Improving TCP Performance in 802'11 Multihop Ad Hoc Networks

1
Improving TCP Performance in 802.11 Multihop Ad
Hoc Networks

• Xu Yuedong (ANSR_at_CSE)
• December 12, 2006

2
Outline

• Problem Overview
• Related Works
• DCC A Cross-Layer Design
• DTCP A Utility Based Algorithm
• Future Works

3
Overview

• TCP is designed for wired networks
• A window-based reliable end-to-end transmission.
• Reliable physical Layer with fixed bandwidth.
• Large bandwidth delay product (BDP).
• TCP over 802.11 multihop ad hoc networks
• Relatively unreliable physical layer. (SINR,
  Mobility, etc.)
• An interference limited, shared medium. (CSMA
  MAC)
• Very small bandwidth delay product (BDP).

Objective Improve TCP Throughput Fairness in
Multihop Ad Hoc Networks
4
Review of TCP Congestion Control

• Primary Goal
• Prevent the network from being overloaded
• TCP Congestion Control
• Slow Start
• Additive Increase Multiplicative Decrease
• Fast Recovery Fast Retransmission.
• A Brief Review of TCP AIMD

Note I will not go into the details of TCP
congestion control in Wired Networks.
5
Review of TCP Congestion Control

• AIMD Mechanism (Tahoe/Reno/Newreno)

6
TCP Performance in MANET

• Throughput Problems
• Throughput instability
• Transmission failure due to mobility noise
• Drastic thru. decrease when hop number increases

7
TCP Performance in MANET

• Throughput Problems

TCP over 802.11
TCP Throughput VS Hops
8
TCP Performance in MANET

• Fairness Problems
• Neighboring node one-hop unfairness
• Unfairness in symmetric topologies
• Unfairness among heterogeneous TCP Flows
• Short-term Long-term unfairness
• TCP in wireless cum wired networks

9
TCP Performance in MANET

• Neighboring node one-hop unfairness
• 2 TCP Connections
• First session starts at 10.0s ( 6 ? 4 )
• Second session starts 20.0s later ( 2 ? 3 )

10
TCP Performance in MANET
First session start
Second session start
11
TCP Performance in MANET

• What are the reasons? and the Solutions?
• Mobility induced packet loss and routing failure
• TCP sender differentiates the congestion induced
  packet losses from those caused by routing
  failure and transmission error.
• The router can be a bit more smart, i.e. dont
  flush the packets when links is broken, or send
  back the routing failure information. We can also
  modify the TCP retransmission timeout in order to
  avoid long freezing time.

12
TCP Performance in MANET

• What are the reasons? and the Solutions?
• The 802.11 MAC contentions, i.e. hidden
  terminals

Cs transmission range
F
E
D
B
C
Congestion
Collision at Node B
Ds interference range
13
TCP Performance in MANET

• What are the reasons? and the Solutions?
• The aggressiveness of TCP AIMD.
• Limit the TCP aggressiveness using various
  methods.
• Contention loss of TCP data sub-flow and ack
  sub-flow.
• Use delayed acknowledgement method to reduce the
  contention loss of data packets and ack packets.
• Some other problems as well as their solutions.

14
Our WorksUtility Based Congestion Control

• Prove that an important algorithm in Infocom03 is
  wrong. Propose a distributed congestion control
  (DCC) algorithm for multihop networks.
• Propose a delay based congestion control
  mechanism, namely DTCP that controls TCP-like
  traffic in multihop ad hoc networks. Present an
  alternative implementation of DTCP in the MAC
  layer.

Our research are the first practical utility
based solutions to improve TCP performance in
multihop ad hoc networks
15
Outline

• Problem Overview
• Related Works
• DCC A Cross-Layer Design
• DTCP A Utility Based Algorithm
• Future Works

16
Classification

• Fruitful research efforts
• Quite a few papers in good conf./Journals since
  1998.
• Layered Design
• Find efficient solutions in the transport layer
  or MAC layer independently.
• CWL, FEW, DACK
• Cross Layer Design
• Learn how to adjust the congestion window, RTO or
  acknowledgement based on the cross-layer
  information.
• Link RED, TCP-AP, ATP, AD-TCP, TCP-ELFN

17
Related Papers
18
Comments

• Problems of previous research
• Existing solutions are insufficient to balance
  throughput and fairness in multihop ad hoc
  networks.
• Previous algorithms concentrated on long-term
  fairness instead of short-term fairness. While
  long-term fairness conceals the unfairness nature
  of wireless ad hoc networks

