llc-mac analysis of general packet radio service in gsm
TRANSCRIPT
Bell Labs Technical Journal ◆ July–September 1999 37
IntroductionThe need to support voice services and user
mobility has led to standardization of one of the most
popular digital cellular standards, the Global System
for Mobile Communications (GSM), by the European
Telecommunications Standards Institute (ETSI). In
addition to speech applications, GSM can provide
some circuit-switched data services such as facsimile
transmission and short messaging services (SMS).
However, SMS volume is expected to exceed the GSM
capacity in two to five years. Due to recent develop-
ments in mobile computing devices and the emerging
market of multimedia communications, there is an
increasing demand for more advanced data service
providers in GSM. This has led to standardization of
two new data-bearer services as part of GSM
Phase 2+, namely, High-Speed Circuit-Switched Data
Service (HSCSD) and General Packet Radio Service
(GPRS), which operate in circuit-switched and packet-
switched modes, respectively. HSCSD allows a mobile
station (MS) to access more than one radio channel
operated in circuit-switched mode simultaneously,
thereby increasing the MS transmission rate to
64 kb/s. For applications generating bursty traffic, a
packet-switched bearer is more bandwidth efficient
than a circuit-switched one. Thus, the purpose of GPRS
is to statistically multiplex bursty data traffic efficiently.
GPRS technology is overlaid on the existing GSM
infrastructure. GPRS and GSM services share dynami-
cally the same radio resources and the same radio
channel structure. The GPRS network is seen as the
packet-switched extension of GSM. In GPRS, up to
eight users can be dynamically multiplexed on the
same radio channel if they require low bit rates, and
eight radio channels can be allocated to one user with
high-bandwidth requirements.1,2 Thus, GPRS provides
flexible bandwidth allocation and is economical to
users. The demand for voice is almost saturated, while
multimedia traffic is going to explode; therefore, the
evolution of the GPRS platform is seen as a lead-in to
third-generation wireless communication.
In order to maximize user data throughput and
decrease data-packet transfer time, the GPRS protocol
stack3,4 shown in Figure 1 should be optimally imple-
mented. In this figure, we highlight in color the proto-
♦ LLC-MAC Analysis of General Packet RadioService in GSMCristian Demetrescu
General Packet Radio Service (GPRS) is one of the major services currently standard-ized by the European Telecommunications Standards Institute (ETSI) for GlobalSystem for Mobile Communications (GSM) Phase 2+. GPRS has been introduced tosupport bursty packet-switched traffic such as e-mail, World Wide Web traffic, andtelemetry. This paper presents a unified analytical model of two GPRS protocols,namely, logical link control (LLC) and radio link control/medium access control(RLC/MAC). The interaction between selective automatic repeat-request (ARQ)schemes at the two layers is investigated. In order to save scarce radio resources, wecalculate a finite optimal number of RLC/MAC retransmissions per LLC frame as afunction of the number of radio link errors and the LLC frame size. A novel channel-dependent ARQ mechanism is introduced at the LLC layer. Finally, throughput anddelay performances of the two ARQ stacks are presented.
38 Bell Labs Technical Journal ◆ July–September 1999
col layers under consideration in this paper—that
is, the logical link control (LLC) and the radio link
control/medium access control (RLC/MAC) layers.
In order to protect LLC peer-to-peer communica-
tion, packet data units (PDUs) coming from higher lay-
ers are segmented into variable-size LLC frames, and 24
cyclic redundancy-check (CRC) bits are added to every
LLC frame for error detection. A stop-and-wait auto-
matic repeat-request (ARQ) mechanism is implemented
at the LLC to retransmit any erroneous LLC frames,
rendering to the LLC peer-to-peer link a high
reliability.3 Each LLC frame is passed to the RLC/MAC
layer, where it is further segmented into RLC/MAC
blocks of fixed size depending on the channel coding
scheme (CS) used at the physical layer.1,5 The
RLC/MAC layer provides a selective ARQ mechanism
for further error recovery over the radio interface (Um).
This paper describes the interaction between the
two ARQ mechanisms at the LLC and RLC/MAC lay-
ers, respectively. We assume that the LLC layer works
in the acknowledgment (ACK) mode so each LLC
frame is acknowledged upon receiving it. The main
objective of this work is to define analytically closed-
form expressions for the packet transfer delay and the
throughput that can be used for on-line quality-of-
service (QoS) management. The packet transfer delay
and the throughput seen at the LLC layer are two of
the most important parameters of the QoS profile;6
their on-line estimation is crucial for network perfor-
mance. We present a novel derivation of the ARQ
parameters (that is, the number of retransmissions) at
the LLC and RLC/MAC layers that strike a balance
between the bandwidth efficiency within the GPRS
access network and the efficiency over the Um.
Most of the papers published on LLC-RLC/MAC
performance analysis in GPRS7-12 make use of numer-
ical simulations to estimate the LLC throughput and
the LLC transfer delay when the LLC works in non-
acknowledgment mode. Ludwig and Turina12 derived
a formula for the average LLC frame transfer delay at
the RLC/MAC layer using an iterative method to cal-
culate the RLC/MAC retransmission probability. They
assumed Poisson arrival traffic and an infinite number
of RLC/MAC retransmissions per LLC frame.
However, their assumptions are general and do not
account for the LLC protocol.
In this paper, we derive closed-form expressions
for the LLC frame transfer delay and the throughput at
the LLC layer (irrespective of the arrival traffic) under
more realistic assumptions—namely, a finite number
of retransmissions per LLC frame and a two-state radio
channel model. This analytical model of the stack of
the two ARQ mechanisms at the LLC and RLC/MAC
layers is, to the best of the author’s knowledge, pre-
sented for the first time in this paper. Further, in order
to optimize radio resources, a channel-dependent ARQ
mechanism is introduced at the LLC layer.
