Linköping University | IDA Bachelor Thesis | Computer and Information Science

Spring 2016| LIU-IDA/LITH-EX-G—16/071--SE

How Different Parameters Affect the Downlink Speed

Martin Claesson Lovisa Edholm

Handledare/Tutor, Niklas Carlsson Examinator, Nahid Shahmehri

Abstract

Today many societies rely on fast mobile networks, and the future seem to place evenlarger demand on the networks performance. This thesis analyzes which parameters af-fects the downlink speed of mobile networks. Various statistical analyses are performedon a large dataset provided by Bredbandskollen. We find that parameters such as the in-ternet service provider, the type of phone, the time of day and the density of populationaffect the downlink speed. We also find that the downlink speeds are significantly higherin urban areas compared to more rural regions.

Acknowledgments

Thanks to Rickard Dahlstrand at .SE for sharing the Bredbandskollen dataset, without thedataset this study would not have been possible. Thanks to our supervisor Niklas Carlssonfor all the great guidance in this project. We would also like to thank our Tim Lestanderand Jakob Nilsson for proof reading our thesis as well as providing us with helpful feedbackduring the process.

iii

Contents

Abstract ii

Acknowledgments iii

Contents iv

List of Figures v

List of Tables vi

1 Introduction 2

1.1 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31.2 Thesis outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2 Related Work 5

3 Method 6

3.1 The dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63.2 Linear regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

4 Results 9

4.1 Urban and rural areas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94.2 Phones and tablets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104.3 Time of day . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134.4 Distance from city center . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174.5 Density of measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

5 Discussion 21

5.1 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225.3 Work in a wider context . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

6 Conclusion 25

6.1 Possible future works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

Bibliography 27

iv

List of Figures

1.1 Grids sized 1600x1600, 800x800, 400x400 and 200x200 meters. . . . . . . . . . . . . 31.2 Doughnut-shaped rings with a width of 100 meters and a radius between 100 and

2500 meters. Placed with the city center in the center of the circles . . . . . . . . . . 4

4.1 Average downlink speed for different ISPs in the data set. . . . . . . . . . . . . . . . 94.2 Average downlink speed for different ISPs in Stockholm, Göteborg, Malmö, Upp-

sala and Linköping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104.3 Average downlink speed for different mobile units for Telia . . . . . . . . . . . . . . 114.4 Average downlink speed for different mobile units for Tele2 . . . . . . . . . . . . . 114.5 Average downlink speed for different mobile units for Telenor . . . . . . . . . . . . 114.6 Total number of measurements per time of day . . . . . . . . . . . . . . . . . . . . . 134.7 Average downlink speed per time of day for Telia . . . . . . . . . . . . . . . . . . . 144.8 Average downlink speed per time of day for Tele2 . . . . . . . . . . . . . . . . . . . 144.9 Average downlink speed per time of day for Telenor . . . . . . . . . . . . . . . . . . 144.10 Linear regression for Telia, with average downlink speed on the y-axis and number

of measurements (interval for each hour of the day) on the x-axis.The equation forthe linear regression is y = 8 ¨ 10´5x + 14.394 and the value for the coefficient ofdetermination(R2) is 0.20671. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

4.11 Linear regression for Tele2, with average downlink speed on the y-axis and num-ber of measurements (interval for each hour of the day) on the x-axis. The equationfor the linear regression is y = ´0.0005x + 25.089 and the value for the coefficientof determination(R2) is 0.86754 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

4.12 Linear regression for Telenor, with average downlink speed in the y-axis and num-ber of measurements (interval for each hour of the day) in the x-axis. The equationfor the linear regression is y = ´0.0004x + 21.475 and the value for the coefficientof determination(R2) is 0.89060 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

4.13 Grids sized 1600x1600, 800x800, 400x400 and 200x200 meters. . . . . . . . . . . . . 194.14 Linear regression with average downlink on the y-axis and number of measure-

ments per region in each square on the x-axis. . . . . . . . . . . . . . . . . . . . . . . 20

v

List of Tables

4.1 Difference in performance between devices. . . . . . . . . . . . . . . . . . . . . . . . 124.2 Linear regression in grids of size 1600x1600, with average downlink speed (Mbps)

in the y-axis and distance to city center (km) in the x-axis . . . . . . . . . . . . . . . 174.3 Linear regression of 25 circles of varying distance from city center, with average

downlink speed (Mbps) in the y-axis and distance to city center (km) in the x-axis 184.4 Linear regression with average downlink speed (Mbps) on the y-axis and density

of measurements per region on the x-axis. . . . . . . . . . . . . . . . . . . . . . . . . 20

vi

List of Tables

Students in the 5 year Information Technology program complete a semester-long soft-ware development project during their sixth semester (third year). The project is completedin mid-sized groups, and the students implement a mobile application intended to be usedin a multi-actor setting, currently a search and rescue scenario. In parallel they study severaltopics relevant to the technical and ethical considerations in the project. The project culmi-nates by demonstrating a working product and a written report documenting the results ofthe practical development process including requirements elicitation. During the final stageof the semester, students form small groups and specialise in one topic, resulting in a bache-lor 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 thesemester, including the conditions and requirements for a bachelor thesis.

1

1 Introduction

In a disaster scenario, the need for fast and reliable network is critical. Today, almost every-one owns a battery powered device such as a laptop, a smartphone or a tablet. Since thetransfer of data is very energy consuming, it is important to reduce the time spent download-ing data in order to maximize the device’s uptime. This is especially important in a disasterscenario when the access to charging points may be very limited.