19
Outline

• Problem Overview
• Related Works
• DCC A Cross-Layer Design
• DTCP A Utility Based Algorithm
• Future Works

20
LRED Algorithm

• Link RED (LRED) Infocom03, TMC05, Cited by 149
  times (Google Scholar)
• Use average number of MAC retries to determine
  ECN prob.
• Introduce Adaptive Pacing of CSMA.
• Algorithm Specification (1)

21
LRED Algorithm

• Algorithm Specification (2)
• If avg_retry lt min_th
• Random Backoff
• elseif avg_retry gt min_th
• Backoff ? RandomBackoff
  ExtraBackoff
• endif
• Improve throughput and fairness!

22
LRED Algorithm
ECN Binary feedback
23
LRED Algorithm
24
DCC Algorithm

• A Counter example
• Collision probability of a meek flow is not less
  than that of an aggressive competing flow.
• Link RED penalizes the wrong flow by marking the
  meek flow with a larger probability.
• Most of previous researches neglect short-term
  fairness.

Cross Topology
Contention Graph
25
DCC Algorithm

• Under certain assumptions, we prove
• Theorem 1 If the airtime of link 1 (x1) is
  larger than that of link 5 (x5), the collision
  probability of link 5 (?5 ) is not less than that
  of link 1 (?1 ), and vise versa. Formally, we
  have
• Congestion Index We define a novel congestion
  index for link ,

26
Evaluation of Congestion Index

• We run fixed-rate TCP to evaluate Congestion
  Index in
• Light traffic in cross-chain network (2Mbps link
  capacity)
• Flow A ? E 100Kbps Flow F ? I 50Kbps

27
Evaluation of Congestion Index

• We run fixed-rate TCP to evaluate Congestion
  Index in
• Heavy traffic in cross-chain network (2Mbps link
  capacity)
• Flow A ? E 200Kbps Flow F ? I 100Kbps

28
The Model

• We formulate the TCP congestion control as a
  unconstraint optimization for four reasons.
  (omitted)
• where and are constant parameters,
  is the mark probability of the
  flow.
• Defining the mark probability of flow in link
  to be ,
• we obtain the mark probability of flow ,

29
The Model
Calculate mark probability of a link
Tl Airtime of link ,
Collision prob. Proportional
controller Coefficient Constant
Based on the sub-gradient method, we have
30
Distributed Algorithm
Theorem 2 Suppose all flows in the networks
have concave, differentiable, strict increasing
utility function U(r) for r gt 0, the fully
distributed congestion control algorithm
converges to the unique equilibrium point r
for some positive constants and .
31
Distributed Algorithm
Distributed Algorithm 1). We adopt rate based
TCP instead of Window based TCP. 2). Algorithm
Description (Updating sending interval Tintv)
If an ECN echo is received, If
a received ack is not ECN echoed,
32
Performance Evaluation (1)

• Chain Topology
• Eliminating the throughput instability.
• High throughput over long chain topologies.
• Robustness of parameter configuration.

33
Performance Evaluation (1)

• Chain Topology
• Eliminating the throughput instability.
• High throughput over long chain topologies.
• Robustness of parameter configuration.

34
Performance Evaluation (1)

• Chain Topology
• Eliminating the throughput instability.
• High throughput over long chain topologies.
• Robustness of parameter configuration.

35
Performance Evaluation (2)

• Cross-Chain Topology
• Achieve Short-term fairness
• The throughput is still larger than standard
  802.11

36
Performance Evaluation (2)

• Cross-Chain Topology
• Achieve Short-term fairness
• The throughput is still larger than standard
  802.11

37
Performance Evaluation (2)

• Cross-Chain Topology
• Achieve Short-term fairness
• The throughput is still larger than standard
  802.11

38
Performance Evaluation (3)

• Grid Topology
• Achieve both long-term and short-term fairness
• A very high instantaneous fairness index for
  competing flows

39
Performance Evaluation (3)

• Grid Topology

40
Performance Evaluation (3)

• Grid Topology

41
Outline

• Problem Overview
• Related Works
• DCC A Cross-Layer Design
• DTCP A Utility Based Algorithm
• Future Works

42
Motivations

• Queuing delay increases when channel is
  congested.
• TCP source adapts its sending rate according to
  average queuing delays.
• Almost all the improved algorithm cannot compete
  with standard TCP Newreno. What if the standard
  TCP is not allowed to be changed or heterogeneous
  TCP flows coexist in the same wireless network?