Panel 1. Abbreviations, Acronyms, and Terms
A-bis—BTS/BSC interfaceACK—acknowledgmentARQ—automatic repeat requestBLER—block error rateBSC—base station controllerBSS—base station subsystemBTS—base transceiver stationCDF—cumulative distribution functionCRC—cyclic redundancy checkCS—coding schemeETSI—European Telecommunications Standards
InstituteGb—BSS/SGSN interfaceGGSN—gateway GPRS support nodeGPRS—General Packet Radio ServiceGSM—Global System for Mobile CommunicationsHSCSD—High Speed Circuit-Switched DataIP—Internet protocolLLC—logical link controlMAC—medium access controlMS—mobile stationNACK—negative acknowledgmentPDF—probability density functionPDU—packet data unitQoS—quality of serviceRLC—radio link controlSGSN—serving GPRS support nodeSMS—short messaging servicesSNDCP—subnetwork dependent convergence
protocolSN-PDU—SNDCP-PDUTCP—transmission control protocolUm—radio interface
Bell Labs Technical Journal ◆ July–September 1999 39
The section “ARQ Mechanism at RLC/MAC”
describes an analytical approach used to derive the
probabilities of RLC/MAC retransmissions when the
radio channel is modeled as a two-state Markov
process. This analysis gives us a closed-form expression
of the most likely number of RLC/MAC retransmis-
sions as a function of the RLC/MAC block error rate
(BLER) and the LLC frame size. The most likely num-
ber of RLC/MAC retransmissions can be used to esti-
mate the LLC frame transfer delay and the LLC
throughput as well as the signaling overhead required
by the ARQ mechanisms at both the RLC/MAC and
the LLC layers. This RLC/MAC protocol improves the
radio bandwidth efficiency by limiting the number of
RLC/MAC retransmissions per LLC frame when the
radio channel is in a “bad” state.
In the section “Channel-Dependent LLC
Protocol,” we introduce a novel LLC protocol that
accounts for the radio channel status. By means of a
time-out error-recovery algorithm, the ARQ mecha-
nism at the LLC layer attempts retransmission of an
erroneous LLC frame only after a time interval, intro-
ducing a time diversity.13 This channel-dependent
ARQ mechanism14 also increases MS battery lifetime
by deferring transmission over a “bad” channel to a
later time when the channel is likely to be in a “good”
state. The section “LLC Throughput” presents analyti-
cal formulas for both the throughput and the delay per
Application
IP/X.25
SNDCP
LLC
RLC
MAC
GSM RF
MS
Relay
RLC
MAC
GSM RF
BSSGP
Frame relay
L1bis
BSSUm Gb
LLC
SGSN
BSSGP
Frame relay
IP
L2
L1bis L1
SNDCP GTP
IP
L2
L1
GTP
Gn
IP
GiGGSN
BSS – Base station subsystemBSSGP – BSS GPRS protocolGb – BSS/SGSN interfaceGGSN – Gateway GPRS support nodeGi – Interface between two GGSNsGn – SGSN/GGSN interfaceGPRS – General Packet Radio ServiceGSM – Global System for Mobile CommunicationsGTP – GPRS tunnel protocolIP – Internet protocolITU-T – International Telecommunication Union— Telecommunicaton Standardization SectorL1 – Layer 1 for Gn interface
L1bis – Layer 1 for Gb interfaceL2 – Layer 2 for Gn interfaceLLC – Logical link controlMAC – Medium access controlMS – Mobile stationRF – Radio frequencyRLC – Radio link controlSGSN – Serving GPRS support nodeSNDCP – Subnetwork dependent convergence protocolUm – Radio interfaceX.25 – ITU-T standard for network layer 3
Figure 1.GPRS transmission plane.
40 Bell Labs Technical Journal ◆ July–September 1999
LLC frame. These closed-form expressions, together
with measurements, are suitable for on-line GPRS
implementation. Several samples of numerical simula-
tions are presented to illustrate the analytical formulas.
Throughout this work, denotes the BLER,
denotes the LLC frame size in RLC/MAC blocks,
is the maximum number of RLC/MAC block retrans-
missions per LLC frame, and is the maximum num-
ber of LLC retransmissions per LLC frame. Henceforth,
the expression “RLC/MAC retransmission per LLC
frame” denotes the retransmission of several
RLC/MAC blocks belonging to the same LLC frame
that have been reported as being in error by a negative
acknowledgment (NACK). These blocks are selectively
repeated. An LLC retransmission occurs only when
the ARQ mechanism at the RLC/MAC layer fails to
correct all the errors of an LLC frame within the maxi-
mum allowed number of RLC/MAC retransmissions.
ARQ Mechanism at RLC/MACWe assume that a radio channel is described by
the Gilbert-Elliott model, which is a two-state Markov
process.15 The two channel states are denoted by
“bad” (where the BLER is equal to ) and “good”
(where the BLER is zero), respectively. Rigorously
speaking, the BLER in the good state is not zero but is
very small compared with that from the bad state.
Since the dominant degradation of network perfor-
mance comes from errors in the bad state, the effect of
errors in the good state on network performance can
be neglected.
