Challenges to Reliable Data Transport Over Heterogeneous

Wireless Networks

Motivation (Ch 1+2)

• Everybody went nuts about wireless (cell phones, etc) and the data networks (the internet) in the 90's

• Then, why are wireless networks not more popular?– Is there no demand?• No

Then, why are wireless networks not more popular?

• Poor performance

• Too large a difference from wired technology

Heterogeneity

• Makes it difficult to identify performance bottlenecks

Three Fundamental Challenges

• Wireless bit errors– TCP assumes losses are due to congestion

• Asymmetric effects and latency variation– TCP relies on consistent rtt's for good

performance

• Low channel bandwidths– Long range channels are often orders of

magnitude slower than the wired alternative

Split-Connection Protocols

• Put a layer under tcp that is error free– Now losses are due to congestion– Asymmetric rtt's lead to poor performance

Wireless Testbed (ch 3)

Simulation Environment

• Initially used REAL– Realistic TCP modules– Inflexible– Written in C with parts in assembler– Hard to extend– Simulation written in propriety script

language

• Now use NS-2

NS-2

• Added LAN object– Formerly only point-to-point link

• Error Models

• Tested on real wireless network to determine error behaviour

BARWAN

• WaveLan– 2Mb/s DS• Throughput between 50k and 1.5M• Usually closer to the low end

• Ricochet– Half-duplex FH

• Cable– 10Mb/s shared up, dialup down

Measurement Techniques

• Wrote netperf– Measures TPC workloads

• Tcpdump– Detailed packet traces

Performance Metrics

• Throughput– Received bytes /unit time

• Goodput– Ratio of useful bytes to number transmitted– Always < 1, closer to 1 - more efficient

• Utilization– How often contended resource is idle

• Fairness– How evenly shared, Jan's fairness index

Jan's Fairness Index

• n connections

• xi = throughput for node I

• f = (xi)2/(nxi

2)

Berkeley Snoop Protocol (Ch 4)

• Significantly improves TCP performance in error-prone cellular networks

• Uses cross-layer protocol optimisations

Topology

• End node(s) connected to Base station via wireless link

• Rest of hops over wired network

• Using TCP Reno a bit error rate of 5% makes a transfer take 4.5 times longer than ideal TCP(2MB transfer)

Extra layer

• Transfer to– Agent at base station

• Uses info in ACKs

• Soft state

• Transport aware link protocol

• Transfer from– Explicit loss notification

• Retransmits lost packets

• No congestion control

• Link aware transport protocol

Design Goals

• Local solution– Transparent to fixed internet host

• Eliminate adverse interaction between layers

• Enable incremental deployment

• Preserve end-to-end semantics

• Use soft state

Transfer From a Fixed Host

• Caches data to be forwarded to MH

• ACKs are forwarded to fixed host if not due to loss– Duplicate ACKs can mean loss• Packet is resent with high priority

• DupACKs after first not forwarded

Transfer From Mobile Host

• Negative ACKs– Built on SACKs

• Dependant on SACK implementation

– Not used

• ELN– BS keeps list of “holes”

• Hole is set only when BS is not receiving close to max # of packets

– If DupACK corresponds to hole ELN bit is set

Mobility

• Handoffs can lead to packet loss

• Multi-cast based buffering– Intermediate “home” agent does snoop and

sends to each base-station

Performance

Asymmetry

• ACK speed on slow link limits throughput on fast link– Compress ACKs– Reduce ACK frequency

Small Windows

• Fast retransmissions are infrequent

• Most due to timeouts– Results in idle channel

• Usually fix with SACKs and ELN

• ER (Enhanced Recovery)– Probe network after <3 Duplicate ACKs