Bandwidth Aggregation in
Heterogeneous Wireless Networks
There are several issues that
need to be addressed to enable simultaneous use
of multiple interfaces. These span different layers
of the protocol stack and need multi-disciplinary
approaches to devise appropriate architecture
and scheduling policies. We have looked in depth
at the problem of bandwidth aggregation. Since
TCP is the dominant transport protocol in use
today, we have evaluated the performance of a
file transfer using TCP over such bandwidth
aggregation. We have also looked at the performance
of real-time interactive applications with strict
QoS requirements.
Towards realizing the objective
of simultaneous use of multiple interfaces, we
have proposed a general framework in the form
of an architecture based on an extension to Mobile-IP,
a popular network layer mobility management protocol
standardized by IETF. In contrast to transport/application
layer solution, such a network layer approach
introduces minimal changes to the infrastructure
while handling mobility well.
While the use of multiple interfaces
allows us to increase throughput, the varying
characteristics of the different paths (corresponding
to different interfaces) introduce problems in
the form of packet reordering. With respect to
TCP, this reordering can degrade application throughput
significantly as TCP misinterprets reordering
as indicative of packet loss and invokes congestion
control cutting down the sending rate. This could
even counter any gains that can be had through
bandwidth aggregation. Therefore, to improve overall
performance of TCP, we took a twopronged
approach: (1) we proposed a scheduling algorithm
that partitions traffic onto the different paths
(corresponding to each interface) such that reordering
is minimized; (2) a buffer management policy was
introduced at the client to hide any residual
reordering from TCP.
Simulation studies have shown that this approach
can achieve good bandwidth aggregation under a
variety of network conditions. The performance
is comparable to "MTCP", an application
layer solution that opens multiple TCP connections
with one on each interface (the best possible).
For real-time interactive applications,
packet reordering introduces excess delay. Because
of stringent QoS requirements, this late arrival
of packets is often equivalent to loss. Thus,
to reduce the delay associated with reordering,
we proposed a scheduling policy Earliest
Delivery First (EDF) that partitions the traffic
onto different paths such that the QoS requirements
of the application are met. We theoretically analyzed
the performance of EDF and show that it performs
close to an idealized Single Link (SL) discipline,
where the multiple interfaces are replaced by
a single interface with the aggregate bandwidth.
We carried out simulations using video and delay
traces and we observed that under a variety of
network conditions, EDF mimics SL closely and
outperforms by a large margin other straight-forward
scheduling policies like weighted roundrobin.
While our approach has been based on simulation
and analysis so far, we plan to implement some
of the ideas on an actual testbed. In addition,
we will look into other aspects of bandwidth aggregation
related to power consumption, security, etc. While
use of multiple interfaces may be a drain on the
battery, the power savings in time had through
bandwidth aggregation need to be quantified. Security
is another important issue in wireless networks.
Use of multiple interfaces means securing a greater
number of paths, but through proper encryption,
connection can be made inherently more secure
as it is difficult for an attacker to snoop on
all the paths.