Throughout this analysis, we consider that the
errors over the radio channel in a bad state are inde-
pendent. This is an optimistic assumption because, in
general, the errors over the radio channel occur in
bursts—exhibiting, therefore, some degree of correla-
tion. In a bad state, the errors may affect several
RLC/MAC blocks in sequence. However, the probabil-
PB
L
MN
PB
Panel 2. Mathematical Variables and Expressions
δ (k)—Dirac function—standard deviation of k—standard deviation of —carrier to interference ratio
E [k]—average number of RLC/MAC retransmissionsper LLC frame
E [R]—average number of retransmissions perRLC/MAC block
HLLC—number of header octets in an LLCframe
HRLC/MAC—number of header octets in anRLC/MAC block
k—number of RLC/MAC retransmissions per LLCframe
L—maximum number of LLC retransmissions perLLC frame
LLC timer—timer that is set when an LLC frame issent to the RLC/MAC layer
m—number of NACKs sent per LLC frameM—maximum number of RLC/MAC block
retransmissions per LLC frameM(N)—number of NACKs sent per LLC frame of
size equal to N RLC/MAC blocksN—number of RLC/MAC blocks in an LLC framenLLC—number of information octets in an LLC
frame
nRLC/MAC—number of data-information octets inan RLC/MAC block
PB—non-zero BLERPM—probability that after RLC/MAC retrans-
missions there are still RLC/MAC blocks in errorbelonging to the same LLC frame of size
Pr—probabilityPrc—conditional probabilityPr( ≤ R)—probability that the radio channel is
in a bad stateq—probability of successful transmission of an
LLC frameQ—number of LLC frames queuing for transmis-
sion at the RLC/MAC layerR— threshold, 12 dBThroughputLLC—throughput at the LLC layerTLLC Frame Service Time—average service time of an
LLC frameTLLC RLC/MAC—service time of an LLC frame at the
RLC/MAC layerTLLC Service Time—service time of an LLC frame at
the LLC layerTNACK—time required to receive and process a
negative acknowledgment, 40 msTRLC/MAC Block—transmission duration of an
RLC/MAC block over the radio interface, 18.46 ms
C/I
C/I
N
MC/IC/Is0
s
Bell Labs Technical Journal ◆ July–September 1999 41
ity of a deep Rayleigh fade lasting more than 20 ms
(that is, more than one RLC/MAC block) is very low,
so we neglect the correlation of errors over many
RLC/MAC blocks.
We assume that the BLER is constant over the
transmission of an LLC frame. This is a reasonable
assumption, since the BLER is derived from a lognormal,
slow-fading distribution of the carrier-to-interference
ratio, C/I, for which a slow fade can last several sec-
onds. We proceed now to calculate the following con-
ditional probabilities:
, (1)
conditioned by the radio channel being in a bad state
(that is, by ). The notation refers to
RLC/MAC blocks belonging to the same LLC frame.
Note that, when the channel is in the good state, no
retransmission is required as we have assumed that
the BLER is zero in this state. The probability that the
channel is in a bad state is given in (7).
The probability that all RLC/MAC blocks
of an LLC frame are received error free at the first
transmission is equal to
, (2)
where we assume that the transmissions of all
RLC/MAC blocks are statistically independent events.
To calculate , we note that the probability that
an RLC/MAC block has errors after two retransmis-
sions is ; therefore, the probability that this
RLC/MAC block is successfully received after two
RLC/MAC retransmissions (including the original
transmission) is . Using this and the independence
of the transmissions of the RLC/MAC blocks yields
(3)
where and are the probabilities that
after the second and first RLC/MAC transmissions,
respectively, the LLC frame is received error free.
Thus, is the probability that two RLC/MAC
transmissions are required to send the LLC frame
because after the first RLC/MAC transmission there
are still erroneous RLC/MAC blocks. Following the
above approach, the general form of is
. (4)
From (4) we have
(5)
and
. (6)
It is worth pointing out that the conditional probabil-
ity, , from (4) is a closed-form expression.
Ludwig and Turina12 have also derived , but in
an implicit, iterative form. The equivalence between
the closed-form expression of given in (4) and
that given by Ludwig and Turina is shown in the
appendix. The benefit of deriving a closed-form
expression for is that we can analytically derive
the most likely number of RLC/MAC retransmissions
per LLC frame, , for a given and a given LLC
frame size, as shown shortly.
To remove the condition on the channel state
from , we use the probability that the radio
channel is in a bad state
(7)
to obtain the unconditional probability of the
form
(8a)
and
(8b)
where is the threshold that delimits good and bad
states (12 dB), , and . Here, x
represents and �0 denotes the standard deviation of
. In (8a), the term represents the prob-
ability that an LLC frame is transmitted over a channel
in a good state and that, therefore, no retransmission is
required. Using the tabulated integral16 from (7), we
1-Pr1C@I# R2C@I
C@I
s0 5 7 dBC@I5 16 dB
C@IR
Pr Pr forB Bk R P P kCI
kN
kN
( ) = ≤( ) −( ) − −( )
≥−1 1 21
Pr Pr Pr forB1 1 1 1( ) = ≤( ) −( ) + − ≤( ) =CI
NC
IR P R k
Pr1k2
Pr1C@I# R251
s0" 2p3
R
-̀
expc-1x-C/I22
2s20
ddx
PrC1k2
PBkopt
PrC1k2
PrC1k2
PrC1k2
PrC1k2
a
`
k5 1
311-PkB2N-11-Pk-1B 2N45 1
limPr limC B Bk k
kN
kN
k P P→∞ →∞
−= −( ) − −( )
=( ) 1 1 01
PrC1k25 11-PkB2N-11-Pk-1B 2N
PrC1k2
PrC122
11-PB2N11-P2B2
N
PrC1225 11-P2B2N-11-PB2
N
N
1-P2B
P2B
PrC122
N
PrC1125 11-PB2N
NPrC112
PrC1k2BLER 5 PB
retransmission ; PrC1k2
in error that are corrected by the kth RLC/MAC
retransmission there are RLC/MAC blocks
Probability that after the 1k-12th RLC/MAC
42 Bell Labs Technical Journal ◆ July–September 1999
obtain . From (8), we note that
.
The number of RLC/MAC retransmissions per LLC
frame, , is a random variable with a probability den-
sity function (PDF) equal to where is
the Dirac function and is given in (8). Then, the
average value of is
(9)
and its variance has the form
(10)
where is the maximum number of RLC/MAC
retransmissions per LLC frame defined in (12) and is
the standard deviation of .
Using (4) through (6), we can also define a con-
ditional PDF for , namely, . As we are
interested in the number of RLC/MAC retransmis-
sions when the radio channel is bad, we plot in
Figure 2 this conditional PDF and the corresponding
conditional cumulative distribution function (CDF)
for various LLC frame sizes expressed in RLC/MAC
blocks. These data are for the case where
and .
The average and standard deviation of random
variable can be obtained using (9) and (10), respec-
tively. For , , and , we obtain
and . We can see from Figure 2a that
there is a maximum value of the , which will be
analytically derived next.
