signal map.
GPing-Pair [4] is proposed as a low cost bandwidth estimator. Using GPing-Pair it is possible to achieve low cost bandwidth estimation for cellu-lar networks. Estimate the bandwidth like GPing-Pair is cheaper than what is done in this thesis.
TCP goodput [5] as an alternative way to present the actual transport quality of cellular networks. This article compares alternatives to signal strength for mapping cellular networks. For this thesis we use RSS but this approach is interesting as an alternative.
A similar work to ours has been done in Ou et al. [6], however while the article mentions crowdsourcing to collect the data they do not do it. What is proposed are different collaborative scheduling algorithms where User Equipments (UEs) collaborate. For example when obtaining location of the UE it can be improved by looking at previous traces in the area. We only use the GPS for determining location. This work extends their work by collecting a real data set from several users and categorizing the energy consumption.
1.6
Structure of the Report
The structure of the report is as follows: Chapter 2 introduces the back-ground and terminology needed for understanding the thesis. Chapter 3 describes the data collection process and the application used. In chapter 4 the energy estimation model based on RSS and data rate is described. Chapter 5 presents the signal map. In chapter 6 the signal map, energy estimation and representation of the collected data is evaluated. In chap-ter 7 conclusions of the work done in this thesis, possible future works are discussed.
2.
Background
This chapter gives a basic understanding of the terminology used in this thesis and a background of current technology used.
2.1
Energy Consumption In 3G
Wireless interfaces on mobile devices consume a large fraction of the total energy. The amount of energy consumed is not proportional to the data transmitted nor the same for all wireless interfaces [7]. This section in-troduces the main factors that impact the energy consumption of the 3G interface for a mobile device.
2.1.1
3G States and State Transitions
The third generation of mobile networks (3G) is a collection name for several different telecommunication technologies such as Universal Mobile Telecommunication System (UMTS). UMTS devices establish network con-nection through the UMTS Terrestrial Radio Access Network (UTRAN). The UTRAN in turn consists of Base Stations (BSs) and Radio Network Controllers (RNCs). An RNC is responsible for the BSs and control their network resources. The network operator controls the state of the UE using the Radio Resource Control (RRC) protocol.
CHAPTER 2. BACKGROUND Idle Standby CELL_PCH URA_PCH Shared CELL_FACH Dedicated CELL_DCH T3 T2 T1 Performance P o w er consumption
Figure 2.1: The different states and transitions that are available for user devices in 3G.
The UE implements a state machine controlled by the RNC. Figure 2.1 shows the different transitions and states that a 3G enabled device imple-ments. The performance in the x-axis indicates the response time and data rate, where higher performance means lower response time and increased achievable data rate. The power consumption in the y-axis indicates the power usage between the different states [8].
Figure 2.2: The power trace and marked RRC states for a ping sending.
Figure 2.2 outlines the different RRC states and inactivity timers for a ping with power readings. The energy consumed during the inactivity timers per RRC state in the figure is called the "Tail Energy". In the dedicated state (CELL_DCH), the UE has a dedicated physical channel allocated for it which means maximum data rate is available, but also the highest con-sumption. A shared channel is used in the shared state, (CELL_FACH), which consumes less power than in CELL_DCH. In both the dedicated and
2.1. ENERGY CONSUMPTION IN 3G
shared state the UE is constantly connected to the RRC. It is possible to transmit data in CELL_FACH but at low data rates. In the idle state the UE is not connected, but still able to check for downlink packets. The im-plementation of the standby state (URA_PCH or CELL_PCH) is optional for operators and has a power profile similar to the idle state. The main difference between the idle and standby state is that the standby state al-lows the UE to transition faster to states with higher performance since it maintains a connection to the RNC [8, 1].
The arrows in figure 2.1 indicate the possible transitions for the UE. The state transitions are controlled using inactivity timers and data buffer occupancy.
The RNC demotes the UE to a lower performing state using inactivity timers, T1-T3 in figure 2.1. When the UE stays for a period of length T1 in CELL_DCH with little or no data transmission, the RNC switches the UE to CELL_FACH according the RRC protocol using the Transmit Power Control (TPC) commands. Similarly, the UE transitions from CELL_FACH to CELL_PCH after T2 and CELL_PCH to idle after T3. The inactivity timers are statically set by the carrier and start counting when there is little or no data traffic. The period during inactivity timers to the state transition is called the "Tail", and the energy consumed during that period is called the "Tail energy" [8, 1].
The RNC uses the Radio Link Control (RLC) in order to report the ob-served traffic volume. The RLC implements data buffers to trigger state transitions. When the occupancy of the buffers exceeds a certain fixed threshold these buffers trigger transitions. There are thresholds for both uplink and downlink. The thresholds depend on the current RRC state. Once the data has been transmitted the buffers are cleared [8, 1].
2.1.2
Received Signal Strength and Data Rate
The power drawn by the UE transmitter depends on the RSS and data rate. There are different techniques used to control the transmission power in order to meet certain standards (e.g., a certain bit error rate, signal to noise ratio). These standards are there to ensure quality of operation on the cellular network, both for the cell and UE.