Smartphone usage is very diverse, and a study made by Falaki et al. [4] has shown thatthe average interactions made by users per day vary from from 10-200 interactions. The samestudy shows that the amount of data received by each user varies between 1-1000 MB.

Network performance varies depending on various conditions. To understand the cur-rent network conditions, models such as performance maps are valuable. In this thesis, weanalyze network capabilities by using crowd-sourced network measurements and networkperformance maps summarizing these measurements. We use this data to create a multi-variate model for mobile download speeds.

This is done by using a large crowd sourced dataset provided by Bredbandskollen, Swe-dens primary internet test provider. By January 2016 it had been used to perform over187 million measurements, and today about 100,000 measurements are performed usingBredbandskollen every day. This thesis focuses on the 16 million mobile non-WiFi measure-ments done between January 2014 and February 2015, using simultaneously collected metainformation such as operator and geographic location we identify factors that impact thedownlink speed. The locations in Sweden have been split up into different regions. In orderto be able to use the techniques and softwares required for the analysis, we use aggregatemeasurements as well as down sample the dataset.

2

1.1. Contributions

1.1 Contributions

This thesis makes two primary contributions. The first one is to identify factors of poten-tial interest that may impact internet speed. For this purpose, we characterize the mobilespeed test usage of Bredbandskollen by creating a performance map that fits our needs.The measurements are diurnal (with a peak-to-valley ratio of 16) and highly concentratedto regions where most people live, with a small amount of the geographical locations beingresponsible for most measurements. To allow for efficient analysis we therefore focus ouranalysis towards the more frequent locations.

The second contribution is to look at how well each of the candidate factors satisfies theassumptions, and if needed, transform variables. This is done by studying each factor indifferent contexts, and determine whether they satisfy the assumptions or not.

Cluster Analysis

To cluster the data we use three methods. First, similar to prior works analyzing this dataset[7], we divide the data into square buckets with sizes between 200x200 and 1600x1600 meters,as seen in Figure 1.1. Second, when considering geographic locality of the measurementsproximity to a city center we analyze the data within doughnut-shaped rings with a width of100 meters and a radius between 100 and 2500 meters (Figure 1.2). The third and last methodis to divide the dataset within the 1600x1600 bucket (the largest bucket used in Figure 1.1)into nine equally sized squares. Since the dataset is sparse, we aggregate many measurementsfrom a geographical area with similar characteristics, and perform analysis on this aggregateddata [11]. For the purpose of our discussion, we call all measurements from such a location agroup of clustered measurements.

Figure 1.1: Grids sized 1600x1600, 800x800, 400x400 and 200x200 meters.

3

1.2. Thesis outline

Figure 1.2: Doughnut-shaped rings with a width of 100 meters and a radius between 100 and2500 meters. Placed with the city center in the center of the circles

1.2 Thesis outline

This thesis is structured as follows. Section 2 presents related works and the theoretical back-ground for the thesis. In Section 3 the methodology used to find the results is presented,together with some explanations of the dataset that is used for the study. Section 4 presentsthe results from the study. Section 5 contains a discussion of the methodologies and the re-sults from the previous chapter. The last section presents a conclusion and suggestions forfurther work.

4

2 Related Work

In a disaster scenario, the need for a reliable network for communication is essential. Byanalyzing previous measurements of download speed in different locations, it is possible tomake predictions that make it possible to improve Quality of Servise (QoS) [13]. Such predic-tions are easier to make within smaller areas, since those areas have a lower covariance [7].Areas with a larger downlink speeds tend to have neighbouring areas with high downlinkspeed, which makes predictions easier in areas with a high downlink speed. Due to this fact,a user can travel through an urban area and experience similar download speeds.

Network prediction maps is useful when trying to increase the download speed [3] whichfacilitates the communication over the network. Yao et al. [13] shows that historical datafrom a given place can provide a good picture of the speed you can expect the same place.

In this thesis we investigate the factors that affect the download speed and in what waythey affect it. The density of the base stations used will increase the success transmissiondensity, but the increasing rates will diminish according to a study by Yu and Kim in 2013[14]. This means that the number of base stations should be more than n-times to increasethe network capacity by a factor of n.

According to Jang et al. [6] the QoS is lower on a device that moves at high speed thanin one that does not. In that paper they studied network performances in a fast moving caron a highway and in a high-speed train running at 300 km/h, and found that mobile nodesexperience far worse performance than stationary nodes over the same network. One sourceof error in our analysis is that our data set does not contain information about whether thedevices are moving or not.

The network technologies used will also affect the QoS [6][7]. The different wireless linkcharacteristics of 3G and 3.5G will make the performance vary greatly between the differentnetwork types, although they all support enough bandwidth for typical Internet applications.

5

3 Method

This chapter focuses on the method used to analyze the downlink speed, and how differentparameters affects it. It begins with an explanation of the dataset used, and continues withan in depth explanation on linear regression and the coefficient of determination.

