Transcript Introduction - Adaptive Systems Lab

Introduction



Future wireless systems will be characterized by
their heterogeneity - availability of multiple access
systems in the same physical space.
Each system differs from others in terms of its
capabilities - data rates, latencies, cost per byte
etc.
Given such choice of systems, end users need not
restrict themselves to any single system but rather
at every instant, depending on their requirements
choose the system or systems that best cater to his
needs.
Example Scenario




A laptop is equipped with multiple interfaces. The
applications running on it have bandwidth
requirements higher than can be offered by any
single system.
A solution is to aggregate the bandwidth offered
by each of these different systems.
For example, if the different interfaces provide a
bandwidth of 56kbps, 128kbps and 144kbps, the
laptop has at its disposal an aggregated bandwidth
of (56+128+144) = 328kbps.
The goal is to minimize the cost while satisfying
the QOS requirements of the applications.
Other Scenarios

A corporate car fleet or rental cars equipped with
multiple wireless interfaces.


The multiple passengers have different profiles and
requirements and have to share the combined
bandwidth of the interfaces among themselves.
An ad-hoc network formed by devices some of
which are equipped with WWAN interfaces.


The devices form the ad-hoc network with their WLAN
interface (bluetooth,802.11).
The WWAN interfaces provide access to the internet
and the aggregated bandwidth is shared among
themselves.
Laptop Scenario
Issues Involved




The multiple access systems display variable data
rate, latency, packet loss.
These variations cause packet reordering.
Buffering helps but it means increased delay.
Situation only worsens when a reliable protocol
like TCP is used.
Approach


Problem can be addresses at different layers of the protocol
stack- Network, Transport.
Network layer approach




A single TCP/UDP connection end to end.
Mobile IP like infrastructure
Scheduler distributes the traffic to the multiple systems at the
network layer.
Transport layer solution



Multiple TCP connections one for each access system being used.
A scheduler distributes application data to the multiple TCP
connections.
The connections can terminate at the home agent or correspondent
node.
Network Layer Solution - Scheduling
Algorithm



Packets of different applications need to be scheduled onto
the different links so that each application gets its share of
bandwidth.
Care should be taken so that delay, reordering and jitter
experienced by the packets is minimized.
Our approach is to use a Weighted Fair Queuing(WFQ)
algorithm (that ensures that each application gets its share
of bandwidth) and follow it with a channel striping
algorithm.
App1
App2
App3
Link1
WFQ
Channel
Striping Alg
Link2
Link3
Channel Striping Algorithm
Channel
Striping
Algorithm
Channel Striping Algorithm: EDF



This algorithm schedules packets from a single
input queue onto multiple links.
In our algorithm (which we call EDF), we
schedule packets on the link which delivers it the
earliest.
Each link l is associated with three quantities



A variable S_l, which is the time the link becomes
available for the next transmission.
D_l, the delay (estimated) associated with the link
BW_l, the bandwidth (estimated) of the link
EDF cont..

If we denote by a_i, the arrival instance of the ith packet
and L_i, the size of the packet, we know that this packet
when scheduled on link l would arrive at the receiver at
R_l, where
R_l = MAX(a_i+D_l,S_l) + L_i/BW_l


EDF schedules the packet on the link for which R_l is the
minimum.
Since the criteria is earliest delivery, it minimizes the
delay, buffering required and jitter experienced by the
packets.
UDP Flows – Simulation Setup
(350kbps,100ms)
(10Mbps,20ms)
Sender
Home
Agent
WAN
1
(100kbps,120ms)
WAN
2
Mobile
Receiver
(Buffering)
(50kbps,150ms)
WAN
3
Higher
Layers
Details of Setup





The sender generates packets according to an arrival and
packet size distribution and forwards them to the home
agent.
The scheduling algorithm in the home agent distributes the
packets onto the multiple interfaces.
The WAN clouds induce delay according to a delay
distribution.
The hop to the mobile receiver is wireless and has a limited
bandwidth (bottleneck). Packets would be dropped
according to a packet loss model.
The mobile receiver collects the packets and sends them in
order to higher layers.
UDP Flows Cont…

We compare the performance of our algorithm
with two other algorithms.


A Single Link Algorithm (SL), where the multiple links
between the home agent and the mobile are replaced
with a single link.
Surplus Round Robin which distributes traffic in
proportion to the bandwidth.
Trace Driven Simulation

Source Details






An MPEG1 video trace ( a cable TV show)
5000 frames
Capture rate 25fps
Mean bit rate – 440kbps, peak bit rate – 1760kbps
4 MTU sizes(bits) – 4000,8000,12000 and no restriction
Network details


Constant and truncated Gaussian delay distribution with standard
deviation 50ms.
No packet loss
Buffering Required – constant delay
Buffering Required – variable delay
Jitter – constant delay
Jitter – variable delay
TCP Flows – Details




Same setup as before except the protocol used is
TCP.
Application - File transfer of 2 Mbytes
Download time is measured
Two scenarios considered


The packets are passed to TCP layer as they arrive.
The packets are buffered and passed in order to TCP
layer
Download time (No buffering)
Algorithm
Constant
Delay
Variable
Delay
SL (500kbps)
33.4
69.9
EDF
33.6
140.9
SRR
111.4
177.0
SL (350kbps)
48.6
71.9
Future Work


UDP flows – evaluate the performance of the system with
packet loss and rate control.
TCP flows





Buffering at the receiver
Opening multiple TCP connections
Multiple Applications using different protocols (UDP,TCP)
, having different QoS requirements sharing the aggregated
bandwidth.
Multiple users sharing the aggregated bandwidth.
Effect of incorrect estimation of the delay and bandwidth
on EDF.