Source of information: http://prime.mines.edu/papers/tutorial-wsc05.pdf
In this paper, a method to analyze GSM network performance
on the basis of massive data records and application domain
knowledge is presented. The available measurements are divided into
variable sets describing the performance of the dierent subsystems of
the GSM network. Simple mathematical models for the subsystems are
proposed. The model parameters are estimated from the available data
record using quadratic programming. The parameter estimates are used
to nd the input-output variable pairs involved in the most severe performance
degradations. Finally, the resulting variable pairs are visualized
as a tree-shaped cause-eect chain in order to allow user friendly analysis
of the network performance.
The radio resource management in current mobile communication networks concentrates
on maximizing the number of users for which the services can be provided
with required quality, while using only limited amount of resources [4].
Once the network is designed and implemented, the goal is to find a network
configuration parameters that use the existing resources as effciently as possible
from the user point of view. In practice, this means that a reasonable tradeo
between the coverage and capacity of the network must be found. Good coverage
allows users to initiate services at any location with acceptable service quality,
while high capacity allows many network subscribers to use services simultaneously.
However, improving the coverage tends to diminish the capacity and vice
versa. A good tradeo between coverage and capacity is obtained when the number
of service denials (blocking) and abnormal service interruptions (dropping)
are at the minimum, i.e the performance of the network is well optimized.
In this paper, the performance of a GSM network is analyzed based on massive
data records and application domain knowledge. Next, the GSM network
infrastructure is shortly outlined. In Section 3, a hierarchical model for describing
the network performance is proposed. Then, the usage of the proposed model as
a part of an analysis process is presented. In Section 5, results of the experiments
are presented.
A GSM network consists of high number of sites, each usually having three base
station transceivers (BTS) positioned to cover separate sectors around the site
(see Fig. 1). Each BTS has one or more transceiver/receiver pairs (TRX), each
allocated on a single physical radio frequency. Base station controller (BSC)
manages the operation of several BTSs connected to it through the Abis interface.
A single mobile services switching center (MSC) is connected to several
BSCs through the A interface.
The performance of the mobile network is measured based on thousands of
counters, describing the numbers of the most important events over a measurement
period (typically one hour). Due to the high number of counters, a set of
high-level key performance indicators (KPIs) are de ned by the network manufacturers
in order to allow more e cient performance monitoring. Typically,
the KPIs describe the success/failure rates of the most important events such as
service blocking, service dropping and handovers.
Such indicators are traditionally used in resource management [2] and are
well suited for performance monitoring, but there are several drawbacks when
they are used in fault diagnosis [3]. Also, the most widely used performance
indicators describe the operation of the network at the BTS level. As a result,
the performance degradations originating from interaction between several BTSs
become very di cult to observe. In many cases, however, the operation of the
close-by BTSs are highly dependent on each other. Examples of operation in
which several BTSs interact are handovers between close-by BTSs and interference
between BTSs having TRXs on the same physical frequency.
The basic idea in the developed model is to measure performance in terms of the
number of failed operations in the network. Examples of such operations are an
attempt to allocate a signaling channel, an attempt to allocate a tra c channel
for a call, or to perform a handover between two neighboring BTSs. In order to
obtain good performance, the number of failing operations within the network
must be minimized.
The set of available measurements seems to consists of counter groups. Within
a counter group, the counters are clearly connected while the counters from
dierent groups are independent on each other. Therefore, the modeling problem
is more easily solved by identifying a separate subsystem for each counter group
separately. In Figure 2(a), the memberships of the variables in dierent counter
groups (subsystems) are shown. The subsystems tend to form a hierarchical
structure (see Figure 2(b)), i.e the outputs of a subsystem describing some lowlevel
phenomena can be an input to a higher-level subsystem. Next, the principles
used to divide the data generating system into subsystems is described.
The main purpose of the analysis is to locate bottlenecks in the network performance that have direct impact on user perceived quality. The system S1;1 describes the number of user perceived quality problems by summing the problems from four dierent categories, each focusing on dierent part of the transaction. This model is of type I (see Table 1), in which the number of user perceived quality problems in the whole network is the output variable y(t) and the set of input variables xi(t) consist of number of blocked channel requests, the number of call setup failures, the number of calls dropped during transaction, and the number of failed handovers, each computed over the whole analyzed network. The contribution of each problem type into the overall user perceived quality is described by the parameter ai, describing the percentage of the failures of type xi to the total number of failures y.
