|
Cross-Layer Optimization
Smart Antennas in Ad-Hoc Network
My current
research is funded by DARPA/Boeing and involves the development of a
high fidelity simulator that incorporates the impact of time-varying
channel conditions on the throughput performance of ad-hoc
communication networks of Future Combat Systems.
The prime objective of this research is to investigate methods of
increasing the fidelity of network simulations in manners which
provide little to moderate penalties in simulation runtime.
Traditional
simulation methodologies have so far followed independent approaches
to the modeling of physical and network layer platforms. The
traditional approach is unsatisfactory when investigating many
advanced communication techniques such as ad-hoc networks, smart
antennas, space-time coding and packet scheduling in wireless data
networks. Hence a cross layer perspective to this problem is needed.
Use of smart
antennas in ad-hoc networks makes a good test-case for this purpose.
Two methods of integrating the physical layer and network layer
simulation environment are considered below:
Baseline/OPNET-Only
Approach
This is the
traditional simulation technique. OPNET network
simulator is used to investigate the performance improvements in
target application by abstracting physical layer. Higher layer
improvements are implemented directly, but physical layer
modifications must be incorporated into abstraction of physical
layer, e.g., changing the packet error rate and/or distribution. Runtime is
relatively short since physical layer is very efficient.
Physical
layer modeling is statistical hence OPNET is not very accurate for
the investigation of the impact of time-varying channel conditions
on the network performance. OPNET is an event driven
simulation tool and is not suitable to accurately reflect real world
scenarios which need to be processed in a time driven method such as
MATLAB.
Method-I: Software-Only Approach (OPNET with C++ or Matlab physical layer simulator)
In this method
we use a software network simulation tool (e.g., OPNET) and build a
software based physical layer simulation tool using MATLAB or C++.
OPNET interacts with separate physical layer simulator through
middleware.
Specifically, we
have investigated the performance of adaptive antenna arrays in
wireless local area networks. For the first part we have looked into
the network performance of an OPNET only model equipped with omni
directional antennas. We have utilized the fixed antenna gain
pattern corresponding to an isotropic antenna provided by the OPNET
antenna pattern editor. This model has been built to quantify the
extent to which a hostile jammer can impact the throughput of the
system. This OPNET only approach was further extended to include the
adaptive antenna array. An interesting but expected result was a
good improvement in network performance with smart antennas. It
should be highlighted here that a separate baseline model applicable
to wireless local area networks was also developed in order to
instructively compare the baseline with Method-I on a common
platform.
Scenario1:
The scenario involves a
radio network consisting of two stationary communication nodes and a
mobile jamming node which moves along a user-defined trajectory.
FCS_Tx_Fixed node transmits at uniform power in all directions.
The mobile jammer causes increasing and decreasing levels of
interference at the receiver which affects the bit error rate (BER)
performance as well as the good throughput of the network. Figure 1
compares the simulations results for bit error rate performance and
throughput performance for the cases: (a) FCS_Rx_Fixed
employs an isotropic antenna and (b) FCS_Rx_Fixed uses
an adaptive antenna array. The bit error rate plot for
isotropic antenna pattern shows that the bit error rate gradually
varies as the jammer moves along its trajectory and reaches its peak
value when the two are closest to each other. On the other hand,
when smart antennas are used, (1, 2)-CMA algorithm,
the bit error rate at the receiver node is always close to zero.
This drop in BER has a direct impact on the receiver throughput as
shown in the figure. Unlike the throughput curve for the isotropic
antenna, the throughput for smart antenna case is almost independent
of the location of the mobile jammer.
Fig. 1 A Radio
Network with 2-statinory communication nodes and a mobile jammer (See
full-size image)
Scenario-II
(Application of Adaptive Antennas to WLAN Models):
This scenario
serves as a baseline for fidelity. The use of transmit beam forming
and its advantages in providing a better abstraction of the physical
layer are studied in conjunction with a WLAN network:
-
mobile_node_0
is a mobile node and nodes 0, 1, 2, 3, 4 and 5 are stationary
nodes
-
mobile_node_0
follows an almost circular trajectory around these stationary
nodes and is configured to transmit packets randomly
-
mobile_node_0
has been configured to form a transmit beam towards node_2
at different instances of time at three different locations
The attributes for each
of the antennas are configured in antenna attributes and the mobile
node is enabled to plot its antenna patterns during simulation.
During the simulation, node movement can be observed in the
animation viewer and corresponding antenna patterns are created by
Matlab (a snapshot of a typical simulation is shown below)
Fig. 2 A snapshot of a
user-distribution scenario with a mobile node (See
full-size image)
The cross-layer
simulation carried out in this case is the most accurate as it
provides the best abstraction possible of the physical layer.
Results obtained from OPNET alone using the standard gain pattern
editor is assumed to be less accurate and hence ranks lower in terms
of approaching simulation fidelity.
Scalability/Run-time Comparisons:
Matlab enables very
accurate modeling of the physical layer parameters and is
characterized by the presence of a vast library of functions for
various blocks of communication systems and channel models. It
should be highlighted however, that the simulation runtime using
Matlab is a major bottleneck. Although Matlab provides increased
fidelity and shorter code development cycles, frequent function
calls and data exchange between the OPNET pipeline stages and Matlab
translate into increasingly longer runtimes. An alternative solution
to this problem and the one followed by us through our code
development cycle was to first create the entire physical layer
simulator in Matlab and later translate the final working code into
a custom C-code compiled directly into OPNET. This translated into
far superior runtime performance (See Figure-3 below).
Fig. 3 Runtime comparison
for different Simulation platforms (See
full-size image)
Method-II: OPNET network
simulator with hardware/software physical layer
The previous
approach while providing the necessary modeling of the interaction
between network layers, may still be run-time limited. The
requirement for simulation arises from two aspects of systems; the
complexity of the system and the complexity of the environment, due
to multipath, fading, nonlinearities and interference, in which the
system operates. In addition, performance requirements often
dictate that these systems operate at low bit error rates. The low
bit error rate, coupled with the complexity of the system, leads to
the requirement for very long simulation runtimes when Monte Carlo
simulation techniques are used in a traditional, single processor,
computing environment.
In order to reduce
the simulation runtime, hardware in the loop simulations are being
explored and proof-of-concept simulations have been developed. The
overall simulation environment is illustrated in Figure 2. This is
clearly a heterogeneous architecture since, in addition to usual
simulation engines consisting of general-purpose computers,
programmable DSP chips, FPGA devices, and communications hardware
are used as part of the simulation. Using this architecture, those
portions of the system that are inefficiently simulated on a
general-purpose computer can be off-loaded to a dedicated device,
such as a FPGA engine.
Fig. 4 Distributed
simulation architecture with hardware
This work is
currently in its early stages. The research team consists of 4 MS
students (including myself) and 2 PhD students. | 
