Multimedia-unfriendly TCP Congestion Control and Home...

19
Multimedia-unfriendly TCP Congestion Control and Home Gateway Queue Management Lawrence Stewart α , David Hayes α , Grenville Armitage α , Michael Welzl β , Andreas Petlund β α Centre for Advanced Internet Architectures (CAIA), Swinburne University of Technology β Department of Informatics, University of Oslo

Transcript of Multimedia-unfriendly TCP Congestion Control and Home...

Page 1: Multimedia-unfriendly TCP Congestion Control and Home …web.cs.wpi.edu/~claypool/mmsys-2011/Day1-4_Multimedia... · 2011. 3. 9. · Multimedia-unfriendly TCP Congestion Control and

Multimedia-unfriendly TCP Congestion Control and Home Gateway Queue

Management

Lawrence Stewart α, David Hayes α, Grenville Armitageα, Michael Welzlβ, Andreas

Petlundβ

α Centre for Advanced Internet Architectures (CAIA), Swinburne University of Technology

β Department of Informatics, University of Oslo

Page 2: Multimedia-unfriendly TCP Congestion Control and Home …web.cs.wpi.edu/~claypool/mmsys-2011/Day1-4_Multimedia... · 2011. 3. 9. · Multimedia-unfriendly TCP Congestion Control and

http://caia.swin.edu.au 23 February 2011 Page 2 ACM Multimedia Systems 2011

Outline 1.  Introduction

2.  Results

3.  Conclusion

Page 3: Multimedia-unfriendly TCP Congestion Control and Home …web.cs.wpi.edu/~claypool/mmsys-2011/Day1-4_Multimedia... · 2011. 3. 9. · Multimedia-unfriendly TCP Congestion Control and

http://caia.swin.edu.au 23 February 2011 Page 3 ACM Multimedia Systems 2011

Detailed outline (section 1 of 3) 1.  Introduction

2.  Results

3.  Conclusion

1.  Introduction •  Motivation •  Not All Queues Are Created Equal •  TCP CC (Congestion Control) Recap •  Methodology

Page 4: Multimedia-unfriendly TCP Congestion Control and Home …web.cs.wpi.edu/~claypool/mmsys-2011/Day1-4_Multimedia... · 2011. 3. 9. · Multimedia-unfriendly TCP Congestion Control and

http://caia.swin.edu.au 23 February 2011 Page 4 ACM Multimedia Systems 2011

Motivation  Gateway devices often have excessive drop-tail queue buffering1,2

 TCP CC algorithms each interact differently with network buffers  Real-time multimedia traffic sharing the gateway bottleneck

experiences collateral damage: latency/jitter and loss  How does choice of CC algorithm and queue implementation impact

the inflicted collateral damage? 1M. Claypool, R. Kinicki, M. Li, J. Nichols and H. Wu, “Inferring Queue Sizes in Access Networks by Active

Measurement”, PAM Workshop '04 2M. Dischinger, A. Haeberlen, K. P. Gummadi and S. Saroiu, “Characterizing residential broadband networks”,

SIGCOMM IMC '07

Page 5: Multimedia-unfriendly TCP Congestion Control and Home …web.cs.wpi.edu/~claypool/mmsys-2011/Day1-4_Multimedia... · 2011. 3. 9. · Multimedia-unfriendly TCP Congestion Control and

http://caia.swin.edu.au 23 February 2011 Page 5 ACM Multimedia Systems 2011

Not All Queues Are Created Equal  When is a drop-tail queue full?