3.1 The dataset

This study is, as previously mentioned, based on a dataset provided by Bredbandskollen.Bredbandskollen is a swedish speed test service that is provided by The Internet Foundationin Sweden (IIS), an independent organization that promotes the positive development ofthe Internet in Sweden. The dataset is crowd sourced by mobile users testing their Internetspeed on their tablet or phone using Bredbandskollen’s IOS or Android application. The ap-plication provides information about the uplink and downlink speed, as well as informationregarding geographic location, latency and timestamps. The test are carried out against thegeographically closest Internet exchange point. The dataset contains data from 2008-2015,but since the speed has increased so much between these years, we will narrow the datasetdown to only the years 2014-2015 to avoid misleading result that may derive from the greatvariations in downlink and uplink speed. The Internet Service Providers (ISP) that are ana-lyzed are: Telia, Tele2 and Telenor. The different network technologies used in dataset are:

• EDGE (Enhanced Data rates for GSM Evolution, 2.5G)

• GPRS (General Packet Radio Services, 2.5G)

• CDMA (Code Division Multiple Access, 3G)

• UMTS (Universal Mobile Telecommunications System, 3G)

• HSPA (High-Speed Packet Access, 3.5G)

• HSPDA (High-Speed Downlink Packet Access, Turbo 3G)

• HSPAP (High-Speed Packet Access Plus, 3G)

• LTE (Long Term Evolution, 4G)

6

3.2. Linear regression

3.2 Linear regression

For most of the analysis made we have used linear regression [8]. Using linear regression wedevelop a function that creates an approximation-based model of the relationship betweenvariables observed in the data from the Bredbandskollen. In statistics, linear regression istypically used to determine if there is a statistical link between a response variable and oneor more explanatory variables. If more than two variables are used it is commonly referredto as multiple linear regression.

The probability distribution between x at every value for y is given by

y = b0 + b1x + e, (3.1)

where b0 is the intercept, b1 is the slope and e is an uncorrelated error component.

The variance is given by the equation

Var(x|y) = Var(b0 + b1x + e) = s2. (3.2)

By comparing a large number of measurements from Bredbandskollen we will approximatethe impact of different variables on the bandwidth.

Regression models are primarily used for four different different purposes, inluding

• data description,

• parameter estimation,

• prediction and estimation and

• control.

Regression analysis is a helpful tool when developing equations to summarize or describe atype of data. Especially when you are working with a large dataset, like we are, a regressionmodel may provide a more useful and convenient summary than a table or a graph[12].A parameter estimation problem may also be solved by using a regression model. In sucha case, regression analysis can be used to fit the data to the model and thus providing anestimation of a variable.

Regression models are also useful for predicting the response variable. For example, ifwe have a lot of data concerning network traffic, the network speed at a specific location canbe predicted by using a regression model. In this case, it is important that the estimationof the model parameters are good since a poor estimation may result in a poor prediction.Errors in the model or the equation may also result in a poor prediction.

The fourth common area of usage for regression models are for control purposes. Whena regression equation is used for this purpose, it is important that the variables are relatedin a casual manner. This cause-and-effect relationship is not needed for prediction, since itin this case is only necessary that the relationships that existed in the original data used forbuilding the model are still valid.

Creating a regression model is an iterative process. An initial regression model is speci-fied by using any knowledge about the data from Bredbandskollen. In the next step theparameters of the model is estimated and evaluated as well as the model adequacy. This isrepeated over again until an appropriate model is obtained.

7

3.2. Linear regression

Coefficient of determination

The coefficient of determination (R2), sometimes referred to as multiple correlation coeffi-cient, is well established in classical statistical analysis. It is defined as the proportion ofvariance explained by the regression analysis, which makes it useful as a measure of successof predicting the dependent variable from the independent variable [9]. The coefficient ofdetermination is defined as follows:

R2 =SSRSST

= 1 ´ SSResSST

. (3.3)

In the equation above, SST is a measure of the variability of y without considering theeffect of the regressor variable x and SSRes is a measure of the variability in y remaining afterx has been considered. Therefore, R2 is often called the proportion of variance explained bythe regressor x.

The value of R2 varies between 0 and 1, and can be seen as a percentage of how much ofthe variability that is accounted for in the model; e.g., if the value of R2 is 0.43, that meansthat 43 percent of the variability in strength is accounted for [8].

8

4 Results

In this chapter we look at how different parameters affect the downlink speed and in whatway they affect it. Looking at the first measurements of the data set we discovered clearpatterns in the downlink speed.

4.1 Urban and rural areas

It turned out that the downlink speed was significantly higher in urban areas compared tothe rest of Sweden. This result suggests that factors that are different in urban locations incomparison to more rural areas affect the speed in some significant way for all of the ISPs. Toshow this we include results for both the entire dataset and for the data collected in five ofthe biggest cities in Sweden: Stockholm, Gothenburg, Malmö, Uppsala and Linköping.

Figure 4.1: Average downlink speed for different ISPs in the data set.

9

4.2. Phones and tablets

Figure 4.1 shows the average downlink speed for the ISPs in all of Sweden. Confidenceintervals with a confidence level of 95 percent have been added to the graph. However, sincethe dataset is so large, the confidence intervals are so small that they are barely visible. Forthis reason it was decided not to include confidence intervals in the rest of the graphs in thisthesis. Note that the average downlink speed for Telia is 18.49 Mbps, the average down linkspeed for Tele2 is 18.15 Mbps, and the average downlink speed for Telenor is 15.47 Mbps. Teliaand Tele2 shows a significantly higher downlink speed than Telenor.

Figure 4.2 shows the difference between the average downlink speed for the differentISPs in urban parts of Sweden.

Figure 4.2: Average downlink speed for different ISPs in Stockholm, Göteborg, Malmö, Upp-sala and Linköping

From 4.2 we see that the behavior in the urban areas is different from the behaviour inSweden overall. In this we see that Tele2 has significantly higher downlink speed than theother ISPs. Tele2 has a average downlink speed of 26.36 Mbps, Telia has a average downlinkspeed of 22.21 Mbps and Telenor has a average downlink speed of 22.19 Mbps.

These two charts indicates that the ISPs Telenor and Telia behave in a similar way on char-acteristics typical for urban areas, while Tele2 and Telia behaves in a similar way in Swedenoverall.