Open Loop Power Control: In the open loop power control there
is no feedback between the UE and BS, thus the UE needs to estimate transmission power to the BS based on the information available on the network (e.g., the RSS from the BS). The UE transmitter can use this information to set the transmission power for uplink and downlink. Open loop power control is used when a UE is initially accessing the network. It has a tolerance of ± 9 dB under normal conditions and ± 12 dB under extreme conditions. The UE is responsible for setting the optimum power level.
CHAPTER 2. BACKGROUND
is feedback between the UE and BS, which is used to set the transmission power for the UE. A dedicated channel from the BS to the UE exists to provide feedback and the BS determines the optimum power for the UE.
• Inner Loop Power Control: Both inner and outer loop power con-trol is part of the closed loop power concon-trol. The UE transmitter receives TPC commands on the downlink. These commands are to make sure that the UE uplink signal is stable, and if it is not make sure that the signal is amplified. If the uplink is stronger than needed the TPC commands will tell the UE transmitter to lower the amplifica-tion. The amplification steps for the uplink is 1, 2 and 3 dB. Increasing the amplification results in an increased power consumption and vice versa for decreasing.
• Outer Loop Power Control: It ensures that the communication quality meets a certain standard, while keeping the energy consump-tion as low as possible. The uplink outer loop power control sets and maintains the targeted signal quality for the uplink. Downlink outer loop power control listens on the network from the UE receiver. It is responsible for changing modulation based on feedback from the RNC. Higher modulation gives higher data rate but also draws more power. The difference between outer and inner loop power control is that for inner the UE gets feedback from an outer node instead of the RNC. This results in a faster feedback for inner since it can adapt at a higher frequency. Since the feedback is faster it also respond faster to rapid signal strength changes.
2.2
Measurement Tools
This section describes the measurement tools used in this thesis.
• EnergyBox [8], is a packet trace driven application for estimating the energy consumption in WiFi and 3G for mobile devices.
It models the operational states of the interfaces and for a given packet trace and of the network, such as the inactivity timers and buffer thresholds described in section 2.1. The tool outputs the estimated energy consumption and an estimation of when the UE is in different RRC states.
It uses a constant power level per RRC state to simplify setting the power levels. However, this means that the energy estimation provided by EnergyBox does not consider RSS or data rate.
EnergyBox is used for evaluating the energy footprint of a mobile application or system software by extracting RRC states from packet traces. This information is combined with our power model described later in chapter 4 for the DCH state to account for the varying RSS.
2.2. MEASUREMENT TOOLS
• Network Monitor, an open source Android diagnostic tool which periodically tests the network connection. It is one of the many net-work monitor applications in the Google Play Store [9]. It is capable of logging additional parameters such as cell tower’s id, location, network type etc.
This application is used and modified in this thesis for collecting data points. Also used for logging RSS when evaluating the signal map.
• Shark [10], is an Android application that utilizes tcpdump [11] to capture network traffic on the users network interface.
It was used for capturing packet traces for evaluation of the power models. The packet traces are fed to EnergyBox for energy estimation.
3.
Data Collection
In this chapter the process of collecting data for the signal map is described. The system overview is first provided describing the data collection mobile application. Then, the collected type of data is presented.
3.1
System Overview
Server Smartphone Smartphone Smartphone Smartphone One-w ay One-w ay One-w ay One-w ayFigure 3.1: System overview.
The system is designed to collect data from many users. Figure 3.1 shows the system overview where each client is named as Smartphone. The client contains a local database. This database is periodically backed up to a centralized server along with an anonymous unique identifier. The server then maintains the copies of the client databases and periodically merges these into one big database, as indicated in figure 3.1.
3.2
Data Application
This section describes the data collection application and the methods used to collect data.
3.2. DATA APPLICATION
3.2.1
Logging Mechanism
The logging application needs to enable automatic collection in an external server. The application is also required to collect GPS location, RSS, data rate, network type (e.g. WiFi, 3G, 2G, LTE) and which interface is used for data transfers. Since data rate is required speed tests needs to be performed by the application.
There are several different Android network logging applications. We selected an open source network logging application that conformed to our requirements and that could be modified for our project. The logging ap-plication used in this thesis work is based on Network Monitor1.
Network Monitor is selected because the other considered applications, OpenSignal2, Network Log3, do not conform to our requirements. OpenSignal does not allow storing of the requested data in our database. Network Log only logs the network data transfers not location or RSS, but is an open source application. Network Monitor is an open source project which means we can modify it to fit the project needs. The code written to adapt and extend Network Monitor to our project needs is available as open source as a fork of the original [12].
Network data Database External service
Log Export
Figure 3.2: Overview of Network Monitor data collection process.
Figure 3.2 shows the basic task that Network Monitor performs. The logging is done every statically set interval and the logged parameters can be configured from the application’s graphical user interface.
The collected database at the mobile device can be exported using differ-ent external services, ranging from mail, WebDAV, FTP to HTTP clidiffer-ents. Automatically exporting the data is only available for mail, which is lim-ited since most mail servers will block email from the built in mail client. Exporting the database is a required functionality for our project and thus the next section explains the extension we develop to export the data to our servers.
