IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 53, NO.

5, SEPTEMBER 2004

1531

Performance Analysis of RLC/MAC and LLC

Layers in a GPRS Protocol Stack
K. Premkumar and A. Chockalingam, Senior Member, IEEE

AbstractIn this paper, we analyze the performance of various

layers of the general packet radio service (GPRS) protocol stack,
including radio link control/medium-access control (RLC/MAC)
layer and logical link-control (LLC) layer on the uplink. In the
GPRS MAC protocol, several time-slotted uplink radio-frequency
channels are shared by the mobiles on a request-reservation-based
multiple-access scheme. Using the theory of Markov chains, we derive expressions for the average throughput and delay performance
of the GPRS MAC protocol. We evaluate the performance of the
RLC layer (in acknowledged mode) using block-level retransmission (BLR), as defined in the current GPRS standard, and compare
it with that of using slot-level retransmission (SLR). We show that
SLR at the RLC layer performs significantly better than the BLR,
particularly when the channel-error rates are moderate to high.
We further investigate the choice of parameters (e.g., number of
retransmission attempts) for the automatic repeat request schemes
at the RLC and LLC layers. Our results show that it is more beneficial to do error recovery by allowing more retransmission attempts
at the RLC layer than at the LLC layer. We also evaluate the performance of transmission-control protocol with BLR and SLR at
the RLC layer.
Index TermsAutomatic repeat request (ARQ), block-level
retransmission (BLR), general packet radio service (GPRS), logical link control (LLC), radio link control/medium-access control
(RLC/MAC), slot-level retransmission (SLR), transmission-control protocol (TCP).

I. INTRODUCTION

HE first- and second-generation (1G and 2G) cellular systems, including advanced mobile phone service (AMPS),
global system for mobile communications (GSM), IS-136, and
IS-95, are circuit-switched (CS) systems designed primarily
for voice communications. In addition to voice services, CS
data communications at low rates (maximum 14.4 kb/s) are
supported by 1G and 2G systems [e.g., cellular digital packet
data (CDPD) on AMPS and digital fax and asynchronous
data transfer on GSM, IS-95]. With the increasing demand for
wireless mobile Internet access at high speeds, the evolution
of 2G systems to next-generation systems has been focused on
supporting packet-switched operation over the air in order
to make more efficient use of radio resources. For example,
general packet radio services (GPRS) and enhanced data rates

Manuscript received May 14, 2003; revised October 6, 2003, and April
20, 2004. This work was supported in part by the Department of Science and
Technology, Swarnajayanti Fellowship, New Delhi, Government of India,
under Project Ref 6/3/2002-S.F. This work was presented in part at the National
Conference on Communications (NCC 2001), Indian Institute of Technology,
Kanpur, India, January 2001.
The authors are with the Department of Electrical Communication Engineering, Indian Institute of Science, Bangalore 560 012, India (e-mail:
[email protected]; [email protected]).
Digital Object Identifier 10.1109/TVT.2004.832396

for GSM evolution (EDGE) are systems that upgrade (with

packet-routing nodes) and use the existing GSM infrastructure
to provide high-speed (maximum 270 kb/s in GPRS and 384
kb/s in EDGE) packet-mode data communications [1][3]. The
design and performance of protocol stacks in such wireless
systems are largely influenced by the erroneous nature of the
wireless channel (due to distance losses, shadowing, and multipath fading), as well as the behavior of the commonly used
higher layer protocols that are designed primarily for wireline
channels that exhibit very low error rates. Transmission-control
protocol (TCP), a widely used transport-layer protocol, is
one such higher layer protocol [4]. It is well known that TCP
performs poorly on wireless channels in which error rates are
typically high [5][7]. This is because the channel-induced
errors often trigger the congestion control mechanism in TCP
in such a way that the transmission window size is reduced,
thus significantly degrading the TCP throughput. Hence, the
lower layers in wireless protocol stacks [e.g., link layer (LL)
and media access-control layer (MAC)] must be designed to
address these issues of high error rates and higher layer-performance concerns.
Forward error correction (FEC) and automatic repeat request
(ARQ) are commonly used error-control techniques on wireless
channels. In FEC, redundant bits are added to information bits
to detect and correct channel-induced errors. In ARQ, on the
other hand, error control is achieved not through error correction, but through the retransmission of erroneous data packets.
While FEC is applied at the physical layer, ARQ can be applied
at different layers of the protocol stack. Errors uncorrectable by
FEC at the physical layer can be handled by the higher layer
ARQs. For example, the protocol stack in GPRS consists of
ARQs at the radio link-control (RLC) and logical link-control
(LLC) layers, in addition to employing FEC at the physical layer
[2]. Our focus in this paper is the performance analysis of some
of the key layers of the GPRS protocol stack.
Several studies have investigated the performance of the various layers of the GPRS protocol stack, but mainly through simulations [8][13]. One of our key contributions in this paper is
the analytical approach to performance evaluation of some of
the key GPRS protocol layers, including MAC and RLC layers
on the uplink. In the GPRS MAC protocol, several time-slotted
uplink radio-frequency channels are shared by the mobiles on
a request-reservation-based multiple-access scheme. Using the
theory of Markov chains, we derive expressions for the average
throughput and delay performance of the GPRS MAC protocol
[14]. We also evaluate the performance of the RLC layer (in
acknowledged mode) using block-level retransmission (BLR),
as defined in the current GPRS standard, and compare it with

0018-9545/04$20.00 2004 IEEE

1532

Fig. 1.