4.2 Phones and tablets

The network speed may vary between the types of phones that are used. For example anewer phone may be able to use newer types of technologies and the network interface cardmay be more advanced. Due to this, we decided to try to see if any significant differencescould be seen between different types of phone. Of course, in a dataset as large as the onefrom Bredbandskollen many different kinds of phone units will be used and some of themare used in just a few measurements. Therefore we decided to exclude the ones with less than500 measurements. The results are separated between the three major ISPs: Telia, Tele2 andTelenor. This is done to avoid differences caused by characteristics in the ISPs being mistakenfor differences caused by the devices.

10

4.2. Phones and tablets

Figure 4.3: Average downlink speed for different mobile units for Telia

Figure 4.4: Average downlink speed for different mobile units for Tele2

Figure 4.5: Average downlink speed for different mobile units for Telenor

As seen in Figures 4.3, 4.4 and 4.5, the tablets from Xperia and the Samsung SM phoneshas a high average downlink speed for all of the ISPs. Iphone, Ipad and HTC has among thelowest average downlink speeds in all of three cases. The results vary significantly between

11

4.2. Phones and tablets

the different types of phones and tablets, and it could be said that the choice of phone ortablet may affect the downlink speed. One limitation of this analysis is that geographiclocations are not taken into account, e.g., a new phone such as OnePlus may have a higheraverage downlink speed if a larger portion of the measurements are made in a urban area. Ingeneral, the type of phone have similar values for each of the three ISPs. The one exceptionis the Ipad that for some reason has a low average downlink speed for Tele2. The differencebetween the two tablets (Ipad and Xperia) are significant and bigger than we expected.

To get a overall picture of the average performance of the devices were ranked within everyISP, where the best performing device was ranked as number 1 and the worst performingdevice was ranked as number 11. Table 4.1 shows the devices with the highest rank first,as evaluated across the three most popular ISPs. We also took the average speed for everydevice by dividing the sum of the average speed within the three ISPs in three.

RankDevice Average Best Worst Average Downlink Quantity

XPERIA Tablets 2 1 3 24.30 Mbps 4,635OnePlus 2.6 1 4 22.42 Mbps 8,945Samsung SM 3 2 4 22.74 Mbps 201,273Nexus 4 2 5 21.34 Mbps 25,986Huawei 5 1 7 21.57 Mbps 244,348XPERIA 5.3 4 7 20.57 Mbps 189,366LG 7.3 6 10 17.50 Mbps 17,241HTC 8.3 6 11 16.48 Mbps 88,779Samsung 9 8 10 16.67 Mbps 134,968Iphone 9.3 9 10 15.36 Mbps 572,178Ipad 9.3 8 11 13.16 Mbps 76,320

Table 4.1: Difference in performance between devices.

In Table 4.1 we see that XPERIA tablets has the best ranking and also has significantlyhigher average downlink speed than the other devices. Looking at for example OnePlus,who has a better rank than Samsung SM, we see that Samsung SM has a higher averagedownlink speed than OnePlus. This also applies to Nexus in comparison to Huawei andHTC in comparison to Samsung. Overall we see that the performance patterns within thedifferent ISPs is very similar, looking at the different devices.

12

4.3. Time of day

4.3 Time of day

The time of day is another parameter that may affect the downlink speed, mainly due to thefact that the number of simultaneous users varies a great deal during the day. This is shownin Figure 4.6. Since the graph looks almost identical for all of the different ISPs, the totalnumber of measurements have been added into just one graph.

Figure 4.6: Total number of measurements per time of day

The affect this has on the downlink speed can be seen in Figures 4.7, 4.8 and 4.9, wherewe show the downlink speed on the y-axis and time of day (divided into intervals of onehour) on the x-axis. We note that Tele2 and Telenor looks fairly similar, and that Telia looksquite different from the other two. Telia has the highest average downlink speeds at the bus-iest times of the day, which could suggest that they adept their network resources on somesort of on demand schedule. Looking at the difference between 23:00-00:00 and 00:00-01:00as well as the difference 10:00-11:00 and 11:00-12:00 the significant changes between this timespans suggest that some sort of adaption is done by Telia. Looking at Telenor and Tele2 we seethat they follow a more expected pattern, where the average downlink is at its peak whenthere are less people using the network resources. We found this difference between the ISPsto be really surprising. The datasets have been checked for abnormalities in order to see ifthat could explain the big difference we see from Telia, but no such abnormalities could befound.

13

4.3. Time of day

Figure 4.7: Average downlink speed per time of day for Telia

Figure 4.8: Average downlink speed per time of day for Tele2

Figure 4.9: Average downlink speed per time of day for Telenor

14

4.3. Time of day

Linear regression

To further investigate how the difference in user activity seen over the day (e.g. as in figure4.7) affects the downlink speed, we performed a linear regression between the number ofmeasurements and average downlink speed for each ISP. Figures 4.10, 4.11 and 4.12 presentthese results. In these graphs the number of measurements for each interval (with one hourtime granularity) is used on the x-axis, and the y-axis still contains the average downlinkspeed. These graphs again highlights the difference in characteristics between the ISPs,where the time of day affects the speed similarly for Tele2 and Telenor, which is shown by thesimilarity in the linear equations. Telia still differs a lot from the other two, and again thetime of day seems to have almost the opposite effect on the downlink speed compared toTele2 and Telenor. Before we conducted these investigations, we expected that the downlinkspeed would decrease when the number of measurements increased. This expected patterncan be found in the graphs for Tele2 and Telenor, but the results from Telia differs significantlyfrom our inital assumption.