The most likely number of RLC/MAC retransmis-
sions per LLC frame—that is, the value of that maxi-
mizes —can readily be derived by differentiating
with respect to in (4) or (8) and setting the obtained
result equal to zero, yielding
(11)
where the square brackets [ ] denote the closest inte-
ger less than or equal to the term inside the brackets.
It can be seen from (11) that the ratio of the two
logarithm functions inside the square brackets is posi-
k
P
P
P
N
N N
opt
B
B
B
ln
ln=
−
−
+
−( )−( )
1
11
1 1
1
k
PrC1k2
k
PrC1k2
s 5 0.9k 5 2.5
M 5 7PB 5 0.2N 5 20
k
M 5 7
PB 5 0.2
PrC1k2d1k2k
k
s
M
E31k-k2245 s2 5 a
M
k5 1
k2Pr1k2
E3k45 k 5 a
M
k5 1
kPr1k2
k
Pr1k2
d1k2Pr1k2d1k2
k
a
`
k5 1
Pr1k25 1
Pr1C@I# 1225 0.3
0.00
0.10
0.20
0.30
0.40
0.50
0.60
1 2 3 4 5 6 7
RLC/MAC retransmissions
0.00
0.20
0.40
0.60
0.80
1.00
1.20
1 2 3 4 5 6 7
RLC/MAC retransmissions
Co
nd
itio
nal
PD
F
Co
nd
itio
nal
CD
F
N = 20N = 50N = 100
(a) (b)
CDF – Cumulative distribution functionLLC – Logical link controlMAC – Medium access controlN – LLC frame size in RLC/MAC blocksPDF – Probability density functionRLC – Radio link control
Figure 2.Statistics of the RLC/MAC retransmissions for and .M 5 7PB 5 0.2
Bell Labs Technical Journal ◆ July–September 1999 43
tive and therefore . From (11), we can derive
two asymptotic results for , one when (that
is, for a large LLC frame) and the other when
(that is, for a small LLC frame). For large LLC frames
we have the limit while for small LLC frame
size the limit becomes , as expected. As an
example, we insert and into (11) to
obtain , which is in agreement with the
numerical results shown in Figure 2a. Thus, the most
likely number of RLC/MAC retransmissions per LLC
frame given in (11) is a function of the size of the LLC
frame, , and the block error rate, . The latter
dependency is further illustrated in Figure 3 for differ-
ent values of . Figure 3 shows that for small the
most likely number of RLC/MAC retransmissions is
not very sensitive to the LLC frame size, while for
larger this sensitivity increases.
Other limiting cases can be derived from (11)
when and , respectively. When goes to
zero in (11) we obtain ; for , we have
. Thus, for either a very bad radio environ-
ment (that is, ) or for a large LLC frame size
(that is, ), becomes very large. In practice,
when is larger than a predefined error-rate thresh-
old,1,5 either a handover is initiated or, if this is not
possible, a radio-status PDU is sent to the LLC layer to
inform it about the bad status of the radio interface.
When the radio channel is bad (that is, when is
large), a good practice is to defer the RLC/MAC trans-
mission to a later time when the radio channel is good
rather than retransmitting erroneous RLC/MAC
blocks, therefore wasting MS battery lifetime.
It can be seen from Figure 2b that the CDF of
converges very quickly to unity. For ,
, and , the CDF is already
equal to 0.8. In this case, 80% of the RLC/MAC blocks
of an LLC frame are received error-free after only
three retransmissions. Therefore, one can use the most
likely value, , of the number of RLC/MAC retrans-
missions per LLC frame as a measure for the maxi-
mum allowed number of RLC/MAC retransmissions,
, per LLC frame. In a GPRS implementation, could
be chosen as
. (12)
The calculation of can be done on line based on
(11), the measured value of , and the LLC frame size
in RLC/MAC block units. It is worth pointing out that
PB
kopt
M 5 kopt
MM
kopt
k 5 kopt 1 1 5 3N 5 20
PB 5 0.2PrC1k2
PB
PB
kN W 1
PB S 1
kopt S `
PB S 1kopt 5 1
PBPB S 0PB S 1
PB
PBPB
PBN
kopt 5 2
N 5 20PB 5 0.2
kopt S 1
kopt S `
N S 1
N W 1kopt
kopt $ 1
012
3456
78
2 22 42 62 82
LLC frame size (N)
k op
t
PB = 0.2PB = 0.5
102
kopt – Most likely number of RLC/MAC retransmissions per LLC frameLLC – Logical link controlMAC – Medium access controlPB – Block error rateRLC – Radio link control
Figure 3.The most likely number of RLC/MAC retransmissions per LLC frame.
44 Bell Labs Technical Journal ◆ July–September 1999
the maximum allowed number of RLC/MAC retrans-
missions per LLC frame given in (12) balances the user
throughput and the use of the radio channel.
Alternatively, one can use the CDF of to derive
as follows. The CDF of the random variable can be
derived from (8) in the form
. (13)
As a system requirement, we assume now that the
probability of successful transmission of an LLC frame
is equal to . With this system requirement, (13)
becomes
(14)
yielding
(15)
where . Both (12) and (15) can
be used to calculate the maximum allowed number of
RLC/MAC retransmissions per LLC frame.
Channel-Dependent LLC ProtocolAn accurate estimation of the optimum number of
RLC/MAC retransmissions per LLC frame ( ) is of
crucial importance in managing network resources.
For example, once has been estimated using either
(12) or (15), the transmission delay and the through-
put per LLC frame can be evaluated as shown below.
This information can then be used by the admission
control unit to define the QoS profile provided by the
network to a customer. At this point, the value of the
LLC timer that guards this LLC peer-to-peer connec-
tion is set.