Packet/Slot based drop if occupancy ≥ qlen Fixed-byte strict drop if occupancy + pkt ≥ qlen Fixed-byte loose drop if occupancy > qlen

Page 6: Multimedia-unfriendly TCP Congestion Control and Home …web.cs.wpi.edu/~claypool/mmsys-2011/Day1-4_Multimedia... · 2011. 3. 9. · Multimedia-unfriendly TCP Congestion Control and

http://caia.swin.edu.au 23 February 2011 Page 6 ACM Multimedia Systems 2011

TCP Congestion Control Recap  NewReno is ailing defacto standard (high latency/speed, wireless  Many new algorithm proposals: Compound, CUBIC, HTCP, ...  BSD still uses NewReno  Linux uses CUBIC  Windows Vista / 7 have Compound TCP available for testing

Page 7: Multimedia-unfriendly TCP Congestion Control and Home …web.cs.wpi.edu/~claypool/mmsys-2011/Day1-4_Multimedia... · 2011. 3. 9. · Multimedia-unfriendly TCP Congestion Control and

http://caia.swin.edu.au 23 February 2011 Page 7 ACM Multimedia Systems 2011

TCP Congestion Control Recap Downstream NewReno Behaviour

Congestion window oscillation coupled with induced queuing delay. ADSL 1 type link: 1500/256 kbps, min 100ms RTT, 20 000 byte queue.

Page 8: Multimedia-unfriendly TCP Congestion Control and Home …web.cs.wpi.edu/~claypool/mmsys-2011/Day1-4_Multimedia... · 2011. 3. 9. · Multimedia-unfriendly TCP Congestion Control and

http://caia.swin.edu.au 23 February 2011 Page 8 ACM Multimedia Systems 2011

Methodology   Testbed & NS-2   1500/256kbps & 24/1Mbps

(ADSL 1 & 2 speeds)   Byte-based [10k, 100k] &

slot-based [7, 67] queues   RTTs of 24, 50, 100 and 200ms   Dummynet for bandwidth, queue size &

RTT emulation   Single bulk TCP flow A->C   186 byte, N(µ=20ms, σ =1ms) CBR UDP

flow B<->D

Testbed

Page 9: Multimedia-unfriendly TCP Congestion Control and Home …web.cs.wpi.edu/~claypool/mmsys-2011/Day1-4_Multimedia... · 2011. 3. 9. · Multimedia-unfriendly TCP Congestion Control and

http://caia.swin.edu.au 23 February 2011 Page 9 ACM Multimedia Systems 2011

Detailed outline (section 2 of 3) 1.  Introduction

2.  Results

3.  Conclusion

2.  Results •  Goodput •  Collateral Damage to Voice Stream – Latency •  Collateral Damage to Voice Stream - Loss

Page 10: Multimedia-unfriendly TCP Congestion Control and Home …web.cs.wpi.edu/~claypool/mmsys-2011/Day1-4_Multimedia... · 2011. 3. 9. · Multimedia-unfriendly TCP Congestion Control and

http://caia.swin.edu.au 23 February 2011 Page 10 ACM Multimedia Systems 2011

Goodput

Downstream goodput vs PS queue length - measured (testbed) and simulated (NS-2), 1500/256 kbps & 24/1 Mbps links, 100ms min RTT.

ADSL 1 (1500/256kbps) ADSL 2 (24/1Mbps)

Page 11: Multimedia-unfriendly TCP Congestion Control and Home …web.cs.wpi.edu/~claypool/mmsys-2011/Day1-4_Multimedia... · 2011. 3. 9. · Multimedia-unfriendly TCP Congestion Control and

http://caia.swin.edu.au 23 February 2011 Page 11 ACM Multimedia Systems 2011

CUBIC Induces More OWD Than NewReno

Congestion window oscillation coupled with induced queuing delay. ADSL1 type link: 1500/256kbps, min 100ms RTT, 20 000byte queue.

Downstream NewReno Downstream CUBIC

Page 12: Multimedia-unfriendly TCP Congestion Control and Home …web.cs.wpi.edu/~claypool/mmsys-2011/Day1-4_Multimedia... · 2011. 3. 9. · Multimedia-unfriendly TCP Congestion Control and

http://caia.swin.edu.au 23 February 2011 Page 12 ACM Multimedia Systems 2011

Collateral Damage to Voice Stream - Latency

Induced downstream queuing delay - measured (testbed) and simulated (NS-2), 40 000byte FB-loose queue.