Another thing that is worth noting is that the coefficient of determination is much higher forTele2 and Telenor, which means that a larger portion of the variance is included in the model.The value of R2 is 0.87 for Tele2, 0.89 for Telenor and 0.21 for Telia. The variance can be seenin the graph as well where the dots are spread out quite far away from the trendline in thegraph for Telia, and follows the trendline fairly well in the graphs for Telenor and Tele2.

Figure 4.10: Linear regression for Telia, with average downlink speed on the y-axis and num-ber of measurements (interval for each hour of the day) on the x-axis.The equation for the lin-ear regression is y = 8 ¨ 10´5x + 14.394 and the value for the coefficient of determination(R2)is 0.20671.

15

4.3. Time of day

Figure 4.11: Linear regression for Tele2, with average downlink speed on the y-axis and num-ber of measurements (interval for each hour of the day) on the x-axis. The equation for the lin-ear regression is y = ´0.0005x + 25.089 and the value for the coefficient of determination(R2)is 0.86754

Figure 4.12: Linear regression for Telenor, with average downlink speed in the y-axis andnumber of measurements (interval for each hour of the day) in the x-axis. The equa-tion for the linear regression is y = ´0.0004x + 21.475 and the value for the coefficient ofdetermination(R2) is 0.89060

16

4.4. Distance from city center

4.4 Distance from city center

Another parameter that was taken into account was how close the measurement was to thecity center. This study focused on five of the biggest cities in Sweden: Stockholm, Gothen-burg, Malmö, Linköping and Uppsala. In each of the five cities a central point was deter-mined as follows: Central station in Stockholm, Götaplatsen in Gothenburg, Folkets park inMalmö, Stora torget in Linköping and Stora torget in Uppsala.

Grid-based analysis

After deciding the central positions, the dataset was partitioned into a grid with the size1600x1600 meters for each of the cities. From each single measurement within the grid thedistance to the central point was calculated using Pythagorean theorem. On these 15 grids(5 in each ISP) a linear regression was made, resulting in the following equations (4.2) andcoefficients of determination.

Linear regression, Average downlink and distance to city center.City ISP Linear equation R

2

Stockholm Telia -0.035x + 21.736 1ˆ10´4

Stockholm Tele2 0.0033x + 22.183 0.001Stockholm Telenor -0.363x + 25.563 0.006Gothenburg Telia 0.001ln(x)+29.609 0.002Gothenburg Tele2 0.282x + 23.910 0.003Gothenburg Telenor 0.507x + 16.791 0.009Malmö Telia 0.022x + 22.188 5ˆ10´5

Malmö Tele2 0.865x + 16.781 0.026Malmö Telenor 0.607x + 15.153 0.013Linköping Telia 1.435x + 28.847 0.037Linköping Tele2 2.028x + 14.442 0.055Linköping Telenor -0.858x + 20.548 2ˆ10´6

Uppsala Telia 1.429x + 12.459 0.029Uppsala Tele2 -0.515x + 19.646 0.013Uppsala Telenor -0.544x + 28.858 0.005

Table 4.2: Linear regression in grids of size 1600x1600, with average downlink speed (Mbps)in the y-axis and distance to city center (km) in the x-axis

What is worth noting is that 10 of the 15 linear regressions have a positive slope, whichsuggests that the speed declines as it gets closer to the city center. Another thing that isnoteworthy is the relatively low numbers for the coefficients of determination (R2), whichmeans that a lot of the data can not be fitted into the linear model which makes the results abit unreliable. In comparison with the linear regressions made earlier in the report these 15regressions use a lot more data, so a lower value for R2 is to be expected.

17

4.4. Distance from city center

Circular experiment

Since the results from the first experiment regarding distance to city center had a low valuefor the coefficient of determination, it indicated that the results could be a bit unreliable.Because of this we decided to also divide the measurements into 25 circles, with a radiusvarying from 100-2,500 meters and where the smaller circles were not included in the largerones. Since the results from Tele2 and Telenor had been so similar in the previous analyseswe decided to combine the measurements from these two ISPs in this analysis. The results ofthese measurements can be found in Table 4.3.

Linear regression, Average downlink and distance to city center.City ISP Linear equation R

2

Stockholm Telenor and Tele2 0.197x + 22.247 0.099Stockholm Telia 0.1473x + 20.274 0.096Gothenburg Telenor and Tele2 0.0184x + 23.718 0.001Gothenburg Telia -0.079x + 25.915 0.01Malmö Telenor and Tele2 0.4658x + 17.152 0.152Malmö Telia 0.3692x + 17.321 0.208Linköping Telenor and Tele2 0.0025x + 20.498 9.2ˆ10´6

Linköping Telia -0.147x + 22.588 0.035Uppsala Telenor and Tele2 -0.2011x + 23.776 0.059Uppsala Telia 0.2025x + 18.479 0.065

Table 4.3: Linear regression of 25 circles of varying distance from city center, with averagedownlink speed (Mbps) in the y-axis and distance to city center (km) in the x-axis

These regressions show a slightly higher value for the coefficient of determination, whichindicates a more reliable result than the linear regression in Table 4.3. What is worth noting isthat the linear equation has a positive coefficient in seven of the ten linear regression, whichmeans that slope is positive for these regression. With that in mind, it seems to indicate onceagain that the downlink speed increases as you move away from the city center. This seemsto be more true in the bigger cities, i.e, Stockholm, Gothenburg and Malmö since five of thesix linear regression has a positive slope in these cities.

18

4.5. Density of measurements

4.5 Density of measurements

Another parameter that we identify that could potentially affect the downlink speed wasthe number of measurements inside a specified area. For example, does regions with moremeasurements (and potentially more users) see higher or lower downlink speeds within acity? To answer this question, we partitioned the central 1600x1600 areas in each city in twodifferent ways. The first way we divided the areas was by making four squares inside the1600x1600 square. The other way in which we divided the data was by making a 3x3 grid netof each of the 1600x1600 squares.