To ease the readability of the paper, we now
define several notations used henceforth. We denote
by the service time of an LLC frame at the
RLC/MAC layer and by the service time of
an LLC frame at the LLC layer. In addition, we denote
by the total number of LLC frames queuing for trans-
mission at the RLC/MAC layer. Then the LLC timer
(that is, the timer set when an LLC frame is sent to
the RLC/MAC layer) can readily be derived using
from (12) or (15) as follows:
(16)
where
(17)
and
. (18)
Here, denotes the average number of retransmis-
sions per RLC/MAC block; is the trans-
mission duration of an RLC/MAC block over the radio
interface (18.46 ms); is given in (9); and is
given in (12). , which is the time to receive and
process a NACK (40 ms), includes the NACK/ACK
processing delay at both the MS (10 ms) and the base
station subsystem (BSS) (10 ms) as well as the
NACK/ACK transfer delay (20 ms) between the MS
and the BSS. In (16), the term is a positive constant
introduced in order to avoid the fade on the radio
channel at the next LLC retransmission. Thus, for
a time diversity is provided by the LLC time-out
recovery mechanism. In practice, is set by measure-
ments taken on the radio channel. For brevity, we
have included the radio-access delay in the term in (16).
It is worth pointing out that the LLC frame queu-
ing delay at the RLC/MAC layer—that is, the term
in (16)—provides a natural, intrinsic time diversity.
Thus, if an LLC transmission fails, the LLC will retrans-
mit the erroneous LLC frame to the RLC/MAC layer.
This LLC frame should wait in the RLC/MAC queue
until all LLC frames in front of it are served. This
queuing delay introduces an internal time diversity.
However, under a light and bursty traffic load, the
queuing delay at the RLC/MAC layer could be very
small so that there is a need for the additional time
diversity introduced by the term in (16).
The LLC frame queuing delay at the RLC/MAC
queue is explicitly taken into account in (16) by the
presence of the factor . The estimation of is based
on the measured number of LLC frames queuing at
the RLC/MAC queue and therefore is accurate on
average, irrespective of the LLC frame arrival process.
The expression (16) can be computed on line using
t
Q
t
t
t. 0
t
TNACK
ME3k4
TRLC/MAC Block
E3R4
E3R45 Pr3C@I# R411-PB2a
M
k5 1
kPk-1B 1 1-Pr3C@I# R4
TLLC RLC/MAC 5 N #E3R4#TRLC/MAC Block 1 TNACK#E3k4
LLC Timer5 1Q 1 12TLLC RLC/MAC 1 t
M
Q
TLLC Service Time
TLLC RLC/MAC
M
M
Aq R
R
CI
CI
=− + ≤( )
≤( )1 Pr
Pr
MA
P
N
=−( )
+ln
ln B
11
1
CDF1M 25 1-Pr1C@I# R21 Pr1C@I# R211-PMB2N 5 q
q
CDF1M 25 a
M
k5 1
Pr1k2
kM
k
Bell Labs Technical Journal ◆ July–September 1999 45
measurements taken at the RLC/MAC queue.
This type of ARQ mechanism belongs to the fam-
ily of channel-dependent algorithms.13,14 The time
diversity introduced by this ARQ scheme increases MS
battery lifetime by deferring an LLC frame transmis-
sion over a bad radio channel to a later time when the
radio channel is likely to be in a good state. This
reduces the number of LLC retransmissions at the
expense of a slight increase in transfer delay due to the
additional delay introduced by in (16). Henceforth,
we consider that is large enough such that two con-
secutive LLC retransmissions are uncorrelated. Using
this statistical independence of the LLC retransmis-
sions, we next derive the LLC frame error rate.
The probability that after RLC/MAC retransmis-
sions there are still RLC/MAC blocks in error belong-
ing to the same LLC frame of size is equal to
(19)
where the approximation is valid for . Thus,
is the probability of RLC/MAC failure, which leads
to an LLC retransmission of the whole LLC frame;
then the RLC/MAC process is repeated. In other
words, is the LLC frame error rate seen at the LLC
layer. Therefore, the LLC peer-to-peer link has an
error probability of .
We can see from (19) that the reliability of the
LLC peer-to-peer link increases as increases for a
fixed LLC frame size . When , then the LLC
frame error rate decreases exponentially as
increases, as shown in the second part of (19).
As an example, we set , , and
in (19) to obtain . If we now assume
that an MS sends 5 e-mails/hour (1 e-mail =
1 LLC frame = 50 RLC/MAC blocks) and there are
60 GPRS MSs active in one cell, then the average
number of LLC retransmissions per hour is
. Thus, 10% of the LLC
frames are retransmitted in this example. A much
lower number of LLC retransmissions can be achieved
by implementing a channel-dependent RLC/MAC
protocol.13,14 Once the LLC frame error rate has been
derived in (19), we can calculate the delay and the
throughput per LLC frame.
The LLC transfer delay has been calculated by
Ludwig and Turina12 for the case where the
RLC/MAC layer is capable of an infinite number of
retransmissions and where a uniform radio channel is
assumed (that is, is constant). They calculated the
LLC queuing delay by considering a Poisson arrival
process and, therefore, their result is limited.
Numerical simulations for the LLC transfer delay and
throughput have also been presented elsewhere.7-11
Here, we derive the LLC frame service time for a
finite number of RLC/MAC retransmissions per LLC
frame, , and for a radio channel characterized by a
two-state Markov process (that is, it has good and bad
states). The throughput per LLC frame is also derived.
We explicitly account for the headers of the LLC and
RLC/MAC layers.
For clarity, we recall here that the LLC frame size
is equal to RLC/MAC blocks, the maximum number
of RLC/MAC retransmissions per LLC frame is as
given in (12) or (15), the maximum number of LLC
retransmissions is , and the probability that the radio
channel is in a bad state is equal to .
Then, the average LLC frame service time is
(20)
where and are given in (16) and (19).
Equation (20) comes from the statistical independence
of the LLC retransmissions due to the LLC time-
diversity time-out. When the number of LLC retrans-
missions reaches , then the LLC frame is dropped and
higher layers are in charge of recovery (for example,
the transmission-control protocol [TCP] layer).