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 53, NO. 5, SEPTEMBER 2004

Network-layer PDU segmentation into LLC frames, RLC blocks, and MAC bursts.

that of using slot-level retransmission (SLR). We show that SLR

at the RLC layer performs better than the BLR, particularly
when the channel-error rates are high. We further investigate the
choice of parameters (e.g., number of retransmission attempts)
for the ARQ schemes at the RLC and the LLC layers. Our results
show that it is more beneficial to do error recovery by allowing
more retransmissions at the RLC layer than at the LLC layer.
We also evaluate the performance of TCP with BLR and SLR
at the RLC layer.
The rest of this paper is organized as follows. In Section II, we
present the functionalities of the GPRS LLC, RLC, and MAC
layers that are relevant to our performance analysis. The performance analyses of the GPRS MAC and RLC layers are presented in Section III. The performance of BLR and SLR at the
RLC layer are compared. LLC and TCP performance results
with BLR and SLR at the RLC layer are presented in Section IV
and the conclusion is presented in Section V.
II. GPRS LLC/RLC/MAC LAYERS
GPRS networks provide Internet protocol (IP) connectivity
to mobile users through GSM infrastructure using additional
support nodes for packet routing, namely, serving GPRS support node (SGSN) and gateway GPRS support node (GGSN),
and the associated protocol stacks. Refer to [1] and [2] for an
introduction to GPRS architecture, concepts, system operation, and protocol functions. The over-the-air communication
between the mobile station (MS) and the GPRS network is
defined by the physical- and data-link-layer functionalities. The
physical layer functions include modulation, demodulation,
channel coding/decoding, etc. The data-link layer consists
of two sublayers, namely, logical link control (LLC) and the
radio link control/medium-access control (RLC/MAC). In this
section, we present the functionalities of the LLC, RLC, and
MAC layers that are relevant to our performance analysis.
LLC Layer: The LLC layer operates between the MS and the
SGSN and provides a logical link between them [15]. Packet
data units (PDUs) from higher layers (IP layer) are segmented

into variable-size LLC frames (see Fig. 1). The functions of the
LLC layer includes link-level flow control and ciphering. The
LLC layer can operate either in an acknowledged mode or in an
unacknowledged mode. In the unacknowledged mode of operation, the LLC layer does not attempt the recovery of erroneous
LLC frames. LLC frames, erroneously received or otherwise,
are passed on to the higher layers. In the acknowledged mode,
the LLC layer provides an ARQ mechanism to retransmit erroneous LLC frames. A frame-check sequence (FCS) is provided
in each LLC frame to detect LLC frame errors. A retransmission count variable N200 is defined [15]. The LLC is reset and
error recovery is passed on to higher layers (e.g., TCP) if LLC
frames errors could not be recovered within N200 retransmission attempts.
RLC Layer: The RLC layer is provided below the LLC layer
and above the MAC layer [16]. The RLC peers are at the MS and
the base station system (BSS). On the transmit side, the RLC
layer segments each LLC frame into several RLC data blocks
(see Fig. 1). Each RLC data block occupies four time slots, irrespective of the type of channel coding used. Coding schemes
CS-1, CS-2, CS-3, and CS-4 are defined with rate-1/2, rate-2/3,
rate-3/4, and rate-1 (i.e., no coding), respectively [16]. In the
case of coding scheme CS-1, each RLC block consists of 181 information bits, 40 block-check-sequence (BCS) bits, and seven
tail/control bits. With single-slot operation [i.e., only one slot
per GSM time-division multiple-access (TDMA) frame is allotted to a user], the information rate of coding scheme CS-1 is
9.05 kb/s (i.e., 181 b in four TDMA frames, where each TDMA
frame occupies 4.615 ms). Similarly, the maximum information rates possible using other coding schemes are 13.4 kb/s for
CS-2, 15.6 kb/s for CS-3, and 21.4 kb/s for CS-4 (see Fig. 2).
With multislot operation (i.e., allocation of up to eight slots in a
TDMA frame to a user), these maximum possible information
rates are increased eight-fold.
Like the LLC, the RLC too can operate either in an acknowledged or unacknowledged mode. In the acknowledged mode,
the RLC provides a selective repeat ARQ mechanism to recover erroneous RLC data blocks. A BCS is provided in each

PREMKUMAR AND CHOCKALINGAM: PERFORMANCE ANALYSIS OF RLC/MAC AND LLC LAYERS

1533

Fig. 2. RLC/MAC block structure and coding.

RLC data block to enable error detection. A RLC retransmission

counter N3104, which keeps track of the number of times a RLC
block is retransmitted in the case of error, is defined [16]. The
RLC layer is allowed to attempt a maximum of N3104_MAX
retransmissions to recover blocks in error. If an erroneous block
is not recovered within N3104_MAX retransmission attempts,
then control is passed on to LLC to recover the error at the LLC
frame level. In the RLC unacknowledged mode, there is no retransmission of erroneous RLC data blocks. On the receive side,
the RLC performs the reassembly of LLC frames.
MAC Layer: The GPRS MAC protocol operates on a
slotted-ALOHA-based reservation protocol [16]. The MAC
layer peers are at the MS and BSS. The MAC layer requests/reserves resources in terms of the number of traffic data slots. The
MAC function provides arbitration between multiple mobiles
attempting to transmit simultaneously and provides collision
detection and recovery procedures.
The packet random-access channel (PRACH) is used by all
the mobiles, on a contention basis, for the purpose of sending
resource-request packets. Typically, TS0 slot in a GSM frame
of eight slots can be used as PRACH. All mobiles are allowed to
transmit on PRACH slots, following the slotted-ALOHA protocol. Depending on the system load, the number of PRACHs
can be increased. The packet data traffic channels (PDTCH),
on the other hand, are used for the transfer of traffic data
packets. Resource requests are made by the mobiles in terms
of the number of uplink PDTCH slots required. Based on these
requests, PDTCH slots are dynamically assigned to the mobiles
by the base station (BS). Allocation can be done based on
either single-slot or multislot operation. If only one slot per
TDMA frame is assigned to a user, then it is called single-slot
operation, whereas if more than one slot (up to eight slots) per
TDMA frame is assigned to a user, then it is called multislot
operation.

When the MAC at the mobile side receives RLC data blocks
to be transferred to the BS, it sends a request packet on the immediately following PRACH slot. The request packet indicates
, the number of PDTCH slots required. If the BS receives the
request packet without collision or channel errors and if PDTCH
slots are available to honor the request, the BS informs the reservation information to the mobile on the downlink packet-access-grant channel (PAGCH) channel. The reservation information include the PDTCH frequency-time slots that can be used
by the mobile for data transfer. The mobile then sends data in
those reserved slots. On the other hand, if the request packet
is lost (due to collision or channel errors) or if PDTCH slots are
not available, then the mobile will not get the reservation. The
mobile will then reschedule its request packet-retransmission attempt to a later time (typically, after a random backoff time).
The MAC control parameters include MAX_RETRANS,
PERSISTENCE_LEVEL, as defined in [16]. The MAC layer
can send channel requests on the PRACH slots up to a maxretransmission attempts in
imum of MAX RETRANS
the event of loss due to collision or channel errors. The delay
between retransmission attempts is defined by the PERSISTENCE_LEVEL.
Thus, in summary, in terms of error recovery at different
layers: 1) the MAC layer attempts to resolve collision of request packets; 2) RLC layer attempts to recover RLC data block
errors through a selective repeat ARQ mechanism; and 3) LLC
attempts recovery of erroneous LLC frames through another
ARQ mechanism. Link errors unresolved at the LLC layer are
passed on to higher layers (e.g., transport layer) to resolve.
III. ANALYSIS OF GPRS MAC PROTOCOL
In this section, we present a Markov-chain-based analysis to
evaluate the throughput and delay performance of the GPRS

1534

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 53, NO. 5, SEPTEMBER 2004

Fig. 3. GPRS MAC protocol operation.