Grids

Next, we seperated the 1600x1600 grids into three more grids of sizes 800x800, 400x400 and200x200 resulting in four grids in total for each of the ISPs and cities (making it a total of 60grids). Figure 4.13 shows how the different sized grids are placed with the city center in thecenter of every grid.

Figure 4.13: Grids sized 1600x1600, 800x800, 400x400 and 200x200 meters.

19

4.5. Density of measurements

For each one of the grids, a density was calculated as the number of measurements in thatparticular grid divided by the area of that region. In order to see how the density affectedthe downlink speed, we made a linear regression with average downlink speed on the y-axisand the density on the x-axis. This resulted in the following linear equations and coefficientsof determination, shown in Table 4.4

Linear regression, Average downlink and density of measurementsISP Linear equation R

2

Telia 0.592x + 21.264 0.021Tele2 -0.124x + 23.555 0.041Telenor -0.192x + 23.146 0.238Total -0.104x + 22.986 0.059

Table 4.4: Linear regression with average downlink speed (Mbps) on the y-axis and densityof measurements per region on the x-axis.

Squares

To further investigate how the density of the measurements affected the downlink, the1600x1600 regions were seperated into 9 squares of equal size. Another linear regression wasmade with average downlink speed on the y-axis and the number of measurements in eachgrid on the x-axis. Figure 4.14 shows these results.

Figure 4.14: Linear regression with average downlink on the y-axis and number of measure-ments per region in each square on the x-axis.

20

5 Discussion

In this chapter we discuss the results from the previous chapter, as well as the limitations andadvantages of the method we have used to get our results.

5.1 Method

Our method focuses highly on linear regression with one variable. The upside of this ap-proach is that it is easy to see how particular parameters affects the downlink speed. Thedownside is that, compared to a multiple linear regressional model, we do not get as muchinformation about how much the different parameters affect the downlink in comparisonto one another. So to further explore our research questions, we could have focused moreon performing various types of multiple linear regression. We did do a few multi linearregressions, but the information provided by them did not add any new information andthey were done on smaller dataset which made them less reliable. This is the reason that theywere not included in this paper.

Another problem with a multiple linear regression on our dataset is the problem of translat-ing parameters such as time and type of phone into a numeric value. This could be done byweighting a certain type of phone with the average downlink speed and the hourly intervalswith the total number of measurements for that interval. Such trials were made, but wefound the results to be inconclusive, and because of this they were not included in this thesis.Another way in which our results could have been improved would have been by using"category regressor" as in the methodology used by Borghol et al. [1]. That methodologyprovides an extended multi linear regression which gives the following equation,

Yi = b0 +Pÿ

p=1Xi,pbp +

Kÿ

k=1Zi,k + gk + gi, (5.1)

where K is the number of categories, P is the number of predictors and Zi,k is the categoryregressor, encoded as Zi,k = 1 if category i is from category set k, and 0 otherwise. In such acase Zi,k = 1, could be used for parameters such as phone or time of day.

21

5.2. Results

5.2 Results

An analysis that we would have liked to make is one that was based on the measurementsproximity to the base stations. In order to do that we would have needed a map of theirlocations. 10 years ago, such a map was provided by the government agency Post- ochTelestyrelsen (PTS), but it has been taken down from their website. We tried contacting PTS,as well as the different ISPs but our efforts where unfortunatly fruitless. This made such anexamination impossible.

The first thing we noticed from the result was the big difference in downlink speed be-tween rural and urban areas, where the urban areas had a significantly higher averagedownlink speed. The parameters we looked at was:

• Operator

• Type of phone

• Time of day, and the load on each hour

• Population density in 3x3 grids for each city

• Population density in buckets of sizes 1600x1600, 800x800, 400x400 and 200x200

• Distance from central postion of cities

The average downlink speed of the operators differentiated somewhat from one anotherboth in the rural and the urban areas. Tele2 was the ISP that had the highest average downlinkspeed in the urban areas, and Telia had the highest average downlink speed in rural areas(with Tele2 as a close second). The difference between the ISPs is a indication of that theyhandle the downlink speed differently. This might be a result of them reacting different onvarious parameters. In this case it would have been interesting to look at their cell towers.But, as noted above, PTS in Sweden no longer want to share this information for securityreasons.

The next parameter we looked at was how different types of phones affected the speed.Our graphs shows that your phone type most likely will affect your downlink speed, withmodern phones such as OnePlus and the Samsung SM models having higher downlinkspeed in average. Owners of HTC, Ipads and Iphone should be more dissapointed with ourresults. The impact the type of phone has on the downlink speed can be explained by thequality of the Network Inteface Cards (NIC) and what network technologies they are ableto use. E.g, some of the older phones may not be able to use 4G or the other newer networktechnologies which will have negative impact on their downlink speed. The patterns forthe different ISPs are very similar, which indicates that the hierarchy between the phonesare more or less the same, independent of the ISP. We see a difference between the averagedownlink speeds for the same device in different networks. That difference is proportionalwith the average downlink speed differences between the networks. That makes it possibleto compile a common table describing the hierarchy in all ISPs.

