You are on page 1of 4

Energy Efficiency Evaluation of State of the Art

Packet Scheduling algorithms for LTE


Dario Sabella, Marco Caretti, Roberto Fantini
Telecom Italia SpA
Torino, Italy
{dario.sabella, marco.caretti, roberto.fantini}@telecomitalia.it
Abstract In this contribution, the performance of LTE system
with various scheduling algorithm taken from the literature are
considered. In order to assess the performance of state of the art
schedulers a dynamic system level simulator is used that models
in details all the features of the downlink of an LTE system in a
multi-cell environment. The simulator will be used to analyze the
impact of different scheduling policies from an energy efficiency
point of view, considering different traffic models. In particular
the first step involves analyzing the performances of SOTA
algorithms, because up to now these schedulers have been
compared from a more traditional point of view (e.g. SE, spectral
efficiency) and not for the energy efficiency (EE) evaluation.
Keywords: packet scheduling; LTE; energy efficiency.
I. INTRODUCTION
Packet scheduling operation in HSDPA (High Speed
Downlink Packet Access) and in LTE (Long Term Evolution)
networks is widely considered as one of the key parts devoted
to guarantee the QoS whilst maximizing the provided system
capacity. The goal of a resource scheduling algorithm is to
allocate the resources and transmission power to the different
users in each subframe to optimize one or a set of metrics (like
throughput, delay, outage probability) considering different
constraints (for example in the downlink the resource
allocation strategy is constrained by the total transmission
power available at the base station).
Up to now packet scheduling algorithms have been studied
from the capacity and QoS point of view, but in the context of
the present work also other suitable metrics should be
considered in order to assess the energy efficiency of the
system, also because a potential energy saving could be
realized by the design of proper algorithms, especially when
considering advanced power models at the component level.
In this study, the performance of various classical
scheduling algorithms taken from the literature is considered.
Round Robin, Proportional Fair and Max C/I are the starting
point of this study to analyze the different solutions also from
an energy efficiency point of view. In order to assess the
performance of state of the art schedulers a dynamic system
level simulator is used that models in details all the features of
the downlink of an LTE system in a multi-cell environment
(like HARQ Hybrid Automatic Repeat reQuest, link
adaptation and multi-antenna techniques). The simulator
(described in section III) will be used to analyze the impact of
different scheduling policies from an energy efficiency point of
view, considering different traffic models. In particular the first
step involves analyzing the performances of SOTA (State Of
The Art) algorithms, because up to now these schedulers have
been compared from a more traditional point of view (e.g. SE,
spectral efficiency) and not for the energy efficiency (EE)
evaluation. This consideration is valid also for other algorithms
present in literature: in general it could happen that an
algorithm designed for a certain purpose (e.g. SE) could show
different performance or different relative gains if analyzed
from different perspectives (e.g. EE). This consideration is
valid for SOTA schedulers [1][2], but also as a basis for an
assessment of innovative energy aware algorithms.
II. CONSIDERED APPROACH FOR COMPARATIVE ANALYSIS
The considered approach for the comparative analysis of
packet schedulers performances starts from analyzing the
system performances in terms of cell throughput considering
fullbuffer traffic. As a first step we assess the scheduler
efficiency from the point of view of the ratio between radiated
power and served cell throughput; for this purpose we define
an RF consumption indicator:
) (
) (
) (
i
i
RF i
RF
T
P
= (1)
where
) (i
RF
P is the average RF power,
) (i
T is the average
DL cell throughput in the i-th snapshot, and for an N snapshot
based simulation we calculate the RF consumption indicator