3.2.2
Extending Network Monitor
The developed version of Network Monitor is available on GitHub [12]. The original network monitor application is extended to enable the periodic ex-port of the collected database to a server. The File Transfer Protocol (FTP) is selected for the current implementation of automatic synchronization since
1https://play.google.com/store/apps/details?id=ca.rmen.android.networkmonitor 2http://opensignal.com/
CHAPTER 3. DATA COLLECTION
it is easy to implement both in the server and client sides. This facilitates the fast setup of the logging application.
Additional bug fixes and extensions are performed to improve the data rate tests:
• Bugfix, enable higher data rate tests for download and upload tests so they can estimate speeds higher than 1.5 Mbps which are common in 3G.
• Higher RSS sampling rate for data rate tests.
– Allows the use of a higher RSS sample rate while speed tests
are enabled. This means that even though a speed test needs 10 seconds to finish we can concurrently collect RSS during that time (e.g., every second).
– Enables the UE to only perform a new speed test when the
net-work or RSS has changed by a given amount of dBs. This means that users with restricted data plan can still provide some data rate tests without exceeding the data limit while only collecting data in a single location.
3.3
Collected Type of Data
This section describes the different characteristics that are considered when collecting the data for the signal map. Data is collected for 3G, WiFi and LTE, but analysis is only done for 3G due to restrictions in time for this bachelor thesis. We describe the method to collect the different parameters too.
• Location: The RSS values need to be mapped to a location in the signal map. There are several options to obtain the location of a mo-bile device such as Global System for Momo-bile Communications (GSM) triangulation, WiFi and GPS. Their accuracy, ease of implementation as well as energy consumption differ.
Triangulation requires that the UE has several nearby cell towers to pinpoint its location. Otherwise it is very cheap in terms of energy consumption since it only relies on the GSM network. GPS has the highest accuracy and reliability but expensive in terms of energy con-sumption. WiFi positioning relies on measuring the RSS and knowing the location of nearby WiFi Access Points (APs). The accuracy is very good for WiFi positioning but it is limited by overall coverage and restricted access.
We chose GPS because this is a preliminary selection to speed up the data collection. Adopting more advanced methods to keep high accuracy while reducing energy consumption when collecting data such as done in the work by Yin et al. [13] is considered a future step.
3.4. COLLECTED DATA
• RSS: The RSS is always collected by the mobile device since it is utilized to decide when the mobile device needs to either amplify the signal to the cell tower or change the transmission power.
• Data Rates: The data rate impacts the transmission time and thus energy consumption.
We measure the maximum data rate for uplink and downlink by send-ing/receiving a file of a given size and measuring the transfer time. Dividing the file size by the transfer time we get an estimation of the data rate.
The size of the transmitted data also matters and generally a larger file size results in a more accurate estimate. The reason for this is that errors and delays of a transfer are distributed over a larger set and thus estimates the real transfer speed better. The trade off is that exchanging larger files takes longer time to transfer and also more data usage.
• Network type: Since a mobile device can be connected to several different types of networks it is interesting to see which network is used for data transmission. This comes in handy when evaluating the transfer speeds since just checking the connection to a mobile network such as 3G or LTE does not mean that the data is being transmitted over that specific interface. It is for example possible to be connected to both LTE and WiFi at the same time.
• Cell tower id: Identifier for the cell tower the UE is connected to. It is collected since it is always present and it might explain if there are strange aspects in the signal map. This information could also be used in the future for improving the location readings in the collected data set and filtering depending on cells. Another interesting point is to see when the UE switches between cell towers.
• Carrier: Different carriers can have different policies for how the UE interacts with the cellular network. It is up to the carrier to determine the inactivity timers for the RRC states. In the future this could be used to do energy estimation per carrier. A certain user only uses one operator and UEs can experience different RSS at one location depending on the operator. Also inactivity timers impact the time spent in each RRC state.
3.4
Collected Data
Collection of data is done in a collaborative way, this means that there are several users running the logging application which periodically sends an update to the server with new data.
CHAPTER 3. DATA COLLECTION
Data points Cell towers Users Days Count 830929 1258 15 78
Table 3.1: Statistics of the collected data set.
The application was run by 15 people during 78 days collecting a total of 830929 number of data points, as it is shown in table 3.1.
Figure 3.3: Number of data points when there is no WiFi connected.
Figure 3.3 shows the distribution of the RSS values for LTE and 3G. Since the values are diverse, we consider that the requirement for collecting a diverse dataset is satisfied.
3.4. COLLECTED DATA
Figure 3.4: Heatmap mapping logged data points.
Finally, since the data collection is collaborative, it is also interesting to show the users the information they collect. Figure 3.4 shows the developed heatmap in order to visualize the collected raw data points.
The heatmap contains all the logged data points. A data point is rep-resented by a small green circle with a red dot in the middle. Data points that are close are grouped together and represented by a larger red area. This heatmap is meant as an incentive since the user can see the actual data collection progress. The code for setting up the site can be found at the author’s GitHub repository [14].
4.
Energy Estimation for 3G
This chapter describes the power model created in this thesis which is used to estimate the energy consumption of the UE ("High", "Medium" or "Low") at a given location based on the RSS data. The power model estimates the power consumption of the UE for a given RSS value. The model is derived from a dataset containing packet traces of 3G communication with the cor-responding physical power measurements in a mobile broadband module. The dataset was collected by the supervisor.