Another parameter we identified with potential to affect the downlink speed was howmany simultaneous users there was. The first way we examined this was by looking at howmany users there where at different intervals of time. The number of users varied greatlybetween different intervals, with just 6270 in total measurements between 04:00-05:00 and

22

5.2. Results

107958 measurements between 20:00-21:00. With these densities in mind, we made a linearregression for each of the ISPs. The results from this analysis were very similar for Tele2 andTelenor, but differentiated greatly with the results from Telia. Tele2 and Telenor followed apattern where more measurements decreased the average downlink speed linearly, but forTelia the average downlink speed increased for the more densly hourly intervals. One waythis could be explained is that Tele2 and Telenor share their network infrastructure for 2Gand 4G thru a joint venture company called Net4Mobility 1. But the results from Telia areindeed peculiar. It makes sense that a high number of simultaneous users would decrease thedownlink speed, as it does for Tele2 and Telenor, but with Telia we see a completely differentpattern. One explanation for this could be that Telia adjust their capacities for different timesof the day, but we have found no evidence to support this theory so it is pure speculationfrom our side. Another theory about why Telia preforms better than the other when themost traffic arises, is that Telias customers are more spread over the country. That probablymeans that the users do not need to share the tower cells with as many as the users whoare gathered in urban areas. That, on the other hand, does not explain why Telia preformsworse during times when the number of users are low. The fact that Telia has a pattern that isopposite to Telenor and Tele2 might also indicate that time is a parameter that does not affectthe downlink speed itself.

Another way in which we analysed the number of simultaneous user was by looking atthe data from the five cities (Stockholm, Gothenburg, Malmö, Uppsala and Linköping). Westarted by picking a central position in each of the cities (Central station in Stockholm, Gö-taplatsen in Gothenburg, Folkets park in Malmö, Stora torget in Uppsala and Stora torget inLinköping) and using the longitude and latitude for that position. We then excluded all thedata that was further away than 1600 meters from that location, based on the longitude andlatitude of the measurements. The datasets were then divided into 3x3 grids. We also createdbuckets of sizes 1600x1600, 800x800, 400x400 and 200x200 meters. In all of the boxes of thegrids and each of the buckets we calculated the number of measurements and the averagedownlink speed. A linear regression with average downlink in the y-axis and distance thecentral postion in the x-axis was then made for each of the cities and the ISPs, making ita total of fifteen regressions. For ten of the fifteen linear regression the trendline followedthe pattern we expected, i.e a lower downlink speed in average for the measurements madeclose to the central position. But five of the regression had a pattern where the measurementscloser to the central position had a higher average downlink speed. This was especially truefor Telenor with such a pattern in three out of its five linear regressions. Telenor and Telia bothhad the expected pattern in four out of their five linear regressions. The expected pattenwas found in all of the regressions from Gothenburg and Malmö and in all but one fromLinköping. We can not find a good reason for this, but to further investigate this a multilinearregression could be made with both distance to the central position and the area density inorder to see a more clear connection.

A linear regression was also made for each of the operators with average downlink speedin the y-axis and the number of measurements divided by area in the x-axis. Here again,Tele2 and Telenor follows the pattern we expected and Telia has a pattern in contrast to ourexpectations. A linear regression was also made with the data from all of the ISPs, and thisregression followed the pattern we expected. What is a bit peculiar with the regression madewith the data from all of the ISPs is the significantly higher value of R2. This suggest thatmore of the data is included in the linear model, thus making it the linear regression withthe most accurate result. With this in mind, we can conclude that relation between averagedownlink speed and density of measurements in the different areas follows the pattern weexpected.

1http://www.net4mobility.com

23

5.3. Work in a wider context

Density of users per time and density of users nearby in smaller areas can be parame-ters that should be considered in a combined way instead of separately. The distance to thecenter of the city might, in itself, not be a parameter that affects the downlink speed at all.But it is an indicator that often sums the two previously mentioned parameters, density intime and density in geographical area. Users in the absolute center of a urban city is probablyat a place with many other users. And when they are at the same area at the same time, theyprobably have the same pattern of behaviour. In that way they also use their devices in thesame time. That might be why the downlink speed gets lower in cities the closer you get tothe center.

5.3 Work in a wider context

Data sets such as that used here may contain sensitive information that might cause harmif used wrongfully or if they end up in the wrong hands [10]. Therefore we have signed anagreement with Linköping University saying that we agree to keep such information secretand secure. Sensitive information in such data sets could potentially be information A itself,or the information A in combination with some other information B, for example. We choosenot to write about all the information that the data set contains, in order to keep the integrity,confidentiality and secrecy as high as possible [2]. We also only use aggregate values. Everydecision made during this work has been made with various ethical aspects in mind. Duringthis work we have followed the engineering codes of ethics [5].

24

6 Conclusion

We have showed that the dataset from Bredbandskollen can be used to draw some conclu-sions concerning how some parameters affect the downlink speed. The problem with thedataset is, even though it is a large collection of data, that once you start to sort it into smallerpieces the amount of measurements becomes to small to draw accurate and statisticallybacked conclusions. For example, one of the grids in Linköping for Telenor contained onlytwo measurements. This is why an even larger dataset would have been needed to be evenmore sure of the conclusions we have made.

So what are the conclusions? Well, we have found that the ISP, the time of day, the typeof phone and the proximity to a city will affect the downlink speed in general. So all in all,we have found somewhat satisfying answers to our research questions. What we do lackin this analysis is the relation between the different factors, i.e, how much they affect thedownlink speed in comparison to one another. Some of our parameter has a close connectionto each other, e.g, the time of day and population density both relate to how many simul-taneous users there are. The two parameters that stands on their own are the operator andthe type of mobile unit that is used. This means that these two parameters would have to betreated differently than the other parameters in a multi variate model in order to achieve agood result.

In conclusion, if you want to make sure that you have a high downlink speed on yourmobile unit, use one of the Samsung SM phones, place yourself close to Götaplatsen inGothenburg, use Tele2 as your ISP and do this sometime between 04:00-05:00.