=
=
N
i
i
RF RF
N
1
) (
1
of the overall simulation.
As a general approach, in order to provide system level
results, a power model is then needed for analyzing the
performances in terms of total consumed power; for a complete
analysis we consider the energy consumption metric defined in
[6]:
) (
) (
) (
i
i
IN i
IN
T
P
= (2)
where
) (i
IN
P is the average power consumed by the eNB in
the i-th snapshot and depends on the considered power model,
This work has been carried out within the EARTH (Energy Aware Radio
and neTworking tecHnologies) project, which partly funded by the European
Commission under project FP7-ICT-2009-4-247733-EARTH.
European Wireless 2011, April 27-29, 2011, Vienna, Austria ISBN 978-3-8007-3343-9 VDE VERLAG GMBH
Paper 1569421569 717
and the metric

=
=
N
i
i
IN IN
N
1
) (
1
considers all the snapshots
in the simulation.
The
RF
indicator has been introduced for the scheduler
comparison because it could offer additional information
respect to the overall energy consumption metric
IN
, that
instead doesnt provide information about the efficiency in
terms of RF power usage. In other words, the
RF
indicator can
be viewed as an intermediate term of the global metric
IN
, and
tells us how is efficient is the packet scheduler from the point
of view of the transmitted power, while the global metric
IN
indicates which is the impact of hardware and algorithms for
system level results.
The two indicators are strictly related, and in general this
relationship depends on the used power model (the law of
variation of the power consumption respect to the radiated
power). As an example, in first simulations we considered
fullbuffer traffic: in this case the relationship between the two
indicators is straightforward; in fact, for the RF consumption
indicator we have:

=
=
N
i i
RF
RF
T
P
N
1
1
(3)
where the radiated power P
RF
is constant (equal to 40W) for
all the N different snapshots and then also the power
consumption in general can be supposed constant; in particular,
under the hypothesis of considering a linear power model (P
IN
= aP
RF
+ b), we have:
=
+
= =

= =
N
i i
RF
N
i i
IN
IN
T
b aP
N T
P
N
1 1
1 1


RF RF
RF
RF
N
i i
RF
k
P
b aP
T N
b aP =
+
= + =

=
) ( 1 1
) (
1

(4)
Then in this particular case we have a direct proportionality
between the two indicators (
RF IN
k = ), and a relative gain
in terms of P
RF
is equivalent to a relative gain in terms of P
IN
:
RF
RF
RF
RF
IN
IN
k
k