Fig. 4.

MS state transition diagram.

MAC protocol. Consider a single-cell GPRS system with ,

uplink channels and mobile users. Each channel corresponds to a frequency-time slot pair in the MS-to-BS direcchannels, ,
, channels are used
tion. Out of
channels are used as
as PRACHs and the remaining
PDTCHs. Typically, slot TS0 in all GSM TDMA frames on a
given frequency can form a PRACH. Likewise, on a given frequency, slot TS1 in all GSM TDMA frames can form PDTCH-1,
slot TS2 can form PDTCH-2, and so on. We consider a singleslot operation in which only one slot per GSM TDMA frame
is assigned to a user. For example, TS1 slots in consecutive
being assigned to a moTDMA frames
bile for data transfer is a typical illustration of the single-slot
operation. Considering single-slot operation, all uplink chan-

nels can be modeled as synchronized slotted channels, as shown

in Fig. 3. One request packet is one slot in size. One network
(NW) layer PDU, including LLC/RLC headers and checksums,
occupies several slots. Between the successful transmission of
a request packet on a PRACH slot and the corresponding data
transmission on the assigned PDTCH slots, some finite time is
elapsed because of the propagation and processing delays involved. We are interested in analyzing the throughput and delay
performance of the GPRS MAC protocol. In order to carry out
the performance analysis, we assume the following.
1) NW layer PDU arrival process (hence, the new request
packet generation process) is Bernoulli with arrival probability in each slot. The number of LLC frames per NW layer
PDU is assumed to be one. A new NW layer PDU is accepted

PREMKUMAR AND CHOCKALINGAM: PERFORMANCE ANALYSIS OF RLC/MAC AND LLC LAYERS

1535

only after the completion of the transfer of the previously accepted NW layer PDU.
2) Length of the NW layer PDU (including LLC/RLC
headers and checksums), measured in number of slots, is
.
geometrically distributed with parameter ,
3) Loss of request packets on PRACH is only due to collision.
4) Retransmission attempts of request packets following a
collision on PRACH (or nonavailability of PDTCH) are delayed by geometrically distributed random delays with pa. This parameter essentially models
rameter ,
the backoff delay (in the event of request packet loss), which
is characterized by the standards-defined parameter PERSISTENCE_LEVEL.
5) Propagation and processing delays are assumed to be negligible. This assumption can be valid in our considered system
of single-slot operation in which the response from the BS can
come within one TDMA frame time itself.
As per the GPRS MAC protocol and the system model described above, the mobile can be in any one of the following
states in a slot, namely, the idle, backlogged, data_Tx_success,
or data_Tx_failure states, as shown in Fig. 4. In the idle state,
a mobile remains idle with probability
and generates
a PDU with probability . If different mobiles send their request packets on the same PRACH slot, then one or more request packets can be received correctly, depending on the capture probability. The mobiles who lost their request packets go to
the backlogged state. If a mobiles request packet is successfully
received, but there are no available PDTCHs to serve the request, then also the mobile goes to the backlogged state. On the
other hand, if there are available PDTCHs to serve a successfully
received request, then the mobile goes to the data_Tx states,
where it sends data on the assigned PDTCH slots. The data_Tx
states include the data_Tx_success and data_Tx_failure states,
corresponding to the transmission on a PDTCH slot being, respectively, a success or failure.
A. Throughput Analysis
Let
represent the number of mobiles
be the
in the data_Tx_failure state,
number of mobiles in the data_Tx_success state, and
be the number of mobiles in the backlogged state,
at the beginning of slot . The three-dimensional (3-D) process
defines the state of
the system at time . Now, let us analyze the transition of the
system from
in slot to
in slot
,
,
,
,
,
,
. Refer to Fig. 5, which illustrates the
state transitions from slot to slot
.
out of mobiles
During slot , mobiles
in the data_Tx_failure state have data to transmit and the remobiles end the data transmission and go to
maining
,
the idle state. Out of those mobiles, mobiles,
mogo to the data_Tx_success state and the remaining
biles stay in the data_Tx_failure state.

Fig. 5.

State transitions of a mobile from slot t to slot t

+ 1.

Also,
mobiles
out of
mobiles in the
data_Tx_success state have data to transmit and the remaining
mobiles end data transmission and go to the idle state.
Out of those
mobiles,
mobiles
go to the
mobiles stay
data_Tx_failure state and the remaining
in the data_Tx_success state.
mobiles in the backlogged state,
moOut of
biles attempt request packet transmission and the remaining
mobiles stay in the backlogged state. Also,
idle mobiles send request packets giving a total of requests,
succeeds. The remaining
out of which ,
mobiles go to the backlogged state. Out of
succeeding
mobiles go to the data_Tx_success state and the
mobiles,
go to the data_Tx_failure state.
remaining
), the
So, at the end of slot (or at the beginning of slot
total number of mobiles in the data_Tx_failure state is
, in the data_Tx_success state it is
,
. Assuming that slot
and in the backlogged state is
errors can occur independently from slot to slot and denoting
to be the probability of slot error,1 transition probability
1It is noted that the slot-error rate P depends on the channel signal-to-noise
ratio (SNR) and the modulation and coding used. The modulation and coding
schemes can change dynamically in a practical GPRS system depending on the
channel conditions. The mapping of the channel SNR to the slot-error rate is
beyond the scope of this paper. References [10] and [13] give such mapping
between C/I to block error rate for various coding schemes.

