Upload
fawaz-agha
View
216
Download
0
Embed Size (px)
Citation preview
8/6/2019 Directed Retry Blocking Paper
1/9
ANALYTICAL MODEL FOR EVALUATING BLOCKING IN GSM
NETWORKS WITH DIRECTED RETRY
Arnie Neidhardt, Judith Jerkins, and K. R. Krishnan
Telcordia Technologies, New Jersey, U.S.A.
E-mail: {arnie, jlj, krk}@research.telcordia.com
ABSTRACT
We present an analytical model and propose an efficient algorithm for evaluating call blocking
in GSM networks with directed retry, where a call blocked by the sector of its initial attempt
may be redirected to one or more alternative sectors. Since exact solution by Markov-chain
analysis is unfeasible in large networks owing to the size of the state-space, we propose a method
for an approximate solution, which is shown to be highly accurate in two small examples.
.
1. Introduction
In this paper, we propose a method for evaluating
call blocking in GSM networks with Directed Retry,
in which a call blocked by the sector1
of its initial
attempt may be redirected to one or more alternative
sectors that are also accessible to the call. The
purpose of directed retry is to increase the effective
capacity of a network by making it possible for calls
to be served by the frequencies of multiple sectors (so
long as the frequency assignments in sectors reduces
co-channel interference and provides calls with low-
interference access to multiple sectors).
We present an analytical model and propose an
efficient computational method for estimating
blocking in networks with directed retry, which can
be used, in particular, to compare network
performance for different frequency assignments. An
efficient analytical algorithm for performance
evaluation is attractive, even if its results are only
approximate, since the alternative of using detailed
simulations can be a very time-consuming process.
The general problem of estimating blocking in
networks employing Directed Retry has been
addressed in previous literature [1-4]. While themethods proposed in all these papers are based on the
assumption that blockings in different sectors are
independent, a distinctive feature of our method is
that such independence is not assumed. A uniform
loading in all cells is assumed in [1,4], and only one
overflow attempt for calls eligible for directed retry is
actually considered in [1,3], though the possibility of
extension to attempts on multiple overflow paths is
mentioned. The analysis in [2] accounts for the
possibility of more than one overflow path with
directed retry but imposes a given retry sequence on a
1In this paper, the term sector is used to mean a base station andits primary group of frequencies, and, occasionally, also the
geographic area served primarily by the base station.
specified fraction of all call attempts in a cell,including already-redirected calls. Our method can be
applied in arbitrary finite networks, and conforms to
the specified attempt-sequences for calls eligible for
directed retry.
The papers [1-4] above consider handoff calls
explicitly, which, like the fresh calls, are also
assumed to constitute a Poisson arrival process of
known rate. It follows that, with that assumption,
handoff calls can be taken into account by redefining
the source traffic to include such handoff calls aswell. This is the point of view adopted here.
In our analysis, the network is visualized as a set
of non-overlapping geographical bins (e.g., a grid).
We assume we are given the call-origination load foreach bin, and the sequence of primary and overflow
servers that provide service to those calls. The call-
origination load is defined as the traffic load imposed
on the network by calls initiated by users situated in
the geographical region of the bin, as well as by
existing calls that were handed off as the user enteredthe bin. The server-sequence, in general, can
correspond to servers in the primary and overflow
sectors of directed retry that serve the bin, ranked in
decreasing order of signal clarity for calls originatingin the bin. We assume that the concept of Directed
Retry applies to both initial calls and handoffs. Theobjective is to determine the blocking experienced by
calls originating or handed-off in each bin, as well as
the overall network blocking.
The overall problem can be viewed as one where a
set of traffic sources (users in the bins) is served, each by a given subset of servers, attempted in a given
sequence of primary and overflow servers, with the
different sources possibly limited to different subsets
of servers. Thus, the problem corresponds to a
generalized limited-availability system, special formsof which have been investigated in the telephony
literature [5]. The exact solution of such problems by
8/6/2019 Directed Retry Blocking Paper
2/9
the analysis of state-evolution of Markov chains is
rendered unfeasible in realistic-sized networks by the
explosion in the number of possible states. Therefore,we propose a method that allows us to obtain an
approximate solution by solving a system of fixed-
point equations whose solution is easily obtained by
means of an iterative procedure that is guaranteed toconverge.