4.1
Provided Data Set
The dataset collected by Ekhiotz Jon Vergara contains packet traces col-lected while performing systematic data transfers, using a broadband mod-ule (Ericsson F3307). The first set of experiments were performed to obtain the power profile of the 3G for different state transitions:
• Five downloads and five uploads of a 300 kB file with a 15 seconds interval between transfers. Used for analyzing state transitions and power in different states.
• 20 pings with 0 byte data (i.e., only headers) and a 15 seconds pause between transfers. Used for analyzing the transition to FACH (i.e., the DCH tail).
• 20 pings with 1000 bytes data and a 15 seconds pause between trans-fers. Used for analyzing the transition to DCH.
The 15 seconds pause in the previous experiments were enough for the UE to return to IDLE between the transfers. The following tests were performed to study the power and data rate relation:
• Data rate test with different limits to data rate uplink and downlink, with a transfer duration of 20 seconds per data rate (i.e., 20 seconds transferring data at a fixed approximate data rate) and a 3 seconds pause between transfers. The different data rate limits was set using cURL to 1 Mbps, 100 kbps, 10 kbps, and 1 kbps. One test run without
4.2. POWER MEASUREMENT ANALYSIS
any limit for uplink and downlink was also performed to measure the maximum achieved data rate.
Finally, all the above experiments are performed a single time each for different RSS values: -50, -60, -70, -80, -90, -100, and -110 dBm. The experiments were performed at selected locations were the UE experienced the mentioned RSS with as little variation as possible.
The experiments were performed (before this thesis) using the command line tools cURL [15] for generating traffic and tcpdump [11] for capturing traffic. The data rate limits were set in cURL. Capturing the traffic enabled the analysis of the data trace for computing data rate and checking that the test data was indeed the only data transmitted. The power measurements were captured using the same setup as in Vergara et al. [8].
The resulting data set per experiment contained a packet trace and the corresponding power measurement (an array with a power reading every millisecond), which is analyzed in the following section.
4.2
Power Measurement Analysis
In this section we analyze which RRC states were the RSS impact signif-icantly are interesting for analysis and modeling based on the samples we have. We isolate the RRC states where the RSS significantly impact the power consumption in order to model it using the measurements. We parse the RRC states from the power traces with graph analysis and static thresh-olds. The threshold values are only there for getting a consistent isolation process of the RRC states.
We aim to model the energy consumption dependency regarding RSS and data rate variation. The RRC states to consider for this analysis are DCH and FACH. Since no data transmission can be performed in IDLE, this state is not considered. Data rate impacts the power consumption the energy consumption while transmitting or receiving data. The UE performs power control using RSS as a metric, and thus for low RSS values the UE increases the transmission power to the BS.
CHAPTER 4. ENERGY ESTIMATION FOR 3G
Figure 4.1: Energy consumption of pings with different RSS.
Figure 4.1 shows that the energy consumption in FACH does not change notably with RSS. To simplify the analysis only fragments of experiments were all data transfer is done in DCH are chosen. This means that only the power consumption in DCH is considered in the energy model.
The state of the UE is easily distinguished in the power trace and thus isolating DCH from the rest is trivial. Since power readings are volatile it requires manual interference for locating the static thresholds starting and finishing point for the DCH states. We manually select power thresholds to isolate the DCH state, i.e., if the power is higher than 1 W the state is assumed to be DCH. If the power is between 0.3 and 1 W the state is FACH. Then, using the inactivity timers for each state we can isolate the DCH tail from the DCH state.
4.3
Power Consumption of DCH
In order to develop the power model first the provided power trace fragments of DCH collected at different RSS values are analyzed using MATLAB [16]. We selected the traces resulting from the empty pings and data rate with different data rate limits experiments for the analysis. The empty pings are used for analyzing the DCH tail since there is only a small transmission. The power traces resulting from the different data rate tests are used to analysis of data transmission in DCH. All uplink and downlink tests are measured separately when there is data transmission. The traces for the following RSS values were analyzed: -50, -70, -90, and -110 dBm. We chose to limit to these RSS since we believe they are enough to give a general view of the energy consumption for our simplistic approach.
4.3. POWER CONSUMPTION OF DCH
Figure 4.2: Box plot of the power consumption in DCH resulted from the ping experiments. The median value is red and edges of the inner box represents the 25th and 75th percentiles. The upper and lower edges are the 95 % confidence interval.
DCH Tail Power: We analyze the power measurements from the empty
pings experiment in the provided data set to show the power of the DCH tail. The average power for the DCH state when not sending data varies with the RSS. Figure 4.2 shows that the energy consumption in DCH does not change drastically, since the transmissions are so small that the power is dominated by the tail power.
It also shows that the power consumption is quite similar between -50 dBm and -70 dBm, slightly increasing at -90 and greatly increasing at -110 dBm. The slight increase in mean power between -70 dBm and -90 dBm is about 1.8 %. There is a clear difference between -90 and -110 dBm where the percentage increase is 5.2 %.
CHAPTER 4. ENERGY ESTIMATION FOR 3G
Figure 4.3: Average DCH power during 20 seconds of unlimited data rate download and 20 seconds upload.