There is a tradeoff between the number of LLC
retransmissions and the number of RLC/MAC retrans-
missions. As the number of RLC/MAC retransmissions
per LLC frame, , increases, then the probability of
RLC/MAC failure, , decreases; therefore, fewer LLC
retransmissions are necessary per LLC frame, . Thus,
for large and for a radio channel in a bad state, the
radio channel is heavily loaded with RLC/MAC retrans-
missions. This leads to poor radio efficiency and a waste
of MS battery lifetime. On the other hand, a very small
number of RLC/MAC retransmissions, , gives rise toM
M
L
PM
M
L
PMLLC Timer
+ −( ) −
=∑LLC Timer P kPM M
k
k
L
1 1
2
T P LLC Timer tMLLC Frame Service Time = −( ) −( )1
Pr1C@I# R25 0.3
L
M
N
M
PB
LLC retrans5 5#60#PM 5 30
PM 5 0.1PB 5 0.2
N 5 50M 5 3
M
PMB V 1@NN
M
PM
PM
PM
PMB V 1@N
PM 5 Pr1C@I# R231-11-PMB2N4 > N#PMBPr1C@I# R2
N
M
t
t
46 Bell Labs Technical Journal ◆ July–September 1999
many LLC retransmissions (large ), which consequently
leads to heavy network load over the Gb and A-bis inter-
faces. Here Gb refers to the interface between the BSS
and the serving GPRS support node (SGSN), while A-bis
is the interface between the base transceiver station
(BTS) and the base station controller (BSC). This also
increases the LLC frame transfer delay. In counterbal-
ance, the efficiency over the radio interface is increased
due to the inherent time diversity of the LLC retransmis-
sion process (that is, its basis on time-out). We believe
that the expression of given in (12) or (15) strikes a
balance between the number of LLC retransmissions and
the number of RLC/MAC retransmissions.
LLC ThroughputThe throughput seen at the LLC layer can be
defined as follows
.(21)
Equations (20) and (21) enable us to calculate on line
the LLC frame transfer delay and throughput.
We consider that GPRS packet transfer is operated
in the acknowledge mode. Transmission control proto-
col/Internet protocol (TCP/IP) introduces headers of
5 octets each (a 4-to-1 compression algorithm is
applied on a 20-octet TCP/IP header). With reference
to Figure 1, an IP packet is segmented at the subnet-
work dependent convergence protocol (SNDCP) layer,
where a 2-octet header is added to every segment and
the resulting packet is called an SNDCP-PDU
(SN-PDU).4 These SN-PDUs are then passed to the
LLC layer, where they are segmented into LLC frames,
and a 6-octet header is added to each LLC frame. The
LLC frames are passed to the RLC/MAC layer, where
they are again segmented; here, a header of 6 octets is
added to every RLC/MAC block when coding scheme
one (CS-1) is considered. Segmentation at every GPRS
protocol layer is necessary to achieve the residual error
rate required for data transmissions6 (for example, a
residual error rate of for reliability class 2).
Therefore, most of the headers added at any segment
include CRC check bits, among others.
The sizes of the SN-PDU and LLC frames are vari-
able, and only their maximum values are standard-
ized.3,4 In a GPRS implementation, the segmentation
size at any layer is a function of the error rate at the
lower layers, among which the error rate over the
radio interface is of the foremost importance. As
pointed out by Zorzi and Rao,14 a multi-layer protocol
stack should be designed by taking into account the
interactions between the protocol layers and the radio
channel, where most of the errors occur.
In this analysis, the focus is on two protocols—
namely, LLC and RLC/MAC—and we neglect the
overhead due to IP and SNDCP. Therefore, we assume
that the information data field of an IP/SN-PDU packet
is much larger than their headers. Thus, we consider
that SN-PDUs coming to the LLC layer contain only
information bits. The SN-PDUs received at the LLC
layer from the SNDCP layer are segmented into LLC
frames of variable size from 20 to 1500 octets.
We consider that an LLC frame has informa-
tion octets and header octets. An LLC frame
is passed to the RLC/MAC layer, where it is segmented
into RLC/MAC blocks of octets of data
information and header octets for CS-1.
Then, the throughput at the LLC layer is
(22)
with given in (17), where
(23)
and the brackets [ ] denote the closest integer less than
or equal to the term inside the brackets. The
RLC/MAC header, , was already taken into
account in (16), (19), and (21) because the
data-information octets are sent in an RLC/MAC
block, , that also carries the RLC/MAC
header.1
In Figure 4, we plot the throughput at the LLC
layer—as given in (22)—as a function of the LLC
frame size in RLC/MAC blocks—as given in (23)—for
different RLC/MAC block error rates and probabilities.
The simulations in Figure 4 are obtained under the
assumptions that the maximum number of LLC
retransmissions per LLC frame, , is equal to 10 and
that the time shift, , used in the from (16)LLC Timert
L
TRLC/MAC Block
nRLC/MAC
HRLC/MAC
N 5 cnLLC 1 HLLC
nRLC/MACd
TLLC Frame Service Time
ThroughputLLC 5nLLC
TLLC Frame Service Time
HRLC/MAC 5 6
nRLC/MAC 5 20
HLLC 5 6
nLLC
10-6
ThroughputLLC 5information data per LLC frame
TLLC Frame Service Time
M
L
Bell Labs Technical Journal ◆ July–September 1999 47
is equal to 50 ms. For simplicity, we also consider that
there is no LLC frame queuing at the RLC/MAC
layer—that is, in (16). In Figure 4a,
is kept constant and we change .
In Figure 4b, is kept constant and we change
. We see from Figure 4a that, by decreasing
from 0.5 to 0.2, the LLC throughput increases by
almost 2 kb/s. It is worth pointing out that the maxi-
mum achievable throughput for CS-1 is equal to
9.05 kb/s.1,2 Thus, for the case where ,
, and for an average LLC frame size
(about 20 RLC/MAC blocks), we can achieve 77%
bandwidth efficiency at the LLC layer. Notably, this
throughput efficiency is achieved while the MS battery
lifetime is increased because of the time diversity
introduced at the LLC layer.