We wish to arrive at a description of network load
due to call-segments. It is clear that the average
segment duration is shorter than the average sessionduration, but the average initiation rate for segments
is higher than the average initiation rate for sessions
(since a single session can spawn several segments
during its life). The magnitude of the overall load,which is the product of an initiation rate and an
average duration, is independent of whether the ratesand durations refer to segments or to sessions. For the
situation being considered here, the relevant
magnitude of demand is the product of the segment-
initiation rate and average segment duration. In
particular, the traffic demand for each segment will
be assigned to the bin in which the phone is located atthe time the segment is initiated. To allow for the
specification of a well-posed problem without having
to specify the mobility patterns of users, the
assumption will be made that, during a period of
interest for performance (such as a busy hour), thedemand for call-segments takes the form of a Poisson
process, exactly as assumed in [1-4]. In the remainder
of the paper, we use the terms call and call-segment
interchangeably, to mean call-segment.
In Section 2, we introduce the notation needed for
a mathematical description of the problem, and
proceed to develop the general algorithm for
evaluation of call blocking, on the basis of certain
approximations that ensure tractability of the
computations for networks of realistic size. In Section3, we illustrate the application of our method in two
examples, for which, in fact, exact solutions by
Markov chain analysis is also practical because of the
limited number of possible states. In the first example
of a 2-sector network with directed retry, whichinvolves 64 states, our method is shown to achieveexcellent accuracy in the determination of the
capacity increase made possible by directed retry,
with a relative error that is smaller than about 1.25%
over a wide range of overlap between the sectors. For
a symmetric 3-sector network, with 1202
states, the
relative error remains within 3.3% over a range of
overlap among sectors. Section 4 summarizes our
paper and suggests the desirability of applying the
method to realistic-sized cellular networks.
2.2. Server Teams
A call that seeks a traffic channel (time-slot) from
a sector can be served by any available traffic channel
of any frequency in the sector. Therefore, we regard
the combined group of traffic channels corresponding
to all the frequencies in a sector as an indivisible
server-team, presenting itself as an indivisible unit
(which we will sometimes refer to as an atom) to allcalls that seek service from it. This indivisible server-
team corresponding to a sector can be thought of as a
full-access trunk-group in the general teletraffic
parlance, with the number of trunks equal to the
total number of traffic channels in the frequencies
serving the sector.
2.
Mathematical Model for Service withDirected Retry
2.1.Calls and Call-SegmentsFor a user on a mobile phone, a single
communication session may require the support of a
succession of servers as the user traverses a path
through the domain of the network. From the users
perspective, the one long session is a single call, but
from the perspective of the network, the shortersegments of the session (corresponding to different
traffic channels, i.e., servers, being used to support
different segments of the session upon each handoff)
represent different demands on the network. Each
segment represents an occasion on which the network
might refuse to carry the call. Here we view all
occasions of not serving the call (whether in the
initial segment or in handoffs) with equal concern,
and so we shall treat each segment of a
communications session as a separate call; with this
language, every call not served initially, and every
call dropped during handoff, will be described as a
blocked call3.
2.3. Notation
The primary notation used for our analysis is given
here.Let
=T the set of all sectors in the network
=tS the indivisible server-team (atom) for sector
Tt =B the set of bins in the network
=bl the rate of call-segment originations in bin
b (assumed to be a Poisson process)
=bt the mean duration of call-segments
originating in bin b
bbba tl=
=bL the sequence (ordered list) of indivisible2 The 3-sector problem considered has 512 states, but with the
assumption of symmetry, this is reduced to 120.
3 A call not served in its initial segment would typically be
designated a blocked" call, while a call that is not served in a later
segment would typically be designated a call dropped during
handoff.
8/6/2019 Directed Retry Blocking Paper
3/9
server-teams that a call originating in bin b can
attempt
=UL the unordered set of the elements in any
server-team sequence L
=bN =)( bLN the total number of servers in bL
=
)(LN the total number of servers in anyserver-team sequence L
X = the number of elements of the set X
(ordered or unordered)
The aim of the investigation is to estimate , the
blocking probability for call-segments originating in
bin , . In the course of constructing an
algorithm for this purpose, we shall also make
estimates of , the load offered to the indivisible
server-team
bb
b Bb
tz
tSt T, .