1536

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 53, NO. 5, SEPTEMBER 2004

can be written as shown in (1) at the bottom of the page where

, is the number of requests sent, is the
,
number of successful requests in slot ,
,
,
, and
Prob ( successes given that requests are sent and that
PRACHs are available), which can be written as

cap

(2)

is the probability of capturing out of colwhere cap

liding request packets. Equation (2) is terminated by fixing
cap
. It is noted that the capture probability
in the above can be obtained either by its analytical evaluation for a given channel model (for example, as derived in
[17] and [18] for Rayleigh- and Rician-fading channels with
log-normal shadowing) or by using empirical capture models
(for example, as given by the European Telecommunications
Standards Institute (ETSI) in [19]), which has been used in
several papers (e.g., [8]). In this paper, we use the empirical
capture model given in the ETSI standard [19]. As in [19],
, cap
, cap
,
we set cap
cap
, cap
, cap
,
for
, and cap
cap
.
cap
The above analysis assumes a large number of PDTCHs,
so that a successful request always gets an assignment. If the
number of PDTCHs is small compared to the number of users
, (i.e.,
), then a successful request may not get
an assignment, as all the PDTCHs may be busy. This event
and
occurs if
. This implies that
. If
the total number of idle channels
requests succeed in any time slot and the number of idle

is less than , then all requests are

channels
backlogged. For those and satisfying the above condition,
and add to the corresponding
compute the probability
term calculated using (1). To do that, let represent the
.
number of idle channels that is given by
represents the number of data transmissions ended during slot
, which is given by
, and
represents the new arrivals that are backlogged due to nonavailfor the transition
ability of channels. The probability
to
is then given by
from

(3)
where
Thus

(4)
The Markov chain
has a finite number
of states and is positive recurrent [20]. Hence, it has a stationary
steady-state distribution and is found by solving
(5)

new arrivals
request packet transmissions from
out of continue to tx data
out of continue to tx data
out of requests succeed and of

backlogged mobiles
out of go to success
out of go to failure
go to success

(1)

PREMKUMAR AND CHOCKALINGAM: PERFORMANCE ANALYSIS OF RLC/MAC AND LLC LAYERS

where

,
,
is the steady-state probability vector. The
average system throughput is obtained as

1537

the total number of slots required to successfully transmit a PDU

of length slots, is given by

(11)
(6)
The average per-channel throughput

The distribution of
by

is then given by
(7)

B. Delay Analysis
Next, we derive the mean PDU transfer delay performance.
is defined as the average
The mean PDU transfer delay
number of slots elapsed from the slot where a PDU arrived to
the slot where the PDU transmission is complete. The number
of users in the nonidle state (i.e., data_Tx_failure, data_Tx_success, and backlogged states) contribute to the mean delay. There
nonidle users in the system and averaging
are
it over steady-state distribution gives us

is geometric (by assumption) and is given

(12)
[ s being independent and identiand the distribution of
cally distributed (i.i.d.), since we consider i.i.d. slot errors] is
given by
(13)
is the slot-error rate. The distribution of in (11) can
where
be evaluated using Transform techniques, as follows:

(8)
idle users and each will generate requests
There are
with probability in each slot. The average arrival rate to the
system is given by
(9)
From Littles theorem, the average time an user spends in the
system is given by the ratio between the number of users in the
system to the average arrival rate. Hence
(10)
Note that the one in (10) is added to ensure that there is one slot
delay for the mobiles to enter into the nonidle state.
(14)
C. RLC LayerAcknowledged Mode
Note that the analysis in Sections III-A and B corresponds to
the RLC/MAC protocol operation with RLC in the unacknowledged mode (i.e., there is no ARQ at the RLC). In the acknowledged mode, however, RLC retransmits erroneous data blocks
using a selective repeat ARQ mechanism. Each RLC data block
consists of four slots. In our RLC layer analysis here, we consider a SLR mechanism by which a slot in error is repeatedly
retransmitted until it succeeds. This implies that the number of
retransmission attempts at the RLC layer is infinity. We derive
the throughput-delay performance of this SLR scheme as follows.
Let the random variable represent the length of PDU in
number of slots and the random variable represent the number
of transmission attempts of the th slot until success. Thus, ,

Therefore

(15)
Now, in order to obtain the average throughput and mean delay
for the RLC (acknowledged mode) with SLR, we need to simply
in (1) and (3).
change the parameter to
Note that the above analysis is for infinite number of retransmission attempts at the RLC with SLR. The effect of finite number of retransmission attempts at the RLC with SLR is
evaluated through simulations. Likewise, the performance using
BLR at RLC is also evaluated through simulations.

1538

Fig. 6.

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 53, NO. 5, SEPTEMBER 2004

RLC/MAC average per channel throughput  versus new request arrival probability . N

D. RLC/MAC Performance Results

In this section, we present the results and discussions of the
GPRS RLC/MAC performance obtained through both of the
analyses above, as well as simulations. In Fig. 6, numerical results for the average per-channel throughput of the RLC/MAC
,
,
,
protocol, obtained from (7), for
, and
are plotted as a function of the new
request arrival probability . It is noted that the value of 0.1
implies that the average backoff delay after collision is ten slots.
value of 0.1 implies that the average length
Likewise, the
of the PDU measured in number of slots is 10. According to
the throughput computation in (7), the PRACH slots do not
contribute to the effective throughput. In other words, for
channels out of which channels are PRACH, the maximum
. The effect of slot errors
capacity is given by
with/without RLC SLR is also plotted for a slot-error rate of
10% (corresponding to poor channel conditions). These results
are compared with the corresponding RLC BLR performance
obtained through simulation.
It is noted that the average per-channel throughput in Fig. 6
gives the average number of successful slots and the average
number of successful blocks for slot level and block retransmissions, respectively. In order to convert this normalized
throughput to an equivalent data rate in kilobits/second, the
number of information and control bits [including BCS and
slot-check sequence (SCS) bits] defined in the various coding
schemes CS-1, CS-2, CS-3, and CS-4 needs to be taken
into account. For example, when CS-1 is used, a normalized
per-channel throughput of, say, 0.45 corresponds to an effective
data rate of 4.07 kb/s (i.e., 0.45 9.05 kb/s) for block-level
retransmission. For SLR, assuming 10 b of SCS in each slot,

= 10, M = 10, L = 1, g = 0:1, and g = 0:1.