6.1 Possible future works

One area of further investigations is one that we have already discussed, i.e, how the differentparameters relates to one another. One way to do this would be to find a way to associatenumeric values to the parameters time of day and type of phone. Another thing that wouldimprove this analysis would be if one could get a hold of a map of the locations of the towercells used by the ISP. Such a map has existed in the past, but is unfortunately no longer public.Judging by the responses we got from PTS and the ISPs, such a map will be quite difficult to

25

6.1. Possible future works

get a hold of, since the ISPs do not seem willing to share that information for security reasons.Hopefully this will change in the future.

26

Bibliography

[1] Youmna Borghol, Sebastien Ardon, Niklas Carlsson, Derek Eager, and Anirban Ma-hanti. “The Untold Story of the Clones: Content-agnostic Factors that Impact YouTubeVideo Popularity”. In: Proceedings of ACM SIGKDD International Conference on KnowledgeDiscovery and Data Mining. (Beijing, China). ACM, 2012, pp. 1186–1194.

[2] Folker den Braber, Ida Hogganvik, Mass Soldal Lund, Ketil Stølen, and FredrikVraalsen. “Model-based security analysis in seven steps — a guided tour to the CORASmethod”. In: BT Technology Journal (2007).

[3] Alberto Garcia Estevez and Niklas Carlsson. “Geo-location-aware emulations for per-formance evaluation of mobile applications”. In: Proceeding of the Annual Conference onWireless On-demand Network Systems and Services. 2014, pp. 73–76.

[4] H. Falaki, R. Mahajan, S. Kandula, D. Lymberopoulos, R. Govindan, and D. Estrin. “Di-versity in smartphone usage”. In: Procedings of the ACM international conference on MobileSystems, Applications and Sevices. ACM, 2010.

[5] Sven Ove Hansson. Teknik och etik. Avdelningen för Filosofi, Institutionen för Filosofioch Teknikhistoria, KTH, Stockholm., 2009.

[6] A. Keon Jang, Mongnam Han, Soohyun Cho, Hyung-Keun Ryu, Jaehwa Lee,Yeongseok Lee, and Sue B. Moon. “G3G and 3.5G wireless network performance mea-sured from moving cars and high-speed trains”. In: Proceedings of the ACM workshop onMobile internet through cellular networks. 2009, pp. 19–24.

[7] Tova Linder, Pontus Persson, Anton Forsberg, Jakob Danielsson, and Niklas Carlsson.“On Using Crowd-sourced Network Measurements for Performance Prediction”. In:Proceedings of the Annual Conference on Wireless On-demand Network Systems and Services(WONS). 2016.

[8] Douglas C. Montgomery, Elizabeth A. Peck, and G. Geoffrey Vining. Introduction toLinear Regression Analysis. Wiley, 2012.

[9] N.J.D. Nagelkerke. “A Note on a General Definition of the Coefficient of Determina-tion”. In: Biometrika 78 (3): 691-692. (1991).

[10] Helen Nissenbaum. Privacy as Contextual Integrity. Washington law review association,2004.

27

Bibliography

[11] Kiri Wagstaff, Claire Cardie, Seth Rogers, and Stefan Schroedl. “Constrained K-meansClustering with Background Knowledge”. In: Proceedings of the International Conferenceon Machine Learning. 2001, pp. 577–584.

[12] Sanford Weisberg. Applied linear regression. Wiley, 2005.

[13] Jun Yao, S.S. Kanhere, and M. Hassan. “Improving QoS in High-Speed Mobility UsingBandwidth Maps”. In: IEEE Transactions on Mobile Computing (2012).

[14] Seung Min Yu and Seong-Lyun Kim. “Downlink Capacity and Base Station Density inCellular Networks”. In: Proceedings of the Int.Symp.Modeling and Optimization in Mobile,Ad Hoc and Wireless networks (WiOpt). (Tsukuba Science City). 2015, pp. 119–124.

28

På svenska

Detta dokument hålls tillgängligt på Internet – eller dess framtida ersättare –under en längre tid från publiceringsdatum under förutsättning att inga extra-ordinära 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 förickekommersiell forskning och för undervisning. Överföring av upphovsrättenvid en senare tidpunkt kan inte upphäva detta tillstånd. All annan användning avdokumentet kräver upphovsmannens medgivande. För att garantera äktheten,säkerheten och tillgängligheten finns det lösningar av teknisk och administrativart.

Upphovsmannens ideella rätt innefattar rätt att bli nämnd som upphovsman iden omfattning som god sed kräver vid användning av dokumentet på ovanbeskrivna sätt samt skydd mot att dokumentet ändras eller presenteras i sådanform eller i sådant sammanhang som är kränkande för upphovsmannens litteräraeller konstnärliga anseende eller egenart.

För ytterligare information om Linköping University Electronic Press seförlagets hemsida http://www.ep.liu.se/

In English

The publishers will keep this document online on the Internet - or its possiblereplacement - for a considerable time from the date of publication barringexceptional circumstances.

The online availability of the document implies a permanent permission foranyone to read, to download, to print out single copies for your own use and touse it unchanged for any non-commercial research and educational purpose.Subsequent transfers of copyright cannot revoke this permission. All other usesof the document are conditional on the consent of the copyright owner. Thepublisher has taken technical and administrative measures to assure authenticity,security and accessibility.

According to intellectual property law the author has the right to bementioned when his/her work is accessed as described above and to be protectedagainst infringement.

For additional information about the Linköping University Electronic Pressand its procedures for publication and for assurance of document integrity,please refer to its WWW home page: http://www.ep.liu.se/

© Martin Claesson and Lovisa Edholm