Throughput Prediction in Cellular Networks:

Experiments and Preliminary Results

Alassane Samba, Yann Busnel, Alberto Blanc, Philippe Dooze, Gwendal
Simon

To cite this version:

Alassane Samba, Yann Busnel, Alberto Blanc, Philippe Dooze, Gwendal Simon. Throughput Predic-
tion in Cellular Networks: Experiments and Preliminary Results. CoRes 2016, May 2016, Bayonne,
France. �hal-01311158v1�

HAL Id: hal-01311158

https://hal.archives-ouvertes.fr/hal-01311158v1
Submitted on 10 May 2016 (v1), last revised 19 May 2016 (v2)

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
Throughput Prediction in Cellular Networks:
Experiments and Preliminary Results
Alassane Samba1 , Yann Busnel2 , Alberto Blanc3 , Philippe Dooze1 ,
Gwendal Simon3
1 Orange Labs 2 Crest (ENSAI) / Inria Rennes 3 Telecom Bretagne - IRISA

Throughput has a strong impact on user experience in cellular networks. The ability to predict the throughput of a
connection, before it starts, will bring new possibilities, particularly to the Internet service providers. They could adapt
contents to the quality of service really reachable by users, in order to enhance their experience. First this study high-
lights the prediction capabilities thanks to different algorithms and data gathered at different network levels. Then we
propose a simple approach based on machine learning to predict the throughput using a few data related to the context
of use.
Keywords: cellular network, throughput prediction, machine learning

1 Introduction
In cellular networks, Quality of Service (QoS), throughput in particular, depends on user context (radio
channel quality, speed, distance from base station, etc.). To enhance the Quality of Experience (QoE),
content providers implement adaptive delivery strategies, where the quality and the characteristics of the
delivered content are adjusted to match the QoS of each user. These adaptive strategies are reactive : at a
given time of the delivery, the throughput for the next x seconds is predicted based on the real throughput
observed during the past y seconds.
Yet, the content providers take some key decisions at the beginning of the delivery. For instance, most web
services have several style sheets for their web page, with a variable number of elements and information.
The decision of which style sheet to deliver should be taken in the early moment of the connection although
no past throughput observation is available. This calls for new mechanisms to provide a rough prediction of
the throughput for the next x seconds, using only contextual information.
Existing solutions present several shortcomings preventing their widespread usage. For instance, a well-
studied approach is to estimate the bandwidth by a series of short path measurements, including Round-Trip
Time (RTT), and packet loss rate, but this approach requires exchanging data before making a decision.
Some proposals rely on instant Channel Quality Indicator (CQI) to estimate the instantaneous bandwidth,
but they do not target throughput prediction on a larger time frame. We discuss these approaches in Sec-
tion 2.
We aim at identifying which contextual parameters are the most relevant to predict the throughput during
a session in a cellular network. We have conducted a large-scale trial where users have performed a file
download test a thousand times in several locations, mobility and radio connection configurations. Data
related to context of use, radio access network performance and traffic quality have been collected for each
test. In this paper, we describe this trial and we provide some early results of the statistical correlations
between the main contextual information and the actual throughput.

2 Background and motivation

It is well known [MSMO97] that the throughput of a Transmission Control Protocol (TCP) connection
can be accurately predicted from a set of measures related to RTT, window size and packet loss rate.
 Alassane Samba, Yann Busnel, Alberto Blanc, Philippe Dooze, Gwendal Simon

E2E Measurements
Download Throughput
IP Setup Time
HTTP Setup Time Operator Network
Average RTT (TCP) Probe
Remote Server

App
Radio
Access
RAN measurements
Context Information Number of RRC attempts
Radio Channel Quality (SNR, RSRQ, RSRP, RSSI) Average Number of Users
Distance from the Base Station Average Number of Active Users
Speed Traffic Volume
Indoor/Outdoor Call Drop Rate
Device Category Call Setup Success Rate

F IGURE 1: Testbed architecture

More recent outcomes [BMW14] adapt analytical throughput models to cellular networks by integrating
the effects of the radio channel quality. These models, while extremely accurate, are not capable of making
long term predictions.
More pratical approaches such as [MSBZ07] and [LDJ+ 15] provide throughput prediction models. The
former paper addresses throughput prediction for fixed line connections with machine learning approaches
on several parameters, which are in part collected on the server side. This line of research does not address
cellular environments. The latter paper addresses the prediction of instantaneous throughput based on some
parameters that are collected at the mobile device side, including CQI and Discontinuous Transmission
(DTX). This study is however restricted to the prediction of instantaneous throughput. The prediction is not
accurate on a longer time frame since the CQI can change quickly, due to the changing radio conditions.

3 Experimental environment and methodologies

As shown in Figure 1, we have used an Android application [One16] to periodically download a file from
a remote server, using an HSPA+ connection. For each download the application logs several parameters at
the beginning and at the end of the connection, including location, speed, radio qualify parameters (such
as Signal-to-Noise Ratio (SNR), Reference Signal Received Quality (RSRQ), Reference Signal Received
Power (RSRP) and Received Signal Strength Indicator (RSSI)). We were also able to collect data directly
from the cellular network from a Radio Access Network (RAN) management system, which includes the
number of active users, so-called RAN measurements. Finally, using network probes inside the operator
network, we collected end-to-end (E2E) measurements for each test, including the download throughput
we are interested in, the average RTT during the test, the IP setup time meaning the time between socket
connection and first Hypertext Transfer Protocol (HTTP) request and the HTTP setup time meaning the
time between first HTTP request and start of the download.
We present in this paper the results from a first set of measurements made by three testers. This campaign
contains about 2600 observations from 40 different cells (without intra-cell handover). On average, the
file was downloaded in eight seconds, which corresponds to the duration of a video segment in adaptive
streaming. Unfortunately we miss some values due to the device firmware and Global Positioning System
(GPS) availability. For example, RAN state data are missing in 13% of cases and speed in 20%. We used the
optimal discretization algorithm defined by [Bou06] to deal with missing values and enhance the models.
Our goal is to predict the average throughput during the download test based only on context information.
Several supervised learning techniques can be used to make the prediction. We compare the results of three
algorithms, namely Generalized Linear Model (GLM), Neural Networks (NNET) and Random Forests
(RF). We gradually add each group of predictors to assess its contribution to the throughput prediction. A
Throughput Prediction in Cellular Networks: Experiments and Preliminary Results

F IGURE 2: Throughput link with some context indicators.