The number of blocked channel requests (rejects) measure the networks
ability to satisfy the demand generated by the network users. System S2;1
with a model of type I describes how many blocked requests y(t) in the network
originate from BTS i of the network (the variable xi(t) is the number of blocked
requests in BTS i at time t). The contribution of BTS i to the number of all
blocked requests in the analyzed network is described by the parameter ai.
The system S3;1 is described by a model of type I, in which the proportions
of stand-alone dedicated control channel (SDCCH) rejects x1(t), full rate tra c
channel (FR-TCH) rejects x2(t) and half rate tra c channel (HR-TCH) rejects
x3(t) to the total number of rejected channel requests y(t) is computed. This
model is estimated for each BTS separately. That is, the parameter ai describes
the contribution of the channel type i to the total number of blocked channel
requests.
Finally, the systems S5;7, S5;8 and S5;9 describe the above mentioned channel
type rejects due to congestion vs. other possible reasons for blocking. These subsystems
are described by a model of type II, in which the output variable y(t) describes
the number of blocked requests and the input variable x(t) = C(t)Rtot(t)
describes the proportion of channel requests assumed to have occurred during
congestion (C(t) denotes the percentage of time in congestion in time period t
and Rtot(t) is the total number of channel requests). These models include a bias
b since it is not expected that all request rejects are due to congestion, but also
other causes may exist (but measurements are not available). The minimization
of the mean square prediction error 1
T Pt e2(t) = 1
T Pt(y(t) ?? ^y(t))2 with the
corresponding constraints (see Table 1) leads to a standard quadratic programming
problem. For more information about algorithms to solve such problems,
see [1].
It is possible, that the user request is not served due to
problems in the resource allocation phase (call setup) of the transaction. As in
the case of service blocking, a model describing the contributions of each BTS
to the total number of call setup failures in the network is de ned (model S2;2).
Basically, the call setup phase includes allocation of a signaling channel in which
the negotiation for the actual tra c channel is performed. Model S3;2 of type
II divides the call setup failures in a single BTS into SDCCH and TCH setup
failures separately.
Both the SDCCH and TCH signaling may fail due to problems in dierent
network elements (or interfaces between them) involving in the signaling procedure.
The number of SDCCH and TCH signaling failures due to problems
in dierent network elements are described by the models S5;10 and S5;11, respectively.
These models are of type III and include a bias since it is possible,
that call setup failures are caused by other reasons for which measurements are
not available. Also, an equality constraint is introduced since it is necessary to
require that a single failure in a network element causes the failure of the call
setup phase of exactly one transaction.
The most common reasons for failures during the call setup phase or actual
service are the inadequate radio signal propagation conditions (problems in radio
channel in the air interface). The failures in radio channel are usually due to bad
signal quality, i.e the transmitted data includes too many bit errors. The models
S6;9 and S6;10 of type III describe the number of SDCCH and TCH radio channel
failures due to bad signal quality vs. other reasons.
The radio signal quality is mostly aected by two components. Firstly, the
propagation environment causes attenuation to the transmitted radio signal due
to path loss, shadow fading and multipath fading. Secondly, the radio signal may
be attenuated by the other radio signals originating from other BTSs having a
TRX on the same physical frequency (interference). The purpose of the models
S7;2 and S7;3 of type III is to compute how many bit errors in uplink (from MS
to BTS) and downlink (from BTS to MS) tra c are due to di cult propagation
conditions in the BTS's coverage area and how many are likely the result of
interference
When the call setup phase is successfully completed, the actual
service (usually speech in GSM networks) is started. However, the service may
be abnormally interrupted (dropped) due to several reasons. The purpose of the
model S2;3 is to describe how many calls are dropped in BTS i w.r.t the number
of dropped calls in the whole analyzed network. The model is of type I and the
contributions of each BTSs to the total number of dropped calls is described by
the parameter ai (similarly to the blocking and call setup failures).
The call may be dropped due to internal failures in the network elements
or interfaces between network elements. The purpose of the model S5;12 is to
describe the contributions of the possible causes (very similar to the causes of
call setup failures). This model is of type III, having a bias since a call may
be dropped due to reasons for which measurements are not available. As in the
case of call setup failures, most of the dropped calls are expected to be due to
radio channel (air interface) problems. Therefore, the same models explaining
the number of call setup failures due to bad signal quality also describe the
number of dropped calls due to radio channel problems.
As in the previous cases, also the (outgoing) handover failures
(HO) are divided into network level variable and BTS level variables. The
model S2;4 is used to compute the value for parameters ai describing the percentage
of handover failures originating from BTS i. The model S3;3 divides the
handover failures according to the handover type (within BTS HO, BSC controlled
HO and MSC controlled HO), and a model of type I is used to obtain the
contributions of the dierent handover types into the total number of handover
failures.