TABLE I
EFFECTIVE PER-CHANNEL THROUGHPUT (IN KILOBITS PER SECOND) FOR BLR
,M
,L
,
VERSUS SLR FOR VARIOUS CODING SCHEMES. N
g
g
: ,
, AND P
. (SEE FIG. 6)

= =01 =1

= 10
= 10%

= 10 = 1

the effective data rate becomes 3.9 kb/s (i.e., 0.45 8.7 kb/s).
Note that the maximum information rate of 8.7 kb/s in the
above is obtained as follows: Each slot carries 114 (rate 1/2)
coded bits, so that 40 information bits, ten SCS bits, and seven
control bits constitute one slot before coding. Forty information
bits sent every 4.615 ms results in an effective information rate
of 8.7 kb/s. The effective data rates for other coding schemes
CS-2, CS-3, and CS-4 can be obtained likewise. It is noted that
when the slot-error rate is zero (or very low values of ) BLR
performs slightly better than SLR (i.e., maximum information
rate of 9.05 versus 8.7 kb/s), which is due to the increased
overheads due to the tail/control bits in SLR. However, as we
will see subsequently, SLR performs much better than BLR
when channel conditions are poor (moderate to high values of
).
From Fig. 6 we observe that the throughput results obtained
from (7) closely match the results obtained through simulation
(in fact, since there are no approximations involved in the analysis, this close match is expected). It is observed that for low
values of new request arrival probability the system spends
most time in the idle state, resulting in very low throughputs.
, the average time duration between arFor the case of
and during this time interval the average number of
rivals is

PREMKUMAR AND CHOCKALINGAM: PERFORMANCE ANALYSIS OF RLC/MAC AND LLC LAYERS

Fig. 7.

1539

Mean PDU transfer delay (in number of slots) at RLC/MAC versus new request arrival probability . N = 10, M = 10, L = 1, g = 0:1, and g = 0:1.

successful slots are given by

throughput is given by

. Thus, the average per channel

For
,
,
,
, and
,
is
0.09, which is very close to the value shown in Fig. 6. As increases, the fraction of time the system spends in the idle state
decreases, which results in increased throughput. When there
), all the slots carrying data
are no slot errors (i.e.,
traffic are successful, which represents the best possible perforthe per-channel throughput
mance. For example, for
achieved is 0.09 (i.e., an effective data rate of 666 b/s using
coding scheme CS-1) and for values closer to unity the perchannel throughput increases to 0.54 (4 kb/s using CS-1).
It is noted that the No Retx plot in Fig. 6 corresponds to
GPRS MAC with RLC in the unacknowledged mode of oper), when there
ation. For high slot-error rates (say,
are no RLC retransmissions, the fraction of successful slots decreases and, hence, the throughput decreases. On the other hand,
for the same slot-error rate (of 10%), SLR at the RLC improves
the throughput performance. This is because the fraction of time
the channel is left idle is reduced due to the retransmission attempts. As long as the slot-error rate is reasonably good, this
would result in increased throughput. Another observation in
Fig. 6 is that the throughput achieved with BLR at the RLC
is much lower than the SLR. This is because, in BLR, even if
one slot in a block is in error, the entire block (of four slots)
will be retransmitted; this considerably reduces the throughput.
, the normalized
For example, for high arrival rates
per-channel throughput , achieved using BLR, is 0.46 (effective data rate of 4.16 kb/s using CS-1), whereas SLR achieves
a per-channel throughput of 0.52 (effective data rate of 4.52

kb/s using CS-1). Likewise, for these SLR and BLR throughput
, the effective data rate (in kilobits per second)
values at
achieved using CS-2, CS-3, and CS-4 are computed and given
in Table I, from which it is observed that gains in throughput are
possible with SLR as compared to BLR.
The mean PDU transfer delay performance of the GPRS
RLC/MAC protocol is evaluated using (10) for the same set
of parameters used in Fig. 6. The mean PDU transfer delay in
number of slots is plotted in Fig. 7 as a function of . In the
case of No Retx (i.e., the RLC unacknowledged mode), it takes
the same number of slots to carry the traffic as in the no-error
case. Hence, the delay is the same for both the no-error case
as well as the error case with no retransmission. The delay for
SLR increases as it takes more slots to successfully deliver the
data slots. In BLR, since the entire block gets retransmitted
even if one slot in a block is in error, the delay performance is
worse than SLR. For example, for values near unity the mean
delay is 19 slots for SLR and 23 slots for BLR.
The effect of the number of channels
on the throughput
characteristics of the RLC/MAC protocol is shown in Fig. 8
,
,
,
, and
. The
for
total number of channels
is varied in the range from two to
ten. From Fig. 8, we observe the following. The per-channel
increases, up to a certain a value of
throughput increases as
, beyond which the throughput decreases. This is because,
at low values of , requests are backlogged due to the nonavailability of PDTCHs, whereas PDTCHs are idle most of the
time at high values of . More interestingly, around the opfor the chosen set of parameters and traffic
timum value of
load, SLR significantly outperforms BLR. For example, when
, BLR gives a per-channel throughput of 0.47, whereas
SLR gives 0.64. The mean PDU transfer delay performance for
the same set of parameters used in Fig. 8 is illustrated in Fig. 9.

1540

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 53, NO. 5, SEPTEMBER 2004

Fig. 8. RLC/MAC average per-channel throughput  versus number of channels M . N = 10, L = 1,  = 1, g = 0:1, and g = 0:1.

Fig. 9.

Mean PDU transfer delay at RLC/MAC versus number of channels M . N = 10, L = 1,  = 1, g = 0:1, and g = 0:1.

The effect of slot-error rate

on the per-channel throughput
,
and delay performance is shown in Figs. 10 and 11, for
,
,
, and
. The plot shows the performance of RLC in unacknowledged mode (No Retx) as well
as the RLC in acknowledged mode with SLR and BLR. As expected, the per-channel throughput decreases as the slot-error
rate increases. As mentioned before, when the slot-error rate is
small (good channel condition), BLR performs slightly better
than SLR due to a larger number of overhead bits in SLR as compared to BLR. However, as the slot-error rate increases, BLR
performs more poorly than even the unacknowledged mode.
This is because even if one slot goes into error, all the four slots
of a block are retransmitted. Also, SLR is found to significantly
outperform both unacknowledged mode and BLR. This performance gain is more when the slot-error rate is high. Thus, from

Fig. 69, we observe that significant improvement in throughput

and delay performance is possible using SLR instead of the BLR
defined in the GPRS standard.

