CS244a: An Introduction to Computer Networks

1
Simulation and Exploration of RCP in the
networks EE384Y Presentation, May 2003
Changhua He Department of Electrical Engineering
Stanford University changhue_at_stanford.edu
2
Outline

• Introduction to RCP
• Rate Control Protocol
• RCP vs. TCP
• Problem Statement
• Simulation results
• Modeling Analysis
• Conclusion

3
Problem Statement

• Queues with M/Pareto Traffic
• Interested flows Background flows
• Performance measured only by flow duration (flow
  response time)
• In this project, try to
• Explore how flow duration relates to flow rate
  and average load in networks

4
Results (case 1)

• Flow duration relates to
• Transmission rate
• Load of interested flows

5
Brief Summary

• Optimal transmission rate exists
• Depends on background load queue length
• Also on Interested load queue increasement
• In networks
• Scalable
• Rate determined by bottleneck link
• Why this way?
• Only consider single link, no background traffic

• Note transmission rate of background flows less
  than interested flows in bottleneck link

6
Modeling

• Recall the queue
• Flow arrival
• Flow size
• Transmission rate
• Average load
• Consider Queue size

7
Analysis(1)

• M/G/1
• Pollaczek-Khinchine Formula
• By PASTA

8
Analysis(2)

• Flow-based FCFS (Reference Model)
• Same queue size with our system
• Same departure with M/G/1

9
Analysis(3)

• (contd)

10
Analysis(4)

• What about ?
• Much more difficult
• Consider
• Compare with n parallel FCFS queues and M/G/n
  

11
Analysis(5)

• No solution for M/G/n
• Upper bound of queue size

12
Analysis(6)

• Proof

13
Conclusion

• RCP works well in networks
• Rate depends on background loadincoming load
• Scalable, not sensitive to topology
• Theoretical analysis
• Computes queue length for
• For Pareto with shape between 1 and 2, infinity
• Proves

14
Future work

• Queue length with ?
• Nothing about flow duration
• We have first packet delay
• What influences last packet delay?

15
Basic Algorithm of RCP
16
Results (case 1)

• Optimal point appears at higher rate

17
Results (case 2)

• Curve shift upwards
• Scalable, bottleneck determines flow duration

18
Results (case 3)

• Not sensitive to topology