As stated earlier, the server-team of a sectorpresents itself as an indivisible trunk-group to all calls
that seek service from it. Hence, , the sequence of
server-teams accessible to a call-segment originating
in bin b , can be viewed as a sequence of trunk-
groups. Call-segments originating in bin b attemptthe trunk-groups in the specified sequence, but are
free to access any idle trunk within each trunk-group.
bL
A Caveat on Notation
The analysis in this paper proceeds on the basis of
several approximations that we introduce in order to
achieve an efficient algorithm for a computational
solution. It is to be understood that the variousquantities we introduce and calculate are, in fact,
estimates of their true values, derived on the basis
of the approximations. In order to avoid introducing
burdensome new notation to distinguish each such
estimate from the corresponding true value, we shall,
in fact, re-use the notation introduced for a true valuealso for its estimate that appears in our calculations,
depending on the context to make our meaning clear.
Notation applied to Two-Sector Example
To illustrate the use of the notation we have
introduced, consider the simple two-sector network
example of Figure 1. The two sectors have a region ofmutual overlap, such that calls originating within the
region of overlap can be served by the servers of both
sectors, while calls originating in the non-overlap
region of each sector are served only by the servers of
that sector. Let the combined Poisson offered load of
the two sectors be erlangs, out of which a fractionAarises from the region of overlap. The frequencies
in Sector 1 correspond to an indivisible server-team
with servers and those in Sector 2 correspond
to an indivisible server-team with servers.
1S 1s
2S 2s
For simplicity, we assume symmetry in the
distribution of demand. Of the portion of
the total load that originates outside the overlap
region, we take
)1( gA -
2
)1( gA -
( 1s +
1S
2S
1S
bLbb },,
}
to arise from each sector,
being served only by the servers of that sector. The
load , originating in the region of overlap, is
served by a team of servers. However, a
mobile in the overlap region does recognize a
difference in signal clarity between the frequencies of
the two sectors, and its call is set up on a frequency ofthe clearer sector if a traffic channel is free, otherwise
it is attempted on the other sector. Thus, half of the
overlap load comes from calls that first attempt
the server-team , and then, if blocked, overflow to
the other server-team , while the other half of the
load is due to calls that attempt the server-teams in
the sequence .
Ag
,bb
)2s
B
Ag
2S
,t
B
=
b
b
b
b
b
=
bnet
b
b
t=b b
=
b
bnet
a
}bt
}
lb
This two-sector network may be viewed as
consisting of four bins, as shown in Figure 1, with
the pattern of offered loads and server-sequences
shown in Table 1. We will return to this example after
an exposition of our performance evaluation
algorithm in the next Section.
2.4.Performance Evaluation
Problem Statement
Given{ , we wish to determine
the blocking for the call-segments of the individualbins{ , and the overall network blocking
, given by
bl
b
netb
b
b
b
bb
b
a
a
t
t
b
l
l
(1)
If it is reasonable to work with the approximationthat the mean holding times are the same for all bins,
i.e., t for all , then we have
b
bb
a
b
b (2)
This is the case we consider in this paper, so that
the analysis can proceed from knowledge of only the
bin-loads{ , without the need for the separate
specification of{ and{ .
ba
}
8/6/2019 Directed Retry Blocking Paper
4/9
g
Bin 4Bin 1 Bin 2 Bin 3
Figure 1: Four Bins in the 2-Sector Network
Bin b Offered Load ba Server Sequence bL
12
)1(1
gAa
-
=
1S
22
2
Aga =
21SS
32
3
Aga =
12SS
42
)1(4
gAa
-
= 2S
TABLE 1: Bin Structure of 2-Sector Network
Motivation for an Approximation in the Solution
In principle, the exact solution to the problem
framed above can be obtained by solving the
fundamental state-transition equations of the
underlying Markov chain. However, since trunk-group
has states, the total number of states, given
by , becomes astronomically large for
realistic-sized networks, and the exact solution defies
computation.
tS
t
[ 1+ts[ +ts 1
]]
T
For this reason, we seek a simpler, approximate
formulation in which we express blockings of sets of
servers (i.e., server sets) in terms of equivalent loads
offered to them. In particular, there is the general case
of a set of limited-access trunk-groups, in which
different classes of calls might have their access limited
to just subsets of the trunk-groups. Various instances oflimited-access servers known as graded multiples
have been studied in the literature of telephony [5]. The
approach we present can be thought of as a method of
obtaining an approximate solution to the general
problem that includes graded multiples.