F IGURE 3: ECDF of error rates for the RF algorithms based on the three different types of inputs.

K-fold cross-validation [Koh95] permits to appraise the generalizability of models.

4 Evaluation and discussion

We first show in Figure 2 some typical correlations between the actual throughput and the measured
context information. Due to space constraints, we cannot detail these correlations in this paper.
The machine learning algorithms predict a throughput, which we then compare to the actual throughput
observed during the download test. Figure 3 shows the ECDF of the prediction error ratio, which is mea-
sured by the absolute value of the difference between the predicted and the actual throughput, divided by
the actual throughput. We highlight two observations. First, the context information can suffice to make
accurate predictions. The accuracy of the prediction is close to the prediction using both context and RAN
information. Second, the machine learning algorithms are effective. Typically, half of the throughput rates
are within 7% of the actual throughput while, for a fixed line, the previous proposals were within 10% of
errors for the best half of predictions [MSBZ07].

GLM GLM NNET NNET RF RF

10-fold 10-fold 10-fold
Context 0.82 0.80 0.84 0.80 0.84 0.84
Context + RAN data 0.88 0.85 0.89 0.86 0.89 0.88
Context + RAN data + other E2E measurements 0.95 0.93 0.95 0.93 0.95 0.94

TABLE 1: Coefficient of determination (R2) of models.

 Alassane Samba, Yann Busnel, Alberto Blanc, Philippe Dooze, Gwendal Simon

In Table 1 we compare the results of the three aforementioned algorithms. We present the coefficient
of determination (R2) obtained on each model and on its 10-fold cross-validation. R2 measures the part
of variance explained by a model from global target variable variance. It is calculated as follows : R2 =
2
∑ni=1 (ŷi −yi )
1− 2 , where n is the number of observations, yi is the target value of the i-th observation (i ∈
∑ni=1 (ȳ−yi )
{1, 2, ..., n}), ȳ is the average value of target and ŷi is the prediction for the i-th observation. Three algorithms
produce roughly equivalent results for throughput estimation according to R2. Nevertheless RF models
show better generalization capability than others, according to the small differences between R2 of models
from entire dataset and R2 of 10-fold cross-validation. In fact, RF algorithms are known to be less prone
to overfitting. These results also confirm that, unsurprisingly, the more data can be used by the machine
learning algorithm, the better the prediction. But the prediction accuracy does not increase significantly with
the RAN measurements (around 5% compared to the context information only). The accuracy increases with
the use of other E2E measurements (especially average RTT) to estimate throughput, which confirms the
results of previous papers. However, these data are not available at the beginning of the connection, while
the context information is.

5 Conclusion
Predicting a transmission throughput through cellular network using a small set of information avai-
lable before the connection is a challenge. Our results confirm the correlation between throughput and the
context information, which opens important perspectives regarding the development of adaptive delivery
techniques. Our approach uses information about the context and the coverage quality, which are available
before the connection. Adaptive behavior can thus be implemented according to predicted throughput by a
service provider to enhance QoE. A remaining open problem consists in taking into account sudden changes
in coverage conditions that should occur during a connection, like a handover. Supplementary methods that
tolerate evolution of coverage conditions and mobility prediction represents good candidate to handle this
challenge. In the future works, we will also study the correlations between each parameter of the context
information, as shown in Figure 2. Finally, we will make a deeper analysis by using the whole set of mea-
surements, which contains more than 50 testers in a wider range of configurations.

Références
[BMW14] Nicola Bui, Foivos Michelinakis, and Joerg Widmer. A model for throughput prediction for
mobile users. In European Wireless Conference, 2014.
[Bou06] Marc Boullé. Modl : a bayes optimal discretization method for continuous attributes. Machine
learning, 65(1) :131–165, 2006.
[Koh95] Ron Kohavi. A study of cross-validation and bootstrap for accuracy estimation and model
selection. In Int. Joint Conf. on Artificial Intelligence (IJCAI), 1995.
[LDJ+ 15] Feng Lu, Hao Du, Ankur Jain, Geoffrey M. Voelker, Alex C. Snoeren, and Andreas Terzis.
CQIC : Revisiting cross-layer congestion control for cellular networks. In ACM HotMobile
Workshop, 2015.
[MSBZ07] Mariyam Mirza, Joel Sommers, Paul Barford, and Xiaojin Zhu. A machine learning approach
to TCP throughput prediction. In ACM Sigmetrics conference, 2007.
[MSMO97] Matthew Mathis, Jeffrey Semke, Jamshid Mahdavi, and Teunis Ott. The macroscopic behavior
of the TCP congestion avoidance algorithm. SIGCOMM Comput. Commun. Rev., 27(3), 1997.
[One16] EQual One. Get a 360◦ vision of mobile service experience ! – http://www.v3d.fr/
solution/equal-one/. Vision 360 Degres (V3D), February 2016.