DCH max Uplink/Downlink Power: In order to model the
maxi-mum power consumed we use the tests with maximaxi-mum data rate for down-load and updown-load. The data is analyzed from the test of downdown-loading or uploading a file for 20 seconds without any data rate limit. We analyze only the power during the upload/download and thus the tail power is not accounted in this test.
Figure 4.3 shows the average power consumption for the download and upload at different RSS. The power consumption is roughly the same for -50 dBm, -70 dBm and slightly higher at -90 dBm. Between -90 dBm and -110 dBm there is an increase of 21.8 % according to the collected dataset.
Similar results are observed for uploading data. The power consumption is similar at -50 and -70 dBm, slightly increasing at -90 dBm. The increase between -90 and -110 dBm is larger for uploading, consuming 42.4 % more power. Uploading consumes more power since the UE transmitter needs to amplify its signal when sending, in order to guarantee a certain QoS.
4.4. POWER MODELS FOR 3G
Figure 4.4: Summary of three different downlink data rates per RSS, 100 kbps, 1 Mbps, and unlimited.
DCH Different Downlink Rates Power: In order to study the
in-crease of power when increasing the data rate, we analyze the power of the DCH state when downloading data at different data rates. The data is an-alyzed from the test of downloading or uploading a file for 20 seconds. We analyze only the power during the download and thus the tail power is not accounted in this test.
Figure 4.4 shows that increasing the data rate increases the power con-sumption at all RSS. However, the increase is more significant at low RSS values. While the power consumption at -50 or -70 dBm does not increase significantly, the increase in power from 100 kbps to 1 Mbps is slightly higher at -90 dBm. The increase is substantial at -110 dBm.
The unlimited download power consumption is strange at -110 dBm, since it has a higher download rate and lower power consumption than at 1 Mbps. Additionally, the achieved data rate is not the highest at high RSS values. This may be due to that data rate changes a lot depending on location, the load of the BS which the UE is connected to, and the scheduling of the BS.
Power consumption is important to model for different RSS. While the DCH tail power is similar until -90 dBm, it drastically increases for -110 dBm. The increase of power from transmitting in DCH is also more signifi-cant for lower RSS. Higher data rates also increase the power consumption especially in low RSS. In the next section we will model power based on data rate and RSS.
4.4
Power Models for 3G
The goal of this section is to create a model for estimating the energy con-sumption in DCH based on RSS and data rate. This model is later used
CHAPTER 4. ENERGY ESTIMATION FOR 3G
with EnergyBox to enable estimation for different RSS locations. For the model we create power models which estimates energy consumption for dif-ferent parts of the DCH, the transmission and the tail. The total estimated energy consumption for DCH is the power model. The power model is a combination of the power models for the tail and transmission. The DCH tail only depends on RSS while the transmission depends on both RSS and data rate. The power model for the tail only has one function since it only depends on RSS. The power model for transmission is a combination of sev-eral functions, one for uplink and three for different downlink rates. We only have one for uplink since we do not have enough physical measurements for different uplink rates to make several.
In order to estimate the energy consumption for a given location, we consider all the provided power measurements from the previous experiments per RSS. Each measured sample per RSS is considered with all of its power readings. The RSS values considered are the same as for previous section analysis: -50, -70, -90, and -110 dBm.
For estimation of the power consumption per RSS we use regression anal-ysis with the provided data set as input. The DCH tail energy is modeled with 2nd degree exponential regression. Data rate is modeled with expo-nential second degree regression, one function per analyzed rate (unlimited uplink, unlimited downlink, 1 Mbps downlink, and 100 Kbps downlink). The choice of regression model is based on analysis from previous section and testing different regression models for the provided data set.
Power Model for the DCH Tail: We analyze the fit and residuals to de-termine if the fit is good.
Figure 4.5: 2nd degree exponential regression of small pings. The red error bars per RSS shows the standard deviation of the power readings, the cross is the mean and the black dots are power readings.
4.4. POWER MODELS FOR 3G
P owerT ail(RSS) = 0.003 ∗ e(−0.037∗RSS)+ 1.152 ∗ e(0.001∗RSS) (4.1)
Figure 4.6: Plot of residuals for the regression of the DCH tail.
The residuals in figure 4.6 do not show any apparent pattern between the data points. This is good since it indicates that the points do not depend on some underlying function. Figure 4.5, the regression line of function 4.1, has a R-square value of 0.088 which means that the fit estimates the data set by 8.8 %. A low R-square value is expected since power readings per sample are volatile. Root-Mean-Square Error (RMSE) of the fit is 0.11 which means that the Standard Deviation (STD) of errors in the fit is small. Since the RMSE is low the points that do not fit the regression line are still close on average. The fit estimates the mean value per RSS of the power readings. We consider the fit good since it matches the expected behavior, estimates the mean per RSS, there is no apparent pattern in the residuals and there is a low error on average for points outside the fit.
Power Model for Transmission In DCH: For the data rates we do not show any residuals since it has the same behavior as for the pings. We chose to only show the uplink regression line instead of all. Regression lines for other data rates can be seen in Appendix D.
CHAPTER 4. ENERGY ESTIMATION FOR 3G
Figure 4.7: Exponential second degree regression of unlimited uplink.