Both the BSC and MSC controlled outgoing handovers may fail due to problems
in the source (serving) BTS or the target BTS. The serving BTS problems
can be various BSS problems (very rare in practice) and the target BTS problems
may be due to lack of resources, BSS level problems (rare) or problems with
the connection (radio link) to the target BTS. Models S4;2 and S4;3 explicitly
describe the dependencies between these failures in dierent BTSs, i.e the cause
for failed outgoing handover may be in lack of resources or connection in any of
the BTSs around the same operation area. Model S4;1 describes the within BTS
handover failures due to BSS problems or lack of resources. These three models
are of type IV in which both the equality constraint and box constraints for the
parameters are used.
Model S5;2 describes the causes for the target BTS radio channel failures and
the model S5;3 describes the reasons for BSS problems in the target BTS that
caused the failed outgoing handover from the serving BTS. Models S5;4 and S5;5
describe the causes for the failing BSC and MSC controlled outgoing handovers
due to lack of resources in the target BTS, respectively. Similarly, model S6;8
describes the causes of the lack of resources in within BTS HO attempts. The
purpose of these three latter models is to analyze the number of HO failures
per HO type (SDCCH-SDCCH, SDCCH-TCH, TCH-TCH) due to SDCCH, HRTCH
and FR-TCH congestion vs. other unmeasured causes. These models are of
type V , and contain signals c1(t), c2(t) and c3(t) that describe the percentages of
SDCCH, FR-TCH and HR-TCH congestion w.r.t the length of the measurement
period (one hour in our case) and xi(t) denotes the number of HO attempts.
Since both the BSC and MSC controlled handovers may be of one of the
three above mentioned types, there are six dierent kind of handovers involving
distinct target and source BTS. Models at the level six all describe the percentage
of incoming handovers originating from dierent source BTSs, one model per
each HO type. These models can be used to nd out which close-by BTSs are
generating the major portion of the handover load to the target BTS during
handovers. Finally, the model S7;1 describes the causes for BSC controlled TCHTCH
handovers (most typical type of handover).
In order to estimate all subsystem models, the data must be carefully preprocessed.
In this work, the preprocessing phase includes outlier removing, data segmentation
and constant variable pruning. The number of outliers (points clearly
diering from other measurements) is quite high in this type of application. Such
samples are generated during network recon gurations or hardware breakdowns,
typically lasting only few hours. Such time points should be removed since they
do not help in nding the major bottlenecks in network performance. In TRX
level model construction, it is also necessary to segmentate the data into subsegments
(time periods) during which the number of TRXs on a certain physical
frequency do not change.
After outlier detection and data segmentation, the model parameters can be
estimated from the data. However, in some network elements rarely suering
from any types of failures, some or most of the signals in the model are nearly
constant. In such a case, the data is not rich enough for estimating a model.
Instead, the (nearly) constant input signals are pruned before the model is estimated.
In the case of constant signal being an output variable, the model is not
estimated at all.
After the data is cleaned, the models can be estimated using standard quadratic
programming techniques. After the parameters of the subsystem models have
been estimated, an item to a dependency list is generated per each input-output
variable pair. Each item in the dependency list include the strength of the dependency
between the input and output variable (the value of the parameter
a), a measure of model accuracy (root mean square prediction error (RMSE)
of the model), and a measure of models importance in overall network performance
analysis (the average number of failures stored in the output variable of
the input-output variable pair).
After all the models have been estimated and the properties of each inputoutput
variable pairs are stored in the dependency list, a tree-shaped graph
is constructed in order to analyze the cause-eect chains generating the major
performance degradations of the network. Since the number of theoretically possible
dependencies is extremely large, only the most important dependencies are
included to the dependency tree.
Three criteria are used to prune uninteresting dependencies from the tree.
Firstly, the model accuracy from which the dependency originates must be at
a reasonable level. Otherwise, the analysis might be mislead by very inaccurate
models having large values for parameter a (which is forced in several models due
to the equality constraints for the parameter vector). Secondly, the output variable
of the dependency must be interesting enough (i.e relatively large number of
failures must be observed in the output variable). Finally, only the dependencies
that belong to the cause-eect chains contributing most to the overall network
performance degradations are included into the dependency tree. For each subsystem,
dierent minimum and maximum values for strength of dependency,
model accuracy and model interestingness are defined.