The Relevance of the Load on the Server-team
Sequence bLWhether a specific call-segment originating in a bin
gets served depends only on whether or not all the
servers of the server-team sequence are
occupied at the time that call-segment is offered,
regardless of the order in which the servers are
attempted. However, the probability that a call
originating in bin is blocked could be influenced by
the server-team sequences of all the bins, because those
sequences affect the total load imposed on the
servers in . This dependence arises from the
possibility of cascades of blocking, when blocking atone team of servers in one location leads to the
b
bN bL
b
bN
bL
8/6/2019 Directed Retry Blocking Paper
5/9
redirection of some load to another team at another
location, which, in turn, might block and redirect other
load to yet another team. t
bt
S
bf
team-serverthetooffered
isthatbinofloadofferedoffraction=.
Now, , then . If , we write
, where is the initial server-sequence
appearing before the appearance of in , and is
the remainder of the sequence after the appearance of
. In this case, the fraction of load that is offered
to in is the fraction that is blocked by the set of
servers in the initial sequence . Since the
number of servers is
bt LS if
bttRS
bL
UbtI
0=btf bt LS
t bL
ba
btI
btb IL =
tS
tS
btI
S btR
btI , we write
We will approximate this complicated dependence
of the blocking probability of bin b on all the
redirected loads by calculating the blocking probability
when a single effective Poisson load Y is offered to
the unordered set of servers . Thus, we estimate
by the following Erlang-B formula (the
appropriateness of the use of the Erlang-B formula will
be discussed later):
bb
U
bL
b
bb
)(,(U
btbtbt IYIBf = , (5),(3).( bbb YNB ,=b )
System of Equations for Determining Loadswhere stands for the yet-to-be-defined load
offered to the set of servers .
)(U
btIY
UbtI
We turn now to the derivation of a system ofequations from which the loads { in (3), associated
with a particular bin b and offered to the set of
servers , can be obtained. This will be a system of
}bY
U
bL
T equations, where T
=
is the number of sector
server-teams, each team being an indivisible atom (full-
access trunk-group) to all calls seeking service from it.
In Section 2.3, these server-team atoms were denoted
by , with the load offered toTtSt, tz TtSt,
In the special case where itself is the first element
of , ; this special case can be included in the
above notation by defining, in this case, and
further defining Y .
tS
(B
bL 1=btf
,btI
1)0,0,0)(
We thus have
( )
=
bt LSb
Ubtbtbt IYIBaz
:
)(, , (6).Tt
In deriving the equations for { , we find that we
also have to consider loads other than only the { .}bY
}bY
For , the right hand side of the system of
equations (6) involves, in general, offered loads on
various unordered sets of servers V , which are
unspecified as yet. We next specify these loads in Step2.
Tt
UbtI=
Load on Server-Team Atoms tS
The equations for determining the load Y on the
unordered set of server-teams are derived in two
steps.
b
U
bL Step 2: The Specification of Y in terms of{ )(V }tz
Consider a general unordered set V of server-team
atoms, say,nttt
SSS ,,21K=V .
Step 1: Equations for { }tz
In the first step, we write equations for the loads
offered to each of the atoms . The
formulation of equations for these atomic loadsrequires us, in general, also to deal with the concept of
the load Y offered to a limited-access set Vof
atoms, in which access for different calls could be
limited to different subsets of the component atoms.
TtSt,
)(V
If , i.e., if V consists of a single server-team,
say, , then, and is among the
loads to be determined by solution of (6).