(5)
In other words at full load the energy efficiency of the
system is limited by the spectral efficiency of the packet
scheduler and the relative gains dont depend on the particular
linear power model considered.
In future studies, with different cell loads and/or more
sophisticated power models, the two parameters will be useful
to assess the energy efficiency of the scheduler and the impacts
of the hardware improvements on the system performances.
III. SIMULATOR DESCRIPTION
The used system level simulator implements the so-called
snapshots method. With this method, snapshots of the system
are taken at random time intervals to obtain a broad view of the
systems performance in general conditions. In practice, the
procedure is carried out by dropping the users in the network at
random positions, uniformly distributed in each cell. Each
snapshot is then simulated for a duration which was verified to
be long enough to produce an adequately large population of
samples. During the simulation of a snapshot, we assume that
the network is stable: no MSs join/leave the network; there is
no inter-cell hand-over; connections are not established/torn-
down; all traffic sessions are continuously active (even if the
considered traffic could have bursty characteristics at TTI
level). Therefore, we did not simulate the management
procedures to handle the previous events. This assumption is
reasonable since each snapshot only represents an
instantaneous state of the system, while time evolution is
inferred by aggregating results from many snapshots, which are
not correlated between each other. The energy efficiency
metrics have been introduced in the simulator to compare the
different schedulers from an energy efficiency point of view.
The dynamic system level simulator models in details the
LTE downlink radio interface. In particular the following
aspects are implemented:
Spatial Channel Model with spatial correlation, in
accordance with the ITU-R model [3]. The channel
model has been calibrated according to the
methodology defined by WINNER+ [4];
Physical layer abstraction in accordance with the
MIESM model, implemented for different
transmission modes (SISO, SIMO, MIMO, BF);
LTE link level performance curves in AWGN channel
evaluated by simulation platform with detailed
physical layer modeling, realistic channel estimation,
different modulation & coding rates and transport
block sizes.
In TABLE I a summary of main simulation parameters is
listed, while a brief description of the assessment methodology
is provided in the following.
718
TABLE I. SYSTEM PARAMETERS FOR LTE SIMULATION.
Parameters Value
Carrier frequency 2000 MHz
Bandwidth
10 MHz
Inter-site Distance
500 m
TTI duration
1 ms
Cell layout Hexagonal grid, 19 3-sectored
cells
Link to system interface MIESM
Traffic model Full buffer
Total number of subcarriers 600
N. of subcarriers per resource
block
12
Total number of resource blocks 50
Access technique OFDMA
Radio Link Downlink
Number of antennas (Tx,Rx) (1,2)
Fast fading model SCM channel, Urban Macro
Interference model
Explicit (the 9 strongest
interference cells are
considered)
Channel estimation Realistic
CQI reporting
Wideband CQI, no PMI on
PUCCH (mode 1-0)
Number of MCS 3GPP LTE standard format
AMC PER
target
10%
CQI report
5 ms periodicity, with 6 ms
total delay (measurement in
subframe n is used in
subframe n+6)
Packet scheduling
Round robin, Max C/I,
Proportional fair
Number of OFDM symbols in
control region
3
HARQ type
synchronous adaptive
stop&wait
Number of HARQ processes 8
Retransmission interval 8 ms
Max number of retransmissions 3
IV. SIMULATION RESULTS
First let us compare the DL cell throughput as the function
of the number of users per cell. In this analysis a range from 5
to 35 contemporary users is considered, and this hypothesis is
coherent with an LTE traffic scenario in which even fewer
users use all the radio resources, e.g. for downloading huge
files or video streaming contents. In this saturation condition it
is important to analyze the behavior of packet scheduling
algorithms, also by varying the number of user that share the
resources. The results are depicted in Fig. 1 and show that the
difference between the three schedulers less in case of few
users per cells, while increasing the number of users Max C/I
scheduler better exploits the multi-user diversity and provides a
throughput up to 2 times compared to the other two schedulers.
0
5000
10000
15000
20000
25000
5 10 15 20 25 30 35
Average number of users per cell
D
L

c
e
l
l

t
h
r
o
u
g
h
p
u
t

[
k
b
p
s
]

Round Robin
Proportional Fair
Max C/I
Figure 1. DL cell throughput for LTE 10 MHz, SIMO 1x2, with full buffer
traffic.
Since for full buffer simulations (where the cell is supposed
at full load and users are infinitely backlogged) radiated power
is constant (40W), the provided results reflect also a dual trend
in terms of energy efficiency. Let us define
RF
= P
RF
/T a RF
consumption indicator and
IN
= P
IN
/T our system level energy
efficiency metric (where T is the throughput at IP layer, P
IN
is
the input power and P
RF
is the radiated power). FIGURE 2
shows a comparison in terms of
RF
, while the relative gain
with respect to Round Robin scheduler is shown in FIGURE 3.
Results are given in terms of
IN
even if, as shown in formulas
of Appendix 8.7, the same trend can be observed if considering

RF
. As we can see, the obtained gains (especially for the Max
C/I scheduler) depend on the number of users in the cell, even
if we are always in full load conditions: this because the
behavior of packet scheduling algorithms depends on the
exploitation of the multi-user diversity.
Even if these simulation results show good performances of
Max C/I in terms of both spectral and energy efficiency, the
counterpart is offered by the fairness provided by the different
scheduling algorithms (see FIGURE 4). That is, the de-facto
standard fairness measure, the Jains fairness index [5] is less
than the half of the other two schedulers independently of the
number of users per cell.
As a consequence, for this preliminary analysis
Proportional Fair scheduling could be identified as a good
compromise between considered schedulers from the point of
view of the energy efficiency and the fairness provided to the
users, even though the obtained gain depends also on the
number of users present in the cell.
Current simulation results are relative to full buffer traffic,
while further analysis will consider also non full buffer traffic
and different traffic loads and could benefit of using the
EARTH power model [6] integrated in the system level
simulator, in order to perform a comparison between
schedulers in suitable scenarios of interest for the energy
efficiency evaluation. As already said, simulation results show
that performance and the amount of EE gain provided by the
schedulers depends on the number of users sharing the cell
radio resources, and this suggests that performances of future
possible EE schedulers should be analyzed not only by varying
719
the total cell load (and then considering different periods of the
day) but the assessment should take into account the amount of
users present in the system and their traffic profile.
0
1
2
3
4
5
6
7
8
9
10
5 10 15 20 25 30 35
Number of average users per cell
P
R
F

