An End-to-end Approach to
Increase TCP Throughput Over
Ad-hoc Networks
Sarah Sharafkandi and Naceur Malouch
Introduction

TCP is designed for wired networks


Congestion control : window-based
With IEEE 802.11 PHY & MAC, TCP over Ad-hoc has a
low performance:

congestion control and not “collision” control:




TCP react to buffer overflow
 "bursty" traffic
inherent reverse traffic
Objective: Improve TCP throughput without modifying
PHY, MAC and NET layers.
2
When collision causes DATA loss?

By hidden nodes:
packets sent by D
collide with A’s
packets at node B
preventing B from
decoding A’s
packets.

By repetitive retries due to “ordinary” collisions: it happens
when   C*  rare event
By buffer overflow : due to increased waiting times  not
considered in this work

3
State of the art

Distributed Link RED and Adaptive pacing [Fu et al.
INFOCOM’2003]

If the average number of retransmission retry > min_thresh :




 Improvement: 10%-30% for the chain topology
Increasing retry limit and optimum packet size [Jiang et al.
DISCEX’O3]



early drop of packets
increase the backoff period
Increasing the retry limit reduces oscillations in the
instantaneous thpt
Increasing the packet size increases the thpt till some thresh
Improving TCP throughput using Delayed ack method
[Altman et al. MADNET’03]

delayed ack factor = 2, 3
4
An end-to-end approach to “collision
control” ?!
5
Simulation Scenario





NS2 network simulator
Chain topology
The source and destination at both ends of the
chain
AODV as a routing protocol
Some modifications to the source code of NS2:



delayed ack > 2
monitoring without file traces
token bucket: packet version
6
TCP Sends the packets in “burst”
 Two
experiments to show the effect
of “burstiness”
Simulation with TCP using RFC3465
 Simulation with CBR traffic

7
Simulation with TCP using RFC3465

The “burstiness” of RFC3465 results in throughput
reduction despite the gain in the window growth
8
Simulation with CBR traffic: Results

i CBR traffics with rate r/i, i = 1, 2, 3, 4.

Best result is when there is packet spacing
 “burstiness” is minimum
9
New approach


Bursty data traffic over Ad-hoc networks results
to performance reduction
Shaping :



Controls the rate of releasing packets to the network
No more aggressive traffic
Plus delayed ack  approaches the optimal channel
reuse
10
Throughput of TCP with shaper and delayed ack

Shaper increases the TCP throughput by 53%-120%
11
Shaper and Delayed ack

Shaper allow delayed ack mechanism to bypass the limit of
d=3 
12
Optimum rate

There is always an optimum rate for the shaper in which
TCP has the best performance
13
TCP throughput as a function of Number
of hops

Optimum rate decreases when number of hops
increases
14
Impact of bucket size


A data can pass through the shaper only if it can get a token
from token buffer.
We can use it to test again the effect of burstiness
15
Tokens

Again allowing “burstiness” results to throughput
reduction
16
Effectiveness of Shaping in presence
of CBR Traffic


Network scenario :
 same source/destination for UDP traffic
 UDP share all the ad-hoc routers with TCP
Compute the gain while increasing the rate of UDP:
17
Conclusion

TCP throughput drops significantly because of:




link contention caused by hidden terminal problem
An "aggressive“ TCP sender causes an increased contention at
the MAC layer
Implementing a shaper at the sender improves
TCP throughput by controlling the aggression of
TCP data traffic
Delayed ack mechanism plus the shaper
→ increase spatial channel reuse
18
Future work

An adaptive algorithm for finding the optimum
rate


difficulties: convergence and stability
Related work: [ElRakabawy et al. MobiHoc’2005]


same idea: end-to-end solution
BUT :



change TCP protocol for the multihop wireless ad-hoc
based on the esimation of the 4-hop transmission delay
Our approach :
19