On the Design of Load Factor based Congestion Control Protocols for - - PowerPoint PPT Presentation

On the Design of Load Factor based Congestion Control Protocols for Next-Generation Networks

Ihsan Ayyub Qazi Taieb Znati Department of Computer Science University of Pittsburgh, PA, USA

Email:{ihsan,znati}@cs.pitt.edu

Agenda

 Motivation  Existing Solutions  Goals of our Work  Insights from Analysis  Design of our Protocol  Conclusion

2 {ihsan,znati}@cs.pitt.edu

Motivation

 Trends in Technology

• increases in link capacities (e.g. 40+ Gbps links)

and delay (e.g. satellite, wireless WANs)

• proliferation of real-time applications (e.g. Skype,

iTunes)

 Issues with TCP

• reacts adversely to increases in capacity and/or

delay

• introduces large delay and throughput variations
• may lead to the situation where most Internet

traffic does not use TCP

3 {ihsan,znati}@cs.pitt.edu

TCP: Signal of Congestion + Window Update Policies

Packet loss (e.g., used by TCP Reno, NewReno,

SACK etc)

• Binary + implicit signal of congestion
• Self-induced packet losses
• Large average queuing delay

Window Update Policies

• Conservative increase (AI: 1 pkt/RTT)
• takes a long time to acquire bandwidth on high BDP

paths

• Aggressive decrease (MD: ½ after packet loss)
• induces large variations in delay and throughput

4 {ihsan,znati}@cs.pitt.edu

Existing Solutions

End-to-End (E2E) schemes with Implicit Feedback

e.g. TCP

• have fundamental performance limitations

Network-based schemes e.g. XCP

• can achieve optimal performance
• high per-packet overhead may constraint router

speed

• high deployment barrier

E2E schemes with Explicit Feedback e.g.

TCP+AQM/ECN, VCP

• perform better than E2E schemes (implicit feedback)
• involves changes at the end-hosts with incremental

support from the routers

5 {ihsan,znati}@cs.pitt.edu

E2E schemes with Explicit Feedback

VCP [Xia, SIGCOMM 05]

• Uses two bits for the feedback
• Slow convergence to efficiency (MI Factor = 1.0625)

– Makes short flows to last longer

• Unfairness in the presence of diverse RTT flows
• Reduced responsiveness to congestion

{ihsan,znati}@cs.pitt.edu 6

Sender Receiver

AI MD VCP Control MI

10 10 ACK

Load

1.0 0.0

Code (11) (10) (01)

0.8

low -load high -load

• verload

Goals of our Work

 Analyze the tradeoff between increasing the amount of feedback information and the resulting performance improvement  Analyze window increase/decrease policies for achieving efficiency and fairness on high BDP paths, while keeping low queues and negligible packet drop rates

7 {ihsan,znati}@cs.pitt.edu

Insights from Analysis

 Convergence to efficient bandwidth allocation

• 3-bit feedback sufficient for near-optimal

performance

• performance follows the law of diminishing returns

 Convergence to fair bandwidth allocation

• using multiple levels (e.g. 8-levels) for the

multiplicative decrease policy enables the protocol to adjust its

• convergence rate to fairness, rate variations and

responsiveness based on the degree of congestion at the bottleneck.

8 {ihsan,znati}@cs.pitt.edu

Convergence to Efficiency

9

Key Observation

• increasing the feedback information beyond 3

bits improves only a small part of the overall time

Helps flows to finish quicker

{ihsan,znati}@cs.pitt.edu

Convergence to Efficiency

10

Key Observation

• increasing the feedback information beyond 3

bits improves only a small part of the overall time

Helps flows to finish quicker

{ihsan,znati}@cs.pitt.edu

Convergence to Efficiency

11

Key Observation

• increasing the feedback information beyond 3

bits improves only a small part of the overall time

Helps flows to finish quicker

Pareto Distributed File Sizes Average File Size: 30KB Link Capacity: 10Mbps RTT: 200ms

{ihsan,znati}@cs.pitt.edu

Convergence to Fairness

12

 In AIMD-like protocols, the following are intrinsically tied

• convergence rate to fairness (CF)
• flow rate variations (RV)
• responsiveness to congestion (RC)

 High CF  high RV  Small FV low RC and CF  Key Observation

• single value of beta (MD) not

suitable

• optimize for the target situations
• small beta when new flows

arrive

• high beta in steady-state

{ihsan,znati}@cs.pitt.edu

MLCP: Control laws

{ihsan,znati}@cs.pitt.edu 13

Symbol

95% 100% 0001 0010 0011 0100 0110 0111 1000 0% 80% 16% 32% 48% 64% 120% 1111 1001 ... 101% …

Multiplicative Increase Additive Increase Multiplicative Decrease Inversely Proportional Increase

Parameters load

Control Laws

14

Fairness in the presence of diverse RTT flows

{ihsan,znati}@cs.pitt.edu

 MI and AI parameters are scaled to handle RTT heterogeneity

Control Laws

15

Fairness in the presence of diverse RTT flows

{ihsan,znati}@cs.pitt.edu

 MI and AI parameters are scaled to handle RTT heterogeneity

Evaluation

 Extensive ns-2 simulations  200Mbps, 80ms, 10 forward flows and 10 reverse flows

{ihsan,znati}@cs.pitt.edu 16

forward traffic reverse traffic

Src 1 Src 2 Src N Dst 1 Dst 2 Dst N

……. …….

MLCP: Achieves high efficiency

17 {ihsan,znati}@cs.pitt.edu

MLCP: Minimizes queuing delay and loss

rate

{ihsan,znati}@cs.pitt.edu 18

MLCP: Fairness + Smooth rate variations

{ihsan,znati}@cs.pitt.edu 19

Ongoing and Future Work

MLCP requires 4-bits of congestion-related feedback, however, only two bits are allocated for ECN in the IP header We are analyzing a 1-bit packet marking scheme called Adaptive Deterministic Packet Marking (ADPM) to convey high resolution congestion price estimates1

20 {ihsan,znati}@cs.pitt.edu

1 Lachlan L. H. Andrew, Stephen V. Hanly, Sammy Chan and Tony Cui,

“Adaptive Deterministic Packet Marking,” IEEE Comm. Letters, 10(11):790-792, Nov 2006.

Conclusion

 We analyzed the tradeoff between increasing the feedback information and the resulting performance improvement  Based on the insights, we designed and implemented a low-complexity protocol, MLCP

 MLCP

• Achieves high utilization on high BDP paths, maintains low

queues and negligible packet loss rate

• Helps flows to finish quicker than 2-bit schemes such as

VCP

• Ensures fairness in the presence of diverse RTT flows
• Enables the sources to adjust convergence to fairness, rate

variations and robustness based on the degree of congestion at the bottleneck

21 {ihsan,znati}@cs.pitt.edu

Thank You!

22 {ihsan,znati}@cs.pitt.edu