43
Noncooperative Game Framework

• A general framework of congestion control for
  multihop ad hoc networks
• A non-cooperative game framework
• Maximize the utility of individual flows
• Find a good penalty function
• maximize
• where is the utility of flow
  , and g(xi) is the penalty function of flow i.

44
An Example

• A Delay Based Method
• The objective function is composed of two parts.
• a). Individual utility function
• b). Average k-hop forward queuing delay as a
  penalty

45
An Example

• A Delay Based Method
• Penalty Function (di is the average per-hop
  queuing delay)

• Why ?
• Avoid over-penalizing the long-hop flow
• Not good to accumulate Delay in different
  cells

46
An Example

• A Delay Based Method
• Required Information
• hop counts obtained from AODV messages
• forward queuing delay (Two methods)
• a) RFC1323 can record the total
  forward delay.
• b) Use RTT to estimate to queuing
  delay. (similar to TCP-AP algorithm)
• Forward transmission delay is known (hops
  per-hop
• transmission time)

47
An Example

• Delay Based TCP (DTCP)
• Utility Function
• Distributed rate adaptation

• Sending Interval

48
An Example

• Delay Based TCP (DTCP)
• A loose Upper-bound of the sending interval
• a). If the number of hops (h) is larger than 4,
• Upper-bound T4 s (T4 4 hop Trans.
  delay)
• b). If the number of hops (h) is less than 4,
• Upper-bound T1 s (T1 one hop Trans.
  delay)
• A loose Lower-bound of the sending interval
• Three exponential weighted moving average RTTs
  (3RTT)
• (not including of extreme large RTTs)

49
Performance Evaluation (1)

• Chain Topologies
• Eliminating the throughput instability.

50
Performance Evaluation (1)

• Chain Topologies
• High throughput over long chain topologies.

51
Performance Evaluation (1)

• Chain Topologies
• Robustness of parameter configuration.

52
Performance Evaluation (2)

• Chain Topology with Multiple Flows
• Flow A ? I and flow D ? F

53
Performance Evaluation (2)

• Chain Topology with Multiple Flows
• Flow A ? I and flow C ? G

54
Performance Evaluation (2)

• Neighboring Node One-Hop Unfairness

55
Performance Evaluation (2)

• Cross-Chain Topology
• Throughput and Fairness Comparison

56
Performance Evaluation (3)

• Large-Scale Grid Topology
• A grid lattice with 169 nodes
• Even number of flows
• Half of them are vertical flows
• The rest ones are horizontal
• Flows are evenly spaced

57
Performance Evaluation (3)

• Large-Scale Grid Topology

58
PQC Pacing Queue Controller

• An alternative implementation in the MAC IFQ
• If we are not allowed to modify TCP or
• No modifications to existing TCP, routing and
  802.11 CSMA protocols.
• Generally, the number of TCP sessions of a node
  is small in multihop ad hoc networks.

?????
Key Decision Can IFQ play the role of flow
control ?
59
PQC Pacing Queue Controller

• How to implement the PQC ?
• TCP protocol is merely used to guarantee the
  reliable end-to-end transmissions.
• The main function of TCP flow control is moved to
  the IFQ of the source nodes.
• The PQC architecture increases the complexity of
  MAC layer.

60
PQC Pacing Queue Controller

• Methodology
• We introduce another queue named Pacing Queue in
  the MAC layer.
• TCP data packets are stored in the pacing queue
  of the source nodes. After waiting for a
  pre-calculated time interval, the head of line
  (HOL) packet is send to the real interface queue
  (IFQ) for transmission.

61
PQC Pacing Queue Controller

• Methodology
• The relaying packets are directly put into the
  IFQ
• A packet table in the pacing queue is maintained
  for recording the departure of a data packet and
  arrival of its ack packet. (The packet table is
  extreme small due to small bandwidth delay
  product.)
• In order to avoid the overflow of the Pacing
  Queue, we introduce a simple scheduling
  (management) method to control the buffer size.

62
PQC Pacing Queue Controller
63
Outline

• Problem Overview
• Related Works
• DCC A Cross-Layer Design
• DTCP A Utility Based Algorithm
• Future Works

64
Future Works

• Cross-Layer Optimization for Utility Fairness in
  Wireless Networks.
• Multi-Channel Multi-Rate Multi-hop Routing.
• Cross-Layer Design of MAC and Congestion Control
  with Network Coding.

65
Questions ?
66
Thank You!