Processor Sharing Flows in the Internet

1
Processor Sharing Flows in the Internet
Nadita Dukkipati Masayoshi Kobayashi Rui
Zhang-Shen Nick McKeown
IWQoS 2005
By
Presenter
Manfred Georg
2
Note Presentation Content

• When this presentation was given, most of the
  figures displaying the motivation for RCP and the
  performance results were taken from the following
  talk given by Nadita Dukkipati.
• "A Clean Slate Design of Internet's Congestion
  Control Algorithm,"Caltech Lunch Bunch Seminar,
  8 November 2005
• http//yuba.stanford.edu/rcp/rcp-caltech.pdf

3
Overview

• Congestion Control
• End-to-End
• Router assisted
• Processor sharing
• RCP
• Description
• Performance Analysis

4
Congestion Control

• Efficiently and fairly use network
• End to End
• Simple (end host) implementation
• Scalability problem
• TCP
• Slow start
• Packet loss
• Flow finished before convergence

5
Congestion Control

• Router Assisted
• Routers use and/or communicatecongestion control
  information
• Per flow state
• Queue for each flow
• Per router state
• XCP
• Too conservative allocation
• Per packet computation

6
Processor Sharing

• One rate for all flows

7
RCP

• Heuristic algorithm
• Flow rate calculated from queue occupancy and
  aggregate input rate
• Each router assigns single rate to all flows
• No per-flow state
• No per-packet calculations (count?)?

8
RCP Algorithm

• Fair rate R(t), update every RTT
• Give R(t) to every flow unless they already have
  a lower rate
• Source transmits at lowest offered rate
• Router updates based on

9
Performance Analysis
10
Conclusion

• TCP better than XCP?
• Flow completion time
• RCP emulates processor sharing
• Every flow receives same rate
• Fast convergence to fair rate
• Low flow completion time
• Stable operation

11
Thank You
Questions?