IV. LLC AND TCP PERFORMANCE

In this section, we evaluate the throughput of the LLC and
TCP, with the RLC/MAC layers below, through simulations.
We consider the LLC acknowledged mode of operation. Of particular interest here from the performance point of view is the
effect and optimum choice of the maximum LLC retransmis(LLC parameter N200 defined in the standard)
sion count
(RLC paramand the maximum RLC retransmission count
eter N3104_MAX defined in the standard). In the performance

PREMKUMAR AND CHOCKALINGAM: PERFORMANCE ANALYSIS OF RLC/MAC AND LLC LAYERS

Fig. 10.

RLC/MAC average per channel throughput  versus slot-error rate P . N = 10, M = 10, L = 1,  = 1, g = 0:1, and g = 0:1.

Fig. 11.

Mean PDU transfer delay at RLC/MAC versus slot-error rate P . N = 10, M = 10, L = 1,  = 1, g = 0:1, and g = 0:1.

evaluation of the LLC, we use a similar system model as described in Section III, except that here we assume that the message length measured in the number of LLC frames is geometrically distributed with parameter . We consider selective acknowledgment of LLC frames [15]. Also, we assume that each
LLC frame consists of five RLC blocks (20 slots). If an erroretransmission
neous LLC frame is not recovered within
attempts, the LLC is reset and reestablished. We assume that
this reset and reestablishment delay is RESET_DELAY in the
number of slots.
A. LLC Performance Results
The average per-channel throughput at the LLC layer as a
is plotted
function of maximum RLC retransmission count

1541

,
,
,
,
,
in Fig. 12, for
, and RESET DELAY
slots. Both BLR as
are considwell as SLR at a slot-error rate of
values considered are 1, 2, and 3. We define
ered. The
the throughput at the LLC layer as the average number of
successful LLC frames. From Fig. 12, we observe that, for a
, increasing
increases throughput. By increasing
fixed
, we try to recover erroneous blocks in the RLC layer itself rather than giving up the entire frame (which contains
the erroneous blocks) to the LLC. We also observe that, as
is increased, throughput increases. When
is small,
LLC resets occur frequently. A reset makes the channel to
stay idle for RESET_DELAY slots. Hence, the per-channel
. By inthroughput at the LLC is less for small values of

1542

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 53, NO. 5, SEPTEMBER 2004

Fig. 12.

LLC average per-channel throughput versus maximum number of RLC retransmissions

= 10, M = 10, L = 1,  = 0:1, P = 10%,

Fig. 13.

Mean PDU transfer delay at LLC versus the maximum number of RLC retransmissions

= 10, M = 10, L = 1,  = 0:1, P = 10%,

= 0:1, and g = 0:2.

= 0:1, and g = 0:2.

creasing
, we try to avoid too many resets, which results
in a better throughput.
The delay performance at the LLC layer for the same system
parameter values in Fig. 12 is plotted in Fig. 13. When both
and
are small, the LLC resets will occur frequently,
,
which will increase the PDU transfer delay. Thus, for small
. When
the delay performance improves by increasing
is large enough (e.g.,
), the delay performance for
2 or 3 is approximately the same. This is because most of
the erroneous RLC blocks are recovered at the RLC layer itself
and fewer errors are being passed to the LLC to recover. From
Figs. 12 and 13, it can be seen that improved throughput and

delay performance are achieved at the LLC layer using SLR at

the RLC than using BLR.
on the per-channel throughput
The effect of slot-error rate
and delay performance at the LLC is shown in Figs. 14 and 15,
again for the same system parameter values in Fig. 12. Fig. 14
and for
shows the throughput performance for
taking values one and
for both BLR and SLR. Note that
and
corresponds to no error recovery at
RLC at complete recovery at LLC. Also,
corresponds to complete recovery at RLC (and no recovery at LLC).
From Figs. 14 and 15, we can infer that SLR performs significantly better than BLR in terms of both throughput and delay.

PREMKUMAR AND CHOCKALINGAM: PERFORMANCE ANALYSIS OF RLC/MAC AND LLC LAYERS

Fig. 14.

LLC average per-channel throughput versus slot-error rate P . N = 10, M = 10, L = 1,  = 0:1, g = 0:1, and g = 0:2.

Fig. 15.

Mean PDU transfer delay at LLC versus slot-error rate P . N = 10, M = 10, L = 1,  = 0:1, g = 0:1, and g = 0:2.

From the average LLC throughput

, shown in Figs. 12
and 14, we can compute the effective data rate at LLC in kilobits
per second
using the relation
kb/s

(16)

where
represents the number of information bits in each
RLC block including the LLC header and checksum. FL represents the LLC frame length in terms of the number of RLC
blocks, BL represents the RLC block length in terms of number
of slots, and OH represents the number of overhead bits per LLC

1543

frame. Note that the 4.615 in the denominator of the above expression accounts for the one slot duration, which is equivalent
to one TDMA frame length of 4.615 ms. For the coding scheme
b for BLR and
b for SLR. In our
CS-1,
simulations, we set FL
and BL
for BLR and FL
and BL
for SLR. The overhead bits OH
corresponds to
in Fig. 14
LLC header and checksum bits. Note that
corresponds to full recovery in the RLC layer. This means that
the per-channel throughput is the same at both RLC and LLC
layers, but the effective data rates in kb/s are different. For SLR
is 0.77,
(see Fig. 14), the per-channel throughput at
which corresponds to an effective data rate of 6.19 kb/s at the

1544

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 53, NO. 5, SEPTEMBER 2004

Fig. 16. Evolution of the TCP window size

packets. rto 5000 slots.

