A Feedback Based Scheme For Improving TCP Performance in AdHoc Wireless Networks

1
A Feedback Based Scheme For Improving TCP
Performance in Ad-Hoc Wireless Networks

• Chandran et al
• UT Dallas

2
What are Ad-Hoc Networks?

• Set of mobile hosts that communicate with each
  other without the support of fixed infrastructure
  like base stations
• Possible uses Battlefields, Rescue missions
• Also known as MANETs (Mobile Ad-Hoc Networks) or
  Multihop wireless networks
• Open areas of research QoS, Multicast routing

3
What is the problem with MANETs?

• We would like to use transport layer protocols
  over such networks obvious choice is TCP
• But MANETs are prone to sudden route changes and
  disconnections
• Physical medium is also wireless
• Results in LOTS of packet losses
• TCP does not fly if packet losses are caused by
  anything other than congestion

4
TCP-F (TCP with feedback)

• Uses explicit route failure notification to
  inform the source of route failures
• Assumptions
• Wireless links are bi-directional
• Reliable link layer protocol
• Suitable routing protocol
• Finite time within which routes are
  re-established
• All packets carry source and destination ID
• Connection is unidirectional

5
The Details
Source
Dest
6
The Details
Source
FP
Dest
7
The Details
Source
RFN
FP
Dest
8
The Details
Source
Discard RFN
Reroute packets
FP
Dest
9
The Details
RFN
Source
Dest
10
The Details
Zzz...
Source

• Stop sending further packets
• Freeze all timers, window
• Starts a route failure timer

Dest
11
The Details
RRN
Source
Dest
12
The Details
Yawn...

• Restarts timers from frozen values
• Resumes transmission

Source
Dest
13
Simulation

• Very simple event driven simulation
• Treat the Ad-Hoc network as a black box with 10
  hops
• Randomly generate route failures
• Route re-establishment time and frequency of
  failures are parameters of simulation
• No congestion

14
Results
15
Results
16
Results
17
Discussions

• Multiple failures along the same route?
• Congestion?
• How to inform multiple sources affected by a
  route failure?
• Notify only when the FP receives a packet from a
  source

18
Future extensions

• Buffering of packets at intermediate nodes
• Save retransmissions
• Expected overhead at intermediate node is low
• RRN can be used to flush buffered packets
• Selective acknowledgement policies might have
  better performance (SACK TCP)

19
Analysis of TCP Performance over Mobile Ad Hoc
Networks

• Gavin Holland and Nitin Vaidya
• TAMU

20
What is this paper about?

• Better simulation platform
• Extensive study of TCP over MANETs
• Study of TCP with Explicit Link Failure
  Notification (ELFN) techniques

21
Simulation Environment

• ns with extensions from MONARCH project at CMU
• routing and 802.11 MAC layer
• ARP
• Node mobility modeled using pre-computed mobility
  patterns
• TCP Reno

22
Dynamic Source Routing (DSR)

• Source specified path
• Path obtained by route discovery
• Route discovery initiated by broadcasting a route
  request
• Route request is propagated through the network
  until it reaches a node that knows of a route to
  the destination

23
DSR

• Route reply is sent back to requester with the
  new route concatenated with the source route in
  the request
• TTL is used to ensure that request does not
  travel far
• Every request has a unique request id used to
  ensure no loops
• Routes are cached at intermediate nodes
• Cache updation is done via route error messages
  that are generated when a packets source route is
  invalid

24
More DSR!

• Packets can be salvaged by a node that detects
  that the next link is broken and if it knows of
  another route to destination
• CMU extensions have two phases
• Local broadcast
• Propagating ring search

25
Performance metric

• TCP throughput
• Upper bound on TCP throughput Expected
  throughput
• First compute throughput versus number of hops
  for fixed wireless hosts in a linear chain
• Ti is the throughput when the number of hops is i.

26
Static throughput
27
Expected throughput
ti is the time for which the shortest path from
the sender to the receiver has i hops
28
Effect of node speed on throughput

• For a given mobility pattern, expected throughput
  is independent of speed (Why?)

• Throughput can increase or decrease!
• Mostly throughput degrades as speed is increased
• Sometimes, throughput can increase

29
Speed effects
30
Mobility induced behaviors

• Some mobility patterns yield very low throughput

31
Anomaly

• Increase in throughput with speed increase

32
Observations

• Characteristics of routing protocol have
  significant effects on TCP performance
• Routing protocol that responds faster to mobility
  induced changes improves TCP performance
• Most problems with DSR are due to stale cache
  entries
• Solution turn off caching! Improves performance
  for a single connection

33
TCP with ELFN

• Piggybacked TCP/IP headers on DSR route failure
  messages
• When TCP sender receives ELFN, it enters standby
  mode
• Packet sent at periodic intervals to see if a
  route has been re-established

34
Results
35
Effect of varying time between probe packets

• The smaller the gap between packets, the better
  the throughput

36
Effect of modifying RTO and window size upon
restoration
37
Effect of choosing different probe packets

• Intuitive - Send the first packet in window
• Optimistic - Send the packet with lowest sequence
  number among the packets signaled as lost
• No significant difference observed ?
• Routes once broken, are rarely restored quickly

38
Conclusions

• TCP throughput drops significantly when node
  movement causes link failures
• Expected throughput is a better means of
  comparing throughput
• ELFN techniques improve TCP performance
• Study performance in congested networks