o
v
e
r

C
e
l
l

T
h
r
o
u
g
h
p
u
t


[
W
/
M
b
p
s
]
Round Robin
Proportional Fair
Max C/I
Figure 2. RF consumption indicators of different SOTA scheduling
algorithms.
0%
10%
20%
30%
40%
50%
60%
70%
5 10 15 20 25 30 35
Number of average users per cell
G
a
i
n

i
n

t
e
r
m
s

o
f

P
IN

/

T


[
W
/
M
b
p
s
]
Proportional Fair
Max C/I
Figure 3. Relative gain respect to Round Robin algorithm.
0,00
0,10
0,20
0,30
0,40
0,50
0,60
0,70
0,80
0,90
1,00
5 users 35 users
Number of average users in the cell
J
a
i
n
'
s

f
a
r
i
n
e
s
s

i
n
d
e
x
Round Robin
Proportional Fair
Max C/I
Figure 4. Jains fairness index of different scheduling algorithms. LTE 10
MHz, SIMO 1x2, fullbuffer traffic.
V. CONCLUSIONS
Up to now packet schedulers have been compared from a
more traditional point of view (SE, spectral efficiency) and not
for the energy efficiency (EE) evaluation purposes. In this
study, the performances of various classical scheduling
algorithms taken from the literature are considered, as a basis
for the assessment of further innovative energy aware
algorithms. Gains in terms of energy consumption index (with
respect to baseline Round Robin scheduler) are analyzed.
Based on the assumption to consider a linear relationship
between RF output power and total consumed power (linear
power model), simulations results showed that at full load the
energy efficiency of the system is limited by the spectral
efficiency of the packet scheduler and the relative gains shown
in the paper do not depend on the particular parameterization of
the considered power model. As shown in this paper a good
compromise between maximum efficiency and fairness among
the considered schedulers is given by Proportional Fair
algorithm.
ACKNOWLEDGMENT
This work has been carried out within the EARTH (Energy
Aware Radio and neTworking tecHnologies) project, which
partly funded by the European Commission under project FP7-
ICT-2009-4-247733-EARTH.
REFERENCES
[1] Petteri Kela, Jani Puttonen, Niko Kolehmainen, Tapani Ristaniemi, Tero
Henttonen and Martti Moisio, Dynamic Packet Scheduling
Performance in UTA Long Term Evolution Downlink, ISWPC 2008.
[2] The impact of QoS support on the end user satisfaction in LTE networks
with mixed traffic, Siomina, I.; Wanstedt, S.; Ericsson Research.
Personal, Indoor and Mobile Radio Communications, 2008. PIMRC
2008.
[3] ITU-R M.2135 Guidelines for evaluation of radio interface
technologies for IMT-Advanced.
[4] http://projects.celtic-initiative.org/winner+/.
[5] Jain, R., Chiu, D.M., and Hawe, W. (1984) A Quantitative Measure of
Fairness and Discrimination for Resource Allocation in Shared Systems.
DEC Research Report TR-301.
[6] EARTH project deliverable, D2.3 Energy Efficiency Analysis of the
Reference Systems, Areas of Improvements and Target Breakdown,
2010.
720

You might also like