W [i] versus the slot index i. M = 2, N = 20, L = 1, R = 3, K = 3, P = 7%, and W

LLC layer. The corresponding effective data rate at the RLC

layer is given by 6.68 kb/s.
B. TCP Performance Results
TCP is a well-known transport-layer protocol in the IP suite
[4]. TCP is a reliable connection-oriented protocol that is
widely used in popular applications such as http, ftp, telnet,
etc. Several studies have analyzed the performance of TCP
on wireless, but without considering ARQ in the link layer
[5][7]. The performance of TCP with an RLP at the link layer
in IS-95A code-division multiple-access (CDMA) system has
been studied in [21]. In this section, we estimate the throughput
performance of TCP on the GPRS uplink with the associated
LLC/RLC/MAC layers. Being a connection-oriented protocol,
TCP has call-setup, data-transfer, and call-clear phases. The
bulk throughput performance of TCP is determined primarily
by the data-transfer phase. In this study, we are interested in
evaluating the TCP throughput in the data-transfer phase and
follow the TCP data-transfer phase model described in [6].
We consider an ONOFF traffic model, such as the web and
e-mail traffic. Internet traffic typically exhibits a significant
probability for long sessions and long interarrival times between
sessions and packets. This leads to heavy-tailed (long-tailed)
complementary cumulative distribution functions (ccdfs) for
these typical measures [22]. Much work is done to find a
suitable description of Internet traffic. A common approach is
to approximate the empirical heavy-tailed ccdfs with Pareto,
Weibull, hyperexponential, or power law distributions. Such
a theoretical distribution allows the easy implementation of a
generator that produces values of the described measure with
a similar characteristic. In our simulations, to evaluate the performance of TCP over GPRS, we consider each TCP source to

= 24 TCP

generate a traffic pattern in which the interarrival time between

packets (i.e., the OFF period) follows a Pareto distribution.
The classical Pareto distribution with shape parameter and
location parameter has the cumulative distribution function
(cdf)

(17)
with the corresponding probability density function (pdf)
. We model the ON period to consist of a
geometrically distributed number of TCP packets. We assume
that one TCP packet consists of five LLC frames and that each
LLC frame contains 536 B. An LLC frame is segmented into 25
RLC blocks, each four slots in size. An erroneous RLC block
times by the ARQ mechanism
can be retransmitted up to
at the RLC layer. If the RLC layer cannot recover the erroneous
blocks, they are passed on to the LLC layer. The ARQ mechanism at the LLC layer then tries to recover erroneous LLC
times. TCP takes care of
frames by retransmission up to
the unrecovered erroneous LLC frames. We simulated 20 TCP
), each sending TCP
sessions from mobile users (i.e.,
packets to some hosts in an external packet-data network and
evaluated the performance. The ON times of the sessions have
a mean of five TCP packets. The Tahoe version of TCP with a
is used in the simulations
fast retransmit threshold of
[7].
as a
Fig. 16 shows the evolution of TCP window size
,
,
function of time for a slot-error rate of
,
,
, maximum window size of
TCP packets, and a round trip timeout (rto) value of 5000

PREMKUMAR AND CHOCKALINGAM: PERFORMANCE ANALYSIS OF RLC/MAC AND LLC LAYERS

Fig. 17.
slots.

TCP throughput performance as a function of

and

R . M = 2, N = 20, L = 1, R

slots. The window evolution of four different cases are plotted

as follows.
1) BLR with
.
.
2) BLR with
.
3) SLR with
4) SLR with
.
versus
for
In Fig. 16, a comparison between
both BLR and SLR indicates that the TCP window size is more
than
. This is because for
open for
the recovery of erroneous blocks can be incomplete, which
can result in more TCP timeouts and fast retransmits, which
shrinks the TCP window size to 1. Since larger instantaneous
window widths are good for achieving higher throughput, the
is preferred over
.
choice of parameter value
Also, a comparison between the window evolutions for BLR
reveals that the SLR results in larger
and SLR with
instantaneous window widths as compared to BLR. This larger
instantaneous window widths in SLR directly translate into
higher TCP throughput compared to BLR, which is illustrated
is increased,
in Fig. 17. From Fig. 17, we observe that as
is large,
TCP throughput increases as expected. When
more erroneous blocks are retransmitted and recovered in the
RLC layer itself, rather than leaving them to LLC or TCP to
recover by retransmitting the LLC frame or the entire TCP
packet. From the TCP throughput values plotted in Fig. 17,
, we can compute the effective data rate in kb/s at the TCP
using the relation
layer
kb/s

(18)

where
represents the number of information bits in each
RLC block including the IP/LLC headers and checksum,
represents the TCP packet length in terms of number of LLC
frames, FL represents the LLC frame length in terms of the

1545

= 3,

= 3, and

= 24 TCP packets. rto=5000

number of RLC blocks, BL represents the RLC block length

in terms of number of GSM slots, and OH represents the total
number of overhead bits per TCP packet (IP header, LLC
header/checksum). The 4.615 in the denominator of the above
expression accounts for the one-slot duration that is equivalent
to one TDMA frame length of 4.615 ms. For the coding scheme
181 for BLR and
40 for SLR. In our
CS-1,
simulations, we set PL 5, FL 25, and BL 4 for BLR and
PL 5, FL 100, and BL 1 for SLR. The number of OH bits
per TCP packet is 376 (320 IP header bits 56 LLC overhead
bits). In Fig. 17, for
the TCP throughput is 0.1021
for BLR, which corresponds to an effective date rate of 0.93
kb/s at the TCP layer. For the same system parameters, for
SLR, the TCP throughput is 0.7087, which corresponds to an
effective date rate of 6.02 kb/s at the TCP layer. Thus, there is
a substantial improvement in the effective data rate at the TCP
layer as we go from block level to SLR at the RLC, particularly
for moderate-to-high channel-error rates.
V. CONCLUSION
We analyzed the performance of various layers
(TCP/LLC/RLC/MAC) of the GPRS protocol stack. Using
the theory of Markov chains, we derived the throughput and
delay performance of the GPRS MAC protocol. We showed
that the SLR scheme at the RLC layer performs better than the
BLR defined in the standard. We investigated the interaction
between the ARQs at the RLC and LLC layers; in particular,
we studied the effect of the choice of the maximum number of
retransmission attempts allowed at the RLC and LLC layers.
Our performance results showed that it is more beneficial to do
error recovery by allowing more retransmissions at the RLC
layer than at the LLC layer. We also evaluated the performance
of TCP with BLR and SLR at the RLC layer and showed
that TCP performed better with SLR than compared to with