P owerU p(RSS) = 0.000002 ∗ e(−0.121∗RSS)+ 1.143 ∗ e(−0.0004∗RSS) (4.2) Figure 4.7 shows the regression line of function 4.2 for unlimited uplink with the energy sample data points. The regression line has a R-square value of 0.85. This is surprising since it means that the power readings for unlimited uplink were quite stable. RMSE of the fit is 0.20 which means that the STD of errors in the fit is small. The fit estimates the mean value per RSS of the power readings. The regression line behave as expected based on previous analysis, with an exponential increase of power consumption for low RSS. We consider the fit good since it matches the expected behavior, estimates most points, no apparent pattern in residuals and low error on average for points outside the fit.
Since we have limited samples for uplink the estimation is always going to be the maximum. For the downlink however we apply linear interpolation for data rates between the regression lines.
Using the power model is easy for upload (P owerU p(RSS)) and tail (P owerT ail(RSS)) since only the RSS is needed for calculation. If we consider a upload with the RSS value -80 dBm, P owerU p(RSS) would give the following estimation:
P owerU p(−80) = 0.000002 ∗ e(−0.121∗−80)+ 1.143 ∗ e(−0.0004∗−80)≈ 1.21W (4.3) If we instead consider a download with an RSS value of -90 dBm and a data rate of 0.5 Mbps. We need to do linear interpolation to get the value with P owerDown1M bps(−90) and P owerDown100kbps(−90) seen in Appendix D. P owerDown(−90) = 0.4 0.9∗P owerDown1M bps(−90)+ 0.5 0.9∗P owerDown100kbps(−90) (4.4)
5.
Signal Map
This chapter presents the signal and energy categorization created using the data collected from the collaborative application. Each logged entry in the collected data contains all parameters described in section 3.3. The goal of the signal map is to provide the UE with a representation of the collected data. The signal map is a simple statistical precomputed model of the collected data from different users to provide a look up table to the UE. This benefits the UE by reducing the needed computation and storage to identify good or bad locations in the map.
Since there is a lot of data points in the signal map, many of which over-lap, there is a need for providing a simplification of the data representation. Section 5.1 explains a coarse grid based method for data representation. The location accuracy of the collected data points differ in the data set, and sec-tion 5.2 explains a weighting algorithm for prioritizing high accuracy data points. Section 5.3 describes the energy categorization and how it relates to the signal map.
5.1
Data Representation
In this report we focus the efforts on a representative area of the total data set, but our approach can easily be extended to a larger area without any problem. Figure 5.1 shows the selected area and the data points. The area corresponds to the university campus and the surroundings containing 47529 data points which is a 5.7 % of the total collected data.
The signal map simplify the data set by grouping data points with similar characteristics. The idea is by combining similar observations to a single one we reduce the search space.
In order to characterize the data and provide a simple representation to group the data we consider a simple approach. The area is divided into a regular grid of a given size were the RSS mean and standard deviation (STD) of observations within each square is provided. Observations with similar characteristics are those that are located within the same square in our simple approach.
CHAPTER 5. SIGNAL MAP
Figure 5.1: A small subset of data points in the signal map for 3G when no WiFi is connected.
• Self-Organizing Maps (SOMs): SOMs are a unsupervised learning algorithm. This technology is capable of learning which areas has similar features by being feed a data set [17, 18].
• K-Means: Clustering method for dividing a set with n points into k clusters. Choosing the number of k clusters which fits well to the data set is hard, and depends on the given data set [18].
These methods are capable of grouping data with similar features without any restriction to size or form. The trade off is however that they are more complex to implement for our data set, and expert knowledge is needed to apply the methods. Given the time limitation of a bachelor thesis, we select the simpler approach for this thesis.
Grid method: Selecting the size of the squares in the grid is a trade off.
A larger square size gives a small number of squares, but it will contain many different observations. A smaller square size increases the number of squares, but the expectation of having similar observations within each square is higher.
In order to select the dimension for each square we try with a set of dimensions. Then we evaluate each square within the grid, selecting the size that minimizes the STD of the observations. We consider square sizes of 10x10, 20x20, 30x30, 40x40, and 50x50 m.
The reason for not testing 5x5 m or smaller squares is that, the maximum achieved location accuracy of all collected data points is within 3 m. This
5.2. LOCATION ACCURACY
means that even though a point is estimated to be in the middle of a square it is not guaranteed to be in that square.
The square size is decided based on analysis of the average STD for squares of a certain size. The square size with the lowest STD is chosen.
Figure 5.2: Mean STD of all the squares with observations when using different dimensions.
Figure 5.3: Total number of squares in small data set for dif-ferent dimensions.
Figure 5.2 shows dividing the signal map into 10x10 squares provides the lowest STD. It also has a higher number of squares compared to the other considered dimensions, seen in figure 5.3, due to the smaller square size. With 10x10 squares the exposed number of data points in the test is reduced from 47529 to 3190, i.e., approximately 93 %.
Representing data points by squares is fine as a simple approach, but it is highly limiting due to the rigid form of the squares that does not correspond to the RSS. By dividing the area into squares we find two major issues:
• Squares without any data point: We need to provide some heuris-tics to assign a value to such squares, or leave them without any value. We choose the latter in our thesis for simplicity.
• Location accuracy: The RSS changes drastically indoors. In our approach we consider the simplifying assumption that a mobile device will be connected to a WiFi network indoors. This means that we do not consider points where the UE is connected to WiFi. It does not mean that we completely avoids indoor readings since those can happen even without WiFi.
5.2
Location Accuracy
Data points within our collected data set can have different location accu-racy. GPS provides the location of the UE with different accuaccu-racy. Since the location is used for mapping data points in the signal map, points with
CHAPTER 5. SIGNAL MAP
high accuracy should have a high significance. In our small subset used in this chapter the location accuracy of the data points varies between 3 and 1784 m. To mitigate the impact on RSS in the signal map of data points with low location accuracy we implement a weighting algorithm.
Figure 5.4: CDF of the accuracy for the small data set.
Figure 5.4 shows the GPS accuracy distribution for the points in the considered area. The high amount of low accuracy points is most likely due to indoor readings, since the GPS has a hard time estimating the location when indoors. We mitigate some of this by not counting points where WiFi is active and connected. The location accuracy median value of the whole collected data set is 87 m and 1439 m for the small subset used for analysis in this chapter. This means that about half of the samples show low location accuracy.
The reason for keeping data points with low location accuracy is that we want squares with only low accuracy data points to contain at least some estimation of the data set. This does however mean that points with high accuracy need to have a heavy enough weight that low accuracy points do not impact the result significantly.
In order to reduce the influence of the high GPS error data samples we consider the following heuristic: provide less weight to the samples with lower location accuracy. We choose the following simple weighted mean algorithm since it is easy to implement and favors points with lower error values:
Weighti= (Worst accuracy in square − Accuracyi+ 1)2 (5.1)
Weighted mean = Pn i=1(Weighti∗ RSSi) Pn i=1Weighti (5.2)
The weighting algorithm considers the data points registered at a given square. Where i is a single data point within a square and n the number of data points within a square. For each square, the lowest location accuracy
5.3. CATEGORIZING BY ENERGY CONSUMPTION
is used in order to set the weight of each data point using equation 5.1. The intuition behind this is to make the weight adapt to the accuracy within in the square. The addition of 1 is there so we do not disregard any data point. This will generate a weight which is higher for values with better accuracy than the lowest in that square. The weighted mean RSS per square is the common formula shown in 5.2.
Figure 5.5: weighted average RSS of the small data set.
After applying our weighted mean algorithm to the small subset, we get a representation which heavily favors high location accuracy. In figure 5.5 the result of the algorithm is shown in the signal map, where 0 is unknown.
5.3
Categorizing by Energy Consumption
Based on the analysis in chapter 4, we aim to estimate the expected energy consumption for each square in the signal map using RSS. The categorization is simplistic considering only previously analyzed power of the DCH state. The map categorizes the energy consumption of each square as either "Low", "Medium" or "High".
The total actual energy consumption depends on many aspects such as the particular UE hardware, network configuration parameters (e.g., inactiv-ity timers and state transitions), UE transmissions including data pattern, amount of data or data rate and how stable the RSS is [1, 2, 19, 20]. Since the map does not contain all the information, we do a coarse approximation of the consumption based on the RSS and data rate information.
CHAPTER 5. SIGNAL MAP
By leveraging the created signal map in figure 5.5 and the analysis in sections 4.2-4.3, we can construct an energy categorization, categorizing each square in the map as "High", "Medium", and "Low" for energy consumption.
,
Figure 5.6: Box plot with power readings from all the experiments under analysis in section 4.3.
In order to select the boundaries between the "High", "Medium" and "Low" categories, we first analyze all the power samples in section 4.3 at each RSS. Figure 5.6 shows that the power consumption observed for -50 and -70 dBm are very similar. At -90 dBm the energy consumption is slightly higher and at -110 dBm the energy consumption is much higher.
Next, we need to decide the intervals of "High", "Low", and "Medium". For this we analyze the power consumption resulting from the ping exper-iments, figure 4.1, and look at similarities to the observed power levels for -50, -70, -90, and -110 dBm.
The observed energy consumption is similar from -50 to -80 dBm. At -100 and -110 dBm the energy consumption is much higher than for better RSS. -90 dBm is somewhere in between the other samples. Based on this analysis we choose the interval for "Low" to higher RSS than -85 dBm and lower RSS than -95 dBm as "High". Between "Low" and "High" we have the "Medium" interval.
6.
Evaluation
This chapter evaluates the generated signal map and the location energy categorization. Section 6.1 shows the preliminary RSS prediction evaluation of the signal map. Section 6.2 preliminary evaluates the energy categories by comparing the energy estimation for different energy categories present in the energy categorization.
6.1
RSS Estimation
The signal map provides means to present the current knowledge about the RSS of a location, and thereby providing an estimate RSS together with the corresponding expected energy consumption level. In order to evaluate the signal map, we need to evaluate how good this estimation is.
For this purpose, we collect new data and evaluate the estimation pro-vided by the signal map compared to the new data. We do this following two different methods: First we collect new data for three representative squares of the signal map: one with high amount of observations and high STD, one with low amount of observations and low STD, another with low amount of observations. Secondly, new data is collected from a single user running along a small path. For each square in the signal map traversed by the path we compare the collected new data to the signal map data.