1=n
tS tt zSYY == )(V)(
When , it is possible that different subsets of
serve different groups of calls, perhaps with no calls
given full access to all the servers in . Various
instances of such limited-access service systems have
been studied in the early literature of telephony [5]. Weintroduce the following definition of offered load forsuch a system:
2n
V
V
We begin by writing
(4),
=
Bb
bbtt afz
whereFor a general set V of two or more atoms
(server-teams), a bin-load is considered
offered to V if an atom of V appears as the
ba
8/6/2019 Directed Retry Blocking Paper
6/9
8/6/2019 Directed Retry Blocking Paper
7/9
practical terms, if one halts the iterative calculation at a
finite stage k at which the differences are
satisfyingly small, then the bounds show
that the solution is basically identified to within ones
desired accuracy. Since we have no proof of
uniqueness in general, cases of multiple solutions mayexist, and in such cases, the differences would
decrease to limits that are not all zero. In any case, ourproposed calculations can determine whether or not thesolution is unique for a network with given loads and
server-sequences.
kt
kt z-
ktt
ktz-
kt zz
kt
2)
),(
1
22
Az
zs
+
bL
}bb
,(
2
1sB
BAg
+
A
A
( 1
2
s
a
4
3
2
1
=
=
=
=
),(
),(
),(
),(
224
213
212
111
zsB
AssB
AssB
zsB
=
+=
+=
=
b
b
b
b
and
[ ]
+
++-
==
),(
),(),(2
)1(
21
2211
AssgB
zsBzsBg
a
a
b
b
b
bb
net
B
B
b
b
Comparison with Exact Solution
3. Examples: Comparison of Model with
Exact Solutions
For the case of two symmetrical sectors, each with
servers, the exact solution can be obtained by
analysis of the Markov chain that describes theevolution of the occupancy-vector of the two server-
teams . The number of possible states
is , and well-established numerical
methods were used to determine the equilibrium stateprobabilities.
7=s
(1+s
),( 21 SS
)(12+s 64)1 =
3.1.Symmetrical Two-sector Network
We now apply the performance evaluation algorithmdeveloped in Section 2 to the simple two-sector
network example that was described earlier at the endof Section 2.3. For this example (for which the loads
and server-sequences are summarized in Table 1), anexact solution is, in fact, feasible, and available forcomparison with the results produced by our method. Instead of making the comparison in terms of the
individual bin blockings in the two solutions, we
compare the capacity that is obtained in the 2-sectornetwork by having enabled a portion of the traffic to beserved by server-teams of both sectors. In particular,
under the performance constraint , we
estimate the total network load in erlangs that canbe served, for various values of the degree of overlap
01.0netb
A
g.
Offered Loads on Server-Team Atoms
For this example, 4,2 == BT , and Equations (6)
take the form:
2),
2),(
43122
22311
AgaazBaz
AzsBaaz
=++=
=++=
We present in Table 2 a comparison of the exact
solution with the approximate solution proposed here.For reference, we also include the trivial linearapproximation consisting of mere interpolation
between the known values at
and are readily solved by the iterative methoddescribed earlier.
Offered Load Y to Server-Team Sequenceb cells)t(coinciden1andoverlap)(no0 == gg .
Since we haveThe trivially interpolated result, overestimates the
capacity, and is subject to a maximum error of 7.5%.
24
123
212
11
SL
SSL
SSL
SL
=
=
=
=
We note that the approximation proposed in this
paper is slightly conservative, i.e., slightlyunderestimates the capacity, but is highly accurate: forthe cases considered, the relative error never exceeds
1.25%, and the mean relative error is only about 0.55%.
we obtain,
2
4321
4321
1
zY
aaaaY
aaaaY
zY
=+++
=+++
3.2.Symmetrical 3-Sector NetworkWe also carried out similar calculations for a
symmetric 3-sector network, with 7 servers in eachsector. Such a 3-sector network may be viewed as
consisting of 15 bins. For a range of sector overlaps,we again calculated the percentage error in theestimated total offered load that could be supported at a
1% blocking level. The degree of mutual overlap was
described by the fraction of calls in the network thathad access to only their home sector, to only two
sectors, and to all three sectors. The number of states in
from which we obtain the bin-blockings { and the
network blocking as follows:netb
8/6/2019 Directed Retry Blocking Paper
8/9
a corresponding Markov chain would be 512; however, because we assumed symmetry, we could perform anexact Markov-chain analysis with only 120 states.
Leaving out the details of the formulas andcalculations in the interest of space, we found that the
maximum relative error between the exact solution andthe approximation furnished by our method was lessthan 3.3%.