The analyzed GSM network data contained 120 BTSs, in which 101 most important
variables (counters) were measured during a two-month time period.
Since the variables were divided into 33 subsystems consisting of 5 network level
systems, 24 BTS level systems, 4 TRX level systems and 143 frequency related
segments with non-changing frequency plan, we have 5+ (12024)+ (1434)
models with a corresponding parameter vector a estimated using quadratic programming.
Figure 3(a) shows the root mean square errors (RMSE) of the models. Note,
that the RMSE of the models can be interpreted as the number of user perceived
quality problems that could not be explained by the model. Models in which the
RMSE is below 50 failures can be regarded as accurate enough in order to
make useful inferences about the data. Clearly, there are lots of models that are
accurate enough, but also many models are not accurate enough to allow any
justi ed conclusions to be made. Also, three models seem to be very inaccurate.
Figure 3(b) shows the means of the output variables of the models. These
values measure the number of user perceived failures of the subsystems. Therefore,
it can be regarded as a measure of interestingness or importance of the
subsystem in the performance analysis.
Figures 4(a)-(d) show the pruned dependency trees in four separate cases. In
Figure 4(a), the cause-eect chains of the most signi cant blocking problems are
shown. Clearly, there are 4 BTSs (6,11,18,85) that suer from lack of resources.
BTSs (6,11,18) suer from lack of half rate tra c channels and BTS 85 suers
from lack of full rate tra c channels. Only BTS 6, the causes for blocking can
be said to result regularly from congestion.
In Figure 4(b), the results of the corresponding analysis for the call setup
failures are shown. Here, four BTSs (17,52,66,74) seem to suer from call setup
failures regularly. In all these four BTSs, the failures tend originate during SDCCH
signaling and fail due to radio link problems. In BTSs 17 and 74 the radio
link failures can be said to result from bad downlink signal quality and in BTSs
52 and 66 they are due to bad downlink signal quality.
Figure 4(c) shows the corresponding results for call dropping problems. Again,
the cause-eect chains for describing the reasons for dropped calls in four BTSs
(6,7,69,74) are shown. The reasons for call dropping seem to be radio link failures.
In BTSs 6, 7 and 69 the radio failures are likely due to bad signal quality
in uplink. In two TRXs of BTS 6 and in one TRX of BTS 69 the number of bit
errors seem to correlate with the amount of tra c in both the own BTS as well
as in interfering TRXs on the same radio frequency.
Finally, in Figure 4(d) the analysis of the handover problem sources are
shown. The results for three BTSs (7,9,90) having the worst handover performance
are shown, indicating that the problems are in BSC controlled outgoing
handovers. In BTSs 7 and 9, the problems seem to be in lack of resources in
the target BTSs (6,11,85). In target BTS 6 suering from lack of resources,
there seems to be high amount of incoming TCH-TCH handover attempts from
BTS 7. This same BTS tend to cause problems also for target BTS 11 suering
from lack of resources during handover attempts. For target BTS 11, two other
BTSs (12,18) are found that can be said to generate high number of handover
attempts. The handover attempts of these two BTSs are due to very similar
reasons: the quality of the uplink and downlink radio connection and the uplink
and downlink signal strength are not reasonable in these two BTSs, and some of
the users are switched into more appropriate BTSs. Also, signi cant number of
users are switched to another BTS in order to minimize the energy consumption
of the MSs (power budget).
In this paper, a knowledge-based model for analyzing the performance of the
GSM network is presented. The presented model is based on subsystem hierarchy
that was developed during an explorative data analysis process involving
iterative literature research and data visualization. As a result, a logical division
of the system into subsystems with the corresponding input and output variables
were established. Also, the type of the model per subsystem required a priori
knowledge about the semantics of the subsystem variables.
Nokia Foundation is gratefully acknowledged for their financial support.
1. S. Mokhtar Bazaraa, D. Hanif Sherali, and C. M. Shetty. Nonlinear Programming:
theory and algorithms. John Wiley and Sons, Inc., 1993.
2. Sofoklis A. Kyriazakos and George T. Karetsos. Practical Radio Resource Manage-
ment in Wireless Systems. Artech House, Inc., 2004.
3. Pasi Lehtimaki and Kimmo Raivio. A SOM based approach for visualization of GSM
network performance data. In Proceedings of the 18th Internation Conference on
Industrial and Engineering Applications of Arti cial Intelligence and Expert Systems
(to appear), 2005.
4. Jens Zander. Radio Resource Management for Wireless Networks. Artech House,
Inc., 2001.