Modeling

1
Modeling Management of Bandwidth in a
Differentiated Services Network

• Ikjun Yeom
• Department of Electrical Engineering
• Texas AM University

2
Presentation Overview

• Introduction to DiffServ
• Modeling TCP throughput
• Aggregated Marking

3
Internet QoS

• What is QoS?
• Delay, delay jitter, bandwidth, drop rate, etc.
• Required for real-time/multimedia app.
• Best-effort service is not enough.

4
What is the DiffServ?

• A scalable solution to provide service
  differentiation in IP networks.
• Traffic conditioning, classifying, marking at the
  edges
• Providing different Per-Hop Behaviors (PHBs) in
  the core network

5
Overview of a DiffServ Network
6
Assured Forwarding (AF) PHB
7
AF Example
8
TCP Throughput Modelingin a DiffServ Network
9
Motivation

• Throughput of a TCP flow is not equal to its
  contract rate.

10
TSW Packet Marking
Upon each packet arrival
If avg_rate ? contract rate mark this packet
IN else mark this packet IN with probability
of contract rate/avg_rate
11
An Example of Packet Marking Behavior

• An individual TCP flow
• Contract rate 0.5 Mbps
• Packet size 1 KB
• Average RTT 0.14 sec.
• Average throughput 0.69 Mbps
• win_len in the marker is 1 sec.

12
(No Transcript)
13
Model for Packet Marking
14
Objective

• Develop a TCP throughput model as a function of
• Contract rate
• Probability of IN/OUT packet drop
• Packet size
• Round trip time

15
Assumptions

• No ACK drop
• TCP-Reno
• Random packet drop

16
Network Models

• Undersubscribed path
• No IN packet drops

Throughput
Contract rate
Time
17
Network Models (cont.)

• Oversubscribed path
• IN packet drops

Throughput
Contract rate
Time
18
Network Models (cont.)

• Normal-subscribed path
• Both IN and OUT packet drops

Throughput
Contract rate
Time
19
General TCP Model
20
A Model for Undersubscribed Path
21
A Model forOversubscribed Path
22
Model Validation
Prop. delay 180 msec. Contract rate Randomly
selected from 0 to 1 Mbps
Sender
Receiver
23
Results with 40 Mbps bottleneck
24
Results with 30 Mbps bottleneck
25
Results with 20 Mbps bottleneck
26
Results with 15 Mbps bottleneck
27
Relative Error
28
With Cross Traffic
29
Results with Cross Traffic
30
Merging Topology
31
Results with Merging Topology
32
Split Topology
33
Results with Split Topology
34
A Simplified Model

• For Undersubscribed Path

• For Oversubscribed Path

35
Observations

• Define excess bandwidth as
• throughput - contract rate

36
(No Transcript)
37
Aggregated Flows

• Single contract rate for all the flows in the
  aggregation.
• Packet marking based on
• Aggregate contract rate
• Agg. Sending rate
• Flows impact each other.
• The models are not directly applicable.

38
Models for Agg. Flows
39
Model Validation
Prop. Delay selected from 60 to 160
msec. Contract rate 10 Mbps
40
Results with 20 Mbps Bottleneck
41
Results with 30 Mbps Bottleneck
42
Summary

• TCP throughput can be estimated with models.
• Difficulty in sharing excess bandwidth.
• Model extended to aggregated flows.
• Needed to investigate marking scheme for
  aggregated flows.

43
Marking for Aggregated Flows
44
Motivation

• Customer marks packets of individual flows within
  SLA.
• Example marks packet of flows observing
  congestion at a higher rate.
• Approach employing per-flow state within the
  aggregation.

45
Customers Marking
46
System Model
47
Resource Management
?

• Marking rate Target rate
• Realized throughput ? Marking rate
• May waste resource (marking rate)
• Realized throughput gt Target rate
• Unachievable target rate

? Adaptive Marking
48
Adaptive Marking

• Guarantee at least one of the three
• Target rate
• Maximized throughput without IN drop
• Equally shared contract rate

49
Network Condition

• How to estimate the current condition?
• Achieved rate ? 0.75 Marking rate
• The flow experiences oversubscribed path.

50
Marking Rate Adaptation

• 0.75 Marking rate lt Achieved rate lt Target
• Increase the marking rate
• Achieved rate ? 0.75 Marking rate
• Reduce the marking rate
• Achieved rate gt Target
• Reduce the marking rate

51
Simulation
Cross traffic
M
R
52
Achievable Target
53
Unachievable Target
54
Simulation Results
55
Summary

• The adaptive marking manages resource
  effectively.
• Adapts to network condition to reach target rates
  of individual flows.

56
Conclusion

• Provided analytical TCP model for DiffServ
  networks.
• Showed that excess bandwidth cannot be shared
  equally.
• Flows with larger contract rate may not reach
  their contract rate.
• Provided an adaptive marking that utilizes
  resource effectively.