4. Summary
We have presented an analytical model forestimating the performance of GSM cellular networksthat employ directed retry, in which a call blocked by
the sector of its initial attempt may be redirected to analternative sector that is also accessible to the call,thereby increasing the effective capacity of the
network. In essence, the problem is that of performance
evaluation of a generalized limited-availability system[5], special forms of which have been investigated in
the telephony literature. Since the exact solution ofsuch problems by Markov chain analysis is renderedimpractical by the explosion in the number of possible
states, we have proposed a method that allows us toobtain an approximate solution by solving a system offixed-point equations by means of an iterative
procedure that is guaranteed to converge. Unlikeprevious methods reported in the literature, our methoddoes not assume that blockings in different sectors are
independent. It proceeds by associating an offeredload not only with each sector but also with the set of
servers in each sequence of sectors that corresponds toa permitted retry-sequence in the network, by definingthe notion of an equivalent offered load for a generallimited-availability system of servers in which traffic
streams may have only partial access to sub-groups ofservers. The performance evaluation algorithm can beused to compare GSM network performance and
capacity for different frequency assignment plans.
We have presented a comparison for two simple
symmetric networks for which the exact solution byMarkov chain analysis is also available. In a simple 2-
sector network, we have shown that our methodproduces excellent accuracy in the determination of thecapacity made possible by directed retry, with a relativeerror that is smaller than about 1.25% over a widerange of overlap between sectors. In the case of asymmetric 3-sector network, the relative error from the
exact solution is less than 3.3% over the range ofmutual overlap considered amongst sectors.
While such accuracy is reassuring, such simpleexamples are no substitute for testing the validity ofour approximations and the accuracy of our
calculations against actual measurements of performance from realistic-sized cellular networks,with due account taken of the inherent statisticalvariability in the measurements. Obtaining such datafor comparison from commercial cellular networks presents its own challenges. We hope that the
possibility of simple analytical methods of
characterizing network performance will spur interestin applying our method to actual networks.
REFERENCES
1. B. Eklundh, Channel Utilisation and BlockingProbability in a Cellular Mobile Telephone System
with Directed Retry, IEEE Trans on Comm., vol.COM-34, no.4, April 1986, pp.329-337.2. D. McMillan, Traffic Modelling and Analysis forCellular Mobile Networks, Proceedings, 13thInternational Teletraffic Congress, pp.627-632,Copenhagen (Elsevier, North Holland), 1991.
3. Tak-Shing Peter Yum and Kwan Lawrence Yeung,Blocking and Handoff Performance Analysis ofDirected Retry in Cellular Mobile Systems, IEEETransactions on Vehicular Technology, Vol.44, No. 3,August 1995, pp. 645-650.4. R. Verdone and A. Zanella, "Analytical Evaluationof Blocking Probability in a Mobile Radio System withDirected Retry", IEEE Journal on Selected Areas inComm., pp. 322-331, Vol. 19 No. 2, February 2001.5. R. Syski, Introduction to Congestion Theory inTelephone Systems, Second Edition, Elsevier SciencePublishers B.V., Amsterdam, 1986.
8/6/2019 Directed Retry Blocking Paper
9/9
g
Exact
Solution
(E)
Proposed
Method
(M)
Linear
Interpolation
(L)
%Error in
100*
-
E
ME
%Error in L
100*
-
E
LE
0.0 5.00188 5.00188 5.00188 0.00 0.00
0.1 5.11292 5.10827 5.23686 0.09 -2.42
0.2 5.23910 5.22852 5.47184 0.20 -4.44
0.3 5.38400 5.36582 5.70682 0.34 -6.00
0.4 5.55228 5.52434 5.9418 0.50 -7.02
0.5 5.74981 5.70951 6.17678 0.70 -7.43*
0.6 5.98360 5.92842 6.41176 0.92 -7.16
0.7 6.26071 6.19001 6.64674 1.13 -6.17
0.8 6.58539 6.50520 6.88172 1.22* -4.50
0.9 6.95406 6.88686 7.1167 0.97 -2.34
1.0 7.35168 7.35168 7.35168 0.00 0.00
TABLE 2: Traffic Capacity of Two Symmetrical 7-Server Sectors
Total Offered Load in Erlangs for Overall Blocking of 1%
*Maximum relative error for method