In Figure 4b, the LLC throughput also increases as
(the probability that the radio channel
is in a bad state) decreases. The jumps of the LLC
throughput in Figure 4b along the vertical dotted lines
come from the variation of the maximum number of
the RLC/MAC retransmissions per LLC frame, —see
(12) or (15)—also plotted on the same figure.
According to (19), as increases, the probability of an
LLC retransmission, , decreases exponentially. Thus,
the LLC peer-to-peer link becomes more reliable and
fewer LLC retransmissions are required at the LLC
layer to recover from errors. Fewer LLC retransmis-
sions lead to shorter LLC service time and, therefore,
higher throughput. The compromise is that the MAC
layer should increase its number of retransmissions per
LLC frame, reducing the bandwidth efficiency over the
radio channel. For a constant value of , the LLC
throughput decreases slightly as increases because
the LLC retransmission probability, , increases as the
LLC frame size, , increases—see (19).
In Figure 4b, we also plot LLC throughput for
the case when the LLC layer does not take into
account the radio channel state—that is, when
for non-channel-dependent LLC
protocol. As we see, the throughput in this case is
much lower than that of the channel-dependent LLC
Pr1C@I# R2; Pr 5 1
N
PM
N
M
PM
M
M
Pr ; Pr1C@I# R2
Pr1C@I# R25 0.3
PB 5 0.2
PB
Pr1C@I# R2
PB 5 0.2
PBPr1C@I# R25 0.3
Q 5 0
LLC – Logical link controlM – Maximum number of RLC/MAC retransmissions per LLC frameMAC – Medium access control
PB – Block error ratePr – Probability that the radio channel is in a “bad” stateRLC – Radio link control
0
1
2
3
4
5
6
7
8
2 22 42 62 82
LLC frame size (N)
0
1
2
3
4
5
6
7
8
2 22 42 62 82LLC frame size (N)
PB = 0.2PB = 0.5
LLC
th
rou
gh
pu
t (k
b/s
)
LLC
th
rou
gh
pu
t (k
b/s
)
Pr = 0.3Pr = 0.6
(a) (b)
Pr = 1M
Figure 4.Throughput at the LLC layer versus LLC frame size.
48 Bell Labs Technical Journal ◆ July–September 1999
protocol. It is worth pointing out that a similar channel-
dependent ARQ mechanism can be implemented at
the RLC/MAC layer that will substantially increase the
radio efficiency for delay-tolerant traffic.
ConclusionsIn this paper, we have analyzed the
LLC-RLC/MAC protocol stack for GPRS. The interac-
tion between the two ARQ mechanisms at the two
protocol layers has been investigated. A channel-
dependent ARQ has been proposed at the LLC layer
that reduces the number of LLC retransmissions per
LLC frame at the expense of slight degradation in the
delay per LLC frame. This channel-aware ARQ mecha-
nism can substantially increase MS battery lifetime.
The proposed channel-dependent ARQ mechanism
implemented at the LLC in this paper can be easily
introduced at the RLC/MAC as well. This will substan-
tially reduce the number of LLC retransmissions, and
may even remove the need for an ARQ at the LLC
layer because the channel-related control is moved
entirely to the RLC/MAC layer.
The tradeoff between the numbers of LLC and
RLC/MAC retransmissions has been addressed. In par-
ticular, it has been shown that, for delay-tolerant traf-
fic, an optimal number of RLC/MAC retransmissions
leads to better radio-channel bandwidth efficiency and
better MS energy management—especially in a bad
radio environment.
At the RLC/MAC layer, a closed-form expression
for the probability of RLC/MAC retransmission per
LLC frame has been derived. This has allowed us to
analytically compute the most likely RLC/MAC
retransmission number per LLC frame. This formula
can be used for on-line estimation of the throughput
and the transfer delay per LLC frame. In this way, QoS
parameters can be more accurately characterized,
improving GPRS network performance.
AppendixIn this appendix, we show that the formulas
derived in a recent paper by Ludwig and Turina12 for
the probability of having negative acknowledgments
per LLC frame of a size equal to RLC/MAC blocks
are identical with those derived here in (2) through (4).
Ludwig and Turina used an iterative method to com-
pute these probabilities as follows:
(A1)
(A2)
where is the number of NACKs sent per LLC
frame and is the RLC/MAC block error rate. In
(A1) and (A2), we denote by the proba-
bility of receiving NACKs for an LLC frame having a
size equal to RLC/MAC blocks.
We prove now that the iterative process given by
(A1) and (A2) leads to a closed-form expression for
the probability , namely,
(A3)
where is given in (4). We use the method
of induction to prove (A3), knowing that (A1) and
(A2) are true for any integer . Setting in
(A3) yields
, (A4)
which is identical with (A1) and therefore true.