ADSL 1 (1500/256kbps) ADSL 2 (24/1Mbps)

Page 13: Multimedia-unfriendly TCP Congestion Control and Home …web.cs.wpi.edu/~claypool/mmsys-2011/Day1-4_Multimedia... · 2011. 3. 9. · Multimedia-unfriendly TCP Congestion Control and

http://caia.swin.edu.au 23 February 2011 Page 13 ACM Multimedia Systems 2011

Collateral Damage to Voice Stream - Loss

Proportion of CBR packets lost on the downstream for different packet drop mechanisms - 1500/256kbps link, 100ms minimum RTT.

FB-loose, FB-strict & PS NS-2

FB-loose Testbed & NS-2

Page 14: Multimedia-unfriendly TCP Congestion Control and Home …web.cs.wpi.edu/~claypool/mmsys-2011/Day1-4_Multimedia... · 2011. 3. 9. · Multimedia-unfriendly TCP Congestion Control and

http://caia.swin.edu.au 23 February 2011 Page 14 ACM Multimedia Systems 2011

Detailed Simulation: Impact of Queue Size on Loss

[52500, 67500] in 30 byte increments, 1500/256kbps, 100ms min RTT. Dashed line envelope - 5% & 95% confidence intervals.

NewReno CUBIC

Page 15: Multimedia-unfriendly TCP Congestion Control and Home …web.cs.wpi.edu/~claypool/mmsys-2011/Day1-4_Multimedia... · 2011. 3. 9. · Multimedia-unfriendly TCP Congestion Control and

http://caia.swin.edu.au 23 February 2011 Page 15 ACM Multimedia Systems 2011

Time Series Analysis of FB-Strict

Page 16: Multimedia-unfriendly TCP Congestion Control and Home …web.cs.wpi.edu/~claypool/mmsys-2011/Day1-4_Multimedia... · 2011. 3. 9. · Multimedia-unfriendly TCP Congestion Control and

http://caia.swin.edu.au 23 February 2011 Page 16 ACM Multimedia Systems 2011

Time Series Analysis of PS

Page 17: Multimedia-unfriendly TCP Congestion Control and Home …web.cs.wpi.edu/~claypool/mmsys-2011/Day1-4_Multimedia... · 2011. 3. 9. · Multimedia-unfriendly TCP Congestion Control and

http://caia.swin.edu.au 23 February 2011 Page 17 ACM Multimedia Systems 2011

Time Series Analysis of FB-Loose

Page 18: Multimedia-unfriendly TCP Congestion Control and Home …web.cs.wpi.edu/~claypool/mmsys-2011/Day1-4_Multimedia... · 2011. 3. 9. · Multimedia-unfriendly TCP Congestion Control and

http://caia.swin.edu.au 23 February 2011 Page 18 ACM Multimedia Systems 2011

Detailed outline (section 3 of 3) 1.  Introduction

2.  Results

3.  Conclusion

3.  Conclusion •  Summary & Open Questions

Page 19: Multimedia-unfriendly TCP Congestion Control and Home …web.cs.wpi.edu/~claypool/mmsys-2011/Day1-4_Multimedia... · 2011. 3. 9. · Multimedia-unfriendly TCP Congestion Control and

http://caia.swin.edu.au 23 February 2011 Page 19 ACM Multimedia Systems 2011

Summary & Open Questions  Summary

  CUBIC keeps buffers fuller: latency & loss increase   FB-Loose disadvantages small, real-time multimedia packets   Byte-based queue sizes around integer multiples of TCP packet

size correspond with lowest FB-Loose loss & highest FB-Strict loss   Dummynet implements FB-Loose byte-based queues

 Open Questions   Does using TCP for interactive traffic change things?   Can we develop guidelines for home gateway manufacturers?