The metric used to compare is Pearson Product-Moment Correlation Coefficient (PPMCC). The two data sets which we compare are all new collected samples in X and their corresponding value in the signal map as
Y . The PPMCC value is in the range of -1 to 1. 1 is strong linear correlation,
0 is no correlation, and -1 is negative correlation. In our data set we never expect a PPMCC of 1 since the RSS is not stable, even if the UE is not moving. If we get good correlation the PPMCC should be below 1 but above 0. P P M CC = P xy − P xPy n 2 q (P x2−( Px)2 n ) ∗ (P y 2−( Py)2 n ) (6.1)
CHAPTER 6. EVALUATION
Method 1: Our hypothesis is that a square with low amount of observa-tions is statistically not relevant and thus it will not provide good results. We also expect that a square with more data points will provide a better result when the STD is low. However, if the STD is high, the rigid form of the grid is probably the limiting factor. Thus we select our three squares to prove these hypothesis.
Square(x,y) Observations STD PPMCC Square(147,54) 1013 5.66 0.4356 Square(143,55) 54 9.83 -0.0484 Square(144,52) 4 6 -0.5774
Table 6.1: Comparison of squares with different STD and observations. The x and y refer to the index of the square in the signal map (see figure E.1).
Table 6.1 shows that squares with low amount of observations have a low correlation with the new data. The square with 1013 observations has 889 observations with a precision higher than 10 m. The high amount of points with high accuracy makes us consider that the PPMCC is relatively good, since it shows that there is some positive correlation between the data sets. Most squares however only contains less than 54 observations, which means that we are likely to not have enough data points for most squares to do a proper estimation.
Figure 6.1: Collected new data along the small path mapped with GPS.
Method 2: The small path goes through 24 squares with observations. Figure 6.1 shows the collected path with corresponding GPS reading. Since the GPS does not always give good readings we have some readings that are outside the path.
The accuracy of the path is evaluated calculating the PPMCC for the aggregated points. We get a PPMCC value of 0.2115. This means that the
6.1. RSS ESTIMATION
correlation between the values is not strong since the value is close to 0. In other words the RSS estimation is not very good for the small path.
Figure 6.2: Observations per square in the signal map along the small path.
Figure 6.3: Observations per square for the new data along the small path.
Figure 6.4: RSS per square in the signal map VS the newly collected data.
We believe that the reason behind a not that good estimation is the lack of data in the signal map for the evaluation path now that we study it in detail. Figure 6.2 shows that some of the squares along the path only have 1 logged observation and the maximum number of observations is 72. PPMCC can only be calculated if we get enough data from both the squares in the signal map and the new observations because we need variance. If there is no variance within either of the compared data sets we get a division with 0 which is undefined.
Figures 6.4 shows that we can only compute the PPMCC for the whole path and for one of the squares along the small path. In other words we would need to collect new data with a higher amount of observations per square in order to calculate the PPMCC for most squares. However it is not realistic behavior to stand still every 10 m along a short path, thus we prefer the approach of comparing the PPMCC of the whole path.
CHAPTER 6. EVALUATION
Results: These two different evaluation methods aim to evaluate two differ-ent cases. The first is to evaluate if the signal map can sufficidiffer-ently estimate at a single location. The second method is for checking the estimation along a path.
To sum up, evaluation shows that in the current collected data there are not enough observations to properly estimate the RSS, but there is small evidence that with more observations can lead to a better estimation. We also believe that the rigid grid is also a limiting factor for the accuracy of the estimation.
6.2
Evaluation of the Energy Categorization
In order to evaluate the energy categorization, we perform data communica-tion in three different squares in the signal map, corresponding to a "High", a "Medium", and a "Low" energy level square. The evaluation criterion is that the energy consumed in the "High" square needs to be higher than the others and "Low" needs to lower.
Methodology: We create traffic while capturing the packet trace and the RSS. We watch a video on the Youtube Android application to generate traffic. We compute the resulting energy from the collected packet traces using EnergyBox. Our power models presented in section 4.4 is added to EnergyBox in order to compute energy consumption for different RSS values. This procedure is done for one square with "High", one with "Medium" and one with "Low" energy level. For collecting packet traces on the UE we use an application called Shark [10]. To block other applications from sending packets we use AFWall+ [21] as a firewall on the UE. For generating traffic the Android application Youtube was used, the watched video was the new Star Wars trailer1with a playback quality of 480p. RSS is collected
with the logging application used in this thesis, Network Monitor.
EnergyBox Extension for RSS and Data Rate: We apply the power mod-els by exporting the data from EnergyBox. The exported data contains vectors with timestamp in one column and the corresponding RRC state of the UE in another. The data rate is given since it can be calculated by analyzing the packet trace. Since the states, RSS and data rate are known we can directly apply the power models for DCH. For the other RRC states we keep the estimation provided by EnergyBox.
Results: We picked three squares for the tests by looking at the energy categorization, one "High", one "Medium" and one with "Low" energy level. Then we perform our test at all locations and estimate the energy consump-tion.