Following the induction method, we assume now that
given by (A3) is true, and we prove that
.(A5)
We rewrite (A2) for in the form
, (A6)=
−( ) ( ) =[ ]−
=∑ N
kP P M k mk
N k
k
N
B B Pr11
Pr M N m( ) = +[ ]1
1m 1 12S m
Pr B BM N m P PmN
mN
( ) = +[ ] = −( ) − −( )+ +1 1 12 1
Pr3M 1N25 m 4
Pr3M 1N25 045 11-PB2N
m 5 0m $ 0
PrC1m 1 12
Pr Pr ( )B B CM N m P P mmN
mN
( ) =[ ] = −( ) − −( ) ≡ ++1 1 11
Pr3M 1N25 m 4
k
m
Pr3M 1k25 m 4
PB
M 1N2
Pr PrB BM N mN
KP P M k mk
N k
k
N
( ) =[ ] =
−( ) ( ) = −[ ]−
=∑ 1 1
1
Pr BM N PN( ) =[ ] = −( )0 1
N
m
Bell Labs Technical Journal ◆ July–September 1999 49
which, by using (A3) for , yields
. (A7)
We can rewrite in the form
, (A8)
which, when inserted back into (A7), yields
. (A9)
After some rearrangements, (A9) becomes
(A10)
and, finally,
(Q.E.D.). (A11)
References1. European Telecommunications Standards
Institute, “Digital Cellular TelecommunicationsSystem (Phase 2+); General Packet RadioService (GPRS); Mobile Station (MS) – BaseStation System (BSS) Interface; Radio LinkControl/Medium Access Control (RLC/MAC)Protocol,” EN 301 349 (GSM 04.60), Ver. 6.3.0,May 1999, http://www.etsi.org
2. European Telecommunications StandardsInstitute, “Digital Cellular TelecommunicationsSystem (Phase 2+); General Packet RadioService (GPRS); Overall Description of theGPRS Radio Interface; Stage 2,” TS 101 350(GSM 03.64), Ver. 6.1.0, Nov. 1998,http://www.etsi.org
3. European Telecommunications Standards Insti-tute, “Digital Cellular Telecommunications Sys-tem (Phase 2+); General Packet Radio Service(GPRS); Mobile Station (MS) – Serving GPRSSupport Node (MS-SGSN) Logical Link Control(LLC) Layer Specification,” TS 101 351 (GSM 04.64),Ver. 6.3.0, April 1999, http://www.etsi.org
4. European Telecommunications StandardsInstitute, “Digital Cellular TelecommunicationsSystem (Phase 2+); General Packet RadioService (GPRS); Mobile Station (MS) – ServingGPRS Support Node (MS-SGSN); SubnetworkDependent Convergence Protocol (SNDCP),”TS 101 297 (GSM 04.65), Ver. 6.3.0,April 1999, http://www.etsi.org
5. European Telecommunications StandardsInstitute, “Digital Cellular TelecommunicationsSystem (Phase 2+); Channel Coding,”EN 300 909 (GSM 05.03), Ver. 6.2.0, May 1999,http://www.etsi.org
6. European Telecommunications StandardsInstitute, “Digital Cellular TelecommunicationsSystem (Phase 2+); General Packet RadioService (GPRS); Service Description; Stage 1,”EN 301 113 (GSM 02.60), Ver. 6.2.0, May 1999,http://www.etsi.org
7. J. Cai and D. J. Goodman, “General PacketRadio Service in GSM,” IEEE Commun. Mag.,Vol. 35, No. 10, Oct. 1997, pp. 122–131.
8. G. Brasche and B. Walke, “Concepts, Services,and Protocols of the New GSM Phase 2+General Packet Radio Service,” IEEE Commun.Mag., Vol. 35, No. 8, Aug. 1997, pp. 94–104.
9. D. Turina, P. Beming, E. Schoster andA. Andersson, “A Proposal for Multi-Slot MACLayer Operation for Packet Data Channel inGSM,” Proc. of ICUPC—5th Internat. Conf. onUniversal Personal Commun., Vol. 2, IEEE,Oct. 1996, pp. 572–576.
10. G. Brasche, “A Frame Mode Bearer ServiceProposed for the Cellular Radio Network GSM,”Proc. of 46th Vehicular Technology Conf., Vol. 1,IEEE, April–May 1996, pp. 512–516.
11. S. Hoff, M. Meyer, and A. Schieder, “APerformance Evaluation of Internet Access viathe General Packet Radio Service of GSM,”VTC-98, Proc. of 48th Vehicular Technology Conf.,Vol. 3, IEEE, May 1998, pp. 1760–1764.
12. R. Ludwig and D. Turina, “Link Layer Analysisof the General Packet Radio Service for GSM,”Proc. of ICUPC—6th Internat. Conf. on UniversalPersonal Commun., Vol. 2, IEEE, Oct. 1997,pp. 525–530.
13. M. Zorzi, R. R. Rao, and L. B. Milstein, “ARQError Control for Fading Mobile Channels,”
≡ +PrC( )m 2
Pr B BM N m P PmN
mN
( ) = +[ ] = −( ) − −( )+ +1 1 12 1
− + + + +( )
1 2P P PmN
B B B...
= −( ) + + + +( )
+1 1 2 1P P P PB
Nm
N
B B B...
Pr M N m( ) = +[ ]1
−
+ + +( )
=∑ N
kP P P M
k
k
N
B B B2
0
...
= −( )
+ + +( )
+
=∑1 2 1
0
PN
kP P P
Nm
k
k
N
B B B B...
Pr M N m( ) = +[ ]1
11-PmB2k 5 11-PB2
k 11 1 PB 1 P2B 1 ... 1 Pm-1B 2k
11-PmB2k
=
−( ) −( ) − −( )
−+
=∑ N
kP P P Pk
N km
km
k
k
N
B B B B1 1 11
1
Pr M N m( ) = +[ ]1
Pr3M 1k25 m 4
50 Bell Labs Technical Journal ◆ July–September 1999
IEEE Trans. on Vehicular Technology, Vol. 46,No. 2, May 1997, pp. 445–455.
14. M. Zorzi and R. R. Rao, “The Role of ErrorCorrelations in the Design of Protocols forPacket Switched Services,” 35th Ann. AllertonConf. on Commun., Control, and Computing,Univ. of Illinois, Urbana, Sept.–Oct. 1997,pp. 1–9.
15. E. N. Gilbert, “Capacity of a Burst-NoiseChannel,” Bell System Tech. J., Vol. 39,Sept. 1960, pp. 1253–1266.
16. U.S. National Bureau of Standards, “Handbook ofMathematical Functions with Formulas, Graphs, andMathematical Tables,” 9th ed., edited byM. Abramowitz and I. A. Stegun, DoverPublications, New York, 1972.
(Manuscript approved October 1999)
CRISTIAN DEMETRESCU is a member of technical staffin the Global Wireless Systems ResearchDepartment of Bell Labs in Swindon,Wiltshire, United Kingdom. Prior to joiningBell Labs, he received a B.Sc. in electricalengineering from the University of Craiova
in Romania and a Ph.D. in electrical engineering fromthe University of Birmingham in the United Kingdom.He is currently working on real-time services (such asvoice) over the Enhanced General Packet Radio Service(EGPRS) and is team coordinator for packet voice overGPRS prototyping. ◆