1546

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 53, NO. 5, SEPTEMBER 2004

BLR. As further investigation, the performance analysis can

be further extended for correlated channel errors, multislot
operation in GPRS, and EGPRS. Performance analysis on the
downlink is another topic for further investigation.
REFERENCES
[1] G. Brasche and B. Walke, Concepts, services, and protocols of the new
general packet radio service, IEEE Commun. Mag., vol.
GSM phase
35, pp. 94104, Aug. 1997.
[2] C. Bettstetter, H. J. Vgel, and J. Eberspcher. (1999) GSM
phase
general packet radio service GPRS: Architecture, protocols, and air interface. IEEE Commun. Surv. [Online], pp.
214. Available: http://www.comsoc.org/livepubs/surveys/public/
3q99issue/bettstetter.html
[3] A. Furuskar, S. Mazur, F. Muller, and H. Olofsson, EDGE: Enhanced
data rates for GSM and TDMA/136 evolution, IEEE Pers. Commun.
Mag., vol. 6, pp. 5666, June 1999.
[4] W. R. Stevens, TCP/IP Illustrated. Reading, MA: Addison-Wesley,
1994, vol. 1.
[5] T. V. Lakshman and U. Madhow, The performance of TCP/IP for
networks with high bandwidth-delay products and random loss,
IEEE/ACM Trans. Networking, vol. 5, pp. 336350, June 1997.
[6] A. Kumar, Comparative performance analysis of versions of TCP in a
local network with a lossy link, ACM/IEEE Trans. Networking, vol. 6,
pp. 485498, Aug. 1998.
[7] M. Zorzi, A. Chockalingam, and R. R. Rao, Throughput analysis of
TCP on channels with memory, IEEE J. Select. Areas Commun., vol.
18, pp. 12891300, July 2000.
[8] J. Cai and D. J. Goodman, General packet radio service in GSM, IEEE
Commun. Mag., vol. 35, pp. 122131, Oct. 1997.
[9] G. Brasche and B. Walke, Analysis of multi-slot MAC protocols proposed for the GSM phase
general packet radio service, in Proc.
IEEE Vehicular Technology Conf. (VTC97), 1997, pp. 12951300.
[10] S. Hoff, M. Meyer, and A. Schieder, A performance evaluation of internet access via the general packet radio service of GSM, in Proc. IEEE
Vehicular Technology Conf. (VTC98), vol. 3, May 1998, pp. 17601764.
[11] Y. Argyropoulos, A. Bigloo, and J. van Peursem, GPRS delay and
capacity analysis for web browsing application, in Proc. Int. Conf.
Telecommunications, vol. 2, 1998, pp. 3236.
[12] C. Demeterscu, LLC-MAC analysis of general packet radio service in
GSM, Bell Lab.s Tech. J., pp. 3750, 1999.
[13] R. Kalden, I. Meirick, and M. Meyer, Wireless internet access based on
GPRS, IEEE Pers. Commun. Mag., vol. 7, pp. 818, Apr. 2000.
[14] K. Premkumar and A. Chockalingam, Performance analysis of
RLC/MAC protocol in general packet radio service, in Proc. National Conf. Communications (NCC01), Kanpur, India, Jan. 2001, pp.
173177.
[15] Digital cellular telecommunications system (phase
); General
packet radio service (GPRS); mobile station-serving GPRS support
node (MS-SGSN) logical link control (LLC) layer specification (GSM
04.64). European Telecommunications Standards Institute (ETSI)
TC-SMG GPRS ad hoc, Sophia Antipolis, France. [Online]. Available:
http://www.etsi.org/
[16] Digital cellular telecommunications system (phase
); General
Packet radio service (GPRS); mobile station (MS)-BSS interface;
radio link control/medium access control (RLC/MAC) protocol (GSM
04.60). European Telecommunications Standards Institute (ETSI)
TC-SMG GPRS ad hoc, Sophia Antipolis, France. [Online]. Available:
http://www.etsi.org/

2+

2+

2+

2+

2+

[17] M. Zorzi and R. R. Rao, Capture and retransmission control in mobile

radio, IEEE J. Select. Areas Commun., vol. 12, pp. 12891298, Oct.
1994.
[18] M. Zorzi, Capture probabilities in random-access mobile communications in the presence of Rician fading, IEEE Trans. Veh. Technol., vol.
46, pp. 96101, Feb. 1997.
[19] Evaluation criteria for the GPRS radio channel Tdoc 56/95. European
Telecommunications Standards Institute (ETSI) SMG2 GPRS ad hoc,
Sophia Antipolis, France. [Online]. Available: http://www.etsi.org/
[20] R. W. Wolff, Stochastic Modeling and the Theory of Queues. Englewood Cliffs, NJ: Prentice-Hall, 1989.
[21] A. Chockalingam and G. Bao, Performance of TCP/RLP protocol
stack on correlated Rayleigh fading DS-CDMA links, IEEE Trans.
Veh. Technol., vol. 49, pp. 2833, Jan. 2000.
[22] V. Paxson and S. Floyd, Wide-area traffic: The failure of poisson modeling, IEEE/ACM Trans. Networking, pp. 226244, June 1995.

K. Premkumar was born in Chennai, India, on

December 27, 1975. He received the B.E. degree
in electronics and communication engineering from
the Alagappa Chettiar College of Engineering and
Technology, Karaikkudi, India, in 1997 and the
M.S. (research) degree in electrical communication
engineering from the Indian Institute of Science,
Bangalore, in 2002.
From August 2001 to September 2003, he was with
Motorola India, Bangalore, where he was involved in
the design and development of 3G systems and multiuser receivers. Since September 2003, he has been with the Indian Institute of
Science, doing research in MIMO multiuser information theory. His research
interests include the performance analysis of wireless communications and networks, power control, transmit diversity, and multiuser information theory.

A. Chockalingam (S93M95SM98) received

the B.E. (Hons.) degree in electronics and communication engineering from the P. S. G. College
of Technology, Coimbatore, India, in 1984, the
M.Tech. degree (with a specialization in satellite
communications) from the Indian Institute of Technology, Kharagpur, in 1985, and the Ph.D. degree in
electrical communication engineering (ECE) from
the Indian Institute of Science (IISc), Bangalore,
India, in 1993.
From 1986 to 1993, he was with the Transmission
R & D Division, Indian Telephone Industries Ltd., Bangalore. From December
1993 to May 1996, he was a Postdoctoral Fellow and an Assistant Project Scientist in the Department of Electrical and Computer Engineering, University
of California, San Diego (UCSD). From May 1996 to December 1998, he was
in the Systems Engineering Group, Qualcomm Inc., San Diego, CA, as a Staff
Engineer/Manager. In December 1998, he joined the Faculty in the Department
of Electrical Communication Engineering, IISc, where he is an Associate Professor, working in the area of wireless communications and networks. He was
a Visiting Faculty with UCSD during the summers of 19992002.
Dr. Chockalingam is a Recipient of the Swarnajayanti Fellowship from the
Department of Science and Technology, New Delhi, India. He is an Associate
Editor for the IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY.