SlideShare a Scribd company logo

Design Considerations for RINA Congestion Control over WiFi Links

•
•
A
ARCFIRE ICT

The document discusses design considerations for congestion control over WiFi links using the RINA framework. It begins with background on how traditional Internet congestion control assumptions break down over wireless networks. Measurements of WiFi DCF behavior show it can behave predictably. A machine learning approach is proposed to predict attainable rates. Localized congestion control loops in RINA or with PEPs could rely directly on predicted rates, improving performance over TCP. Benefits include optimized performance over WiFi links, including in wireless mesh networks.

Related slideshows

RINA as a Clean-Slate Approach to Software Networks
RINA as a Clean-Slate Approach to Software Networks RINA as a Clean-Slate Approach to Software Networks
RINA as a Clean-Slate Approach to Software Networks
 
Rina acc-icc16-stein
Rina acc-icc16-steinRina acc-icc16-stein
Rina acc-icc16-stein
 
PRISTINE @ FIA Athens 2014
PRISTINE @ FIA Athens 2014PRISTINE @ FIA Athens 2014
PRISTINE @ FIA Athens 2014
 
1 of 33
Download to read offline
Design Considerations for RINA Congestion Control over WiFi Links Kristian A. Hiorth1, Michael Welzl1 University of Oslo, Norway1 February 18, 2019
Design Considerations for RINA Congestion Control over WiFi Links Kristian A. Hiorth Dept. of Informatics University of Oslo [Photo: Andre Douque]
Overview: we will describe the background for our work, examine measurement results and discuss benefits of our proposed concept Background 20 25 30 35 40 45 Time (seconds) 0.0 0.5 1.0 1.5 2.0 2.5 3.0 Throughput(bitspersecond) ×107 Total Flow 1 Flow 2 Flow 3 Estimate Measurements Benefits 1
Background and motivation Measurements Benefits
Wireless networks break the assumptions baked into traditional Internet congestion control 2
Wireless networks break the assumptions baked into traditional Internet congestion control 2
Wireless networks break the assumptions baked into traditional Internet congestion control [Photos: “Pete” and Andre Douque] 2
The IEEE 802.11 Distributed Coordination Function ultimately determines sending rates over WiFi [Diagram: Wikipedia] 3
The DCF has been extensively studied, however models assume numerous constraints Example assumptions: Saturation Fixed frame size Fixed MAC settings 4
The DCF has been extensively studied, however models assume numerous constraints Example assumptions: Saturation Fixed frame size Fixed MAC settings 4
The DCF has been extensively studied, however models assume numerous constraints Example assumptions: Saturation Fixed frame size Fixed MAC settings 4
To overcome model limitations, we envision a measurement driven machine learning solution Measure → Predict → Cross-check with buffer drainage 5
To overcome model limitations, we envision a measurement driven machine learning solution Measure → Predict → Cross-check with buffer drainage 5
To overcome model limitations, we envision a measurement driven machine learning solution Measure → Predict → Cross-check with buffer drainage 5
RINA is scope-aware and naturally enables localized congestion control loops Application Link Link Link Link Routing 6
Performance Enhancing Proxies allow establishing local control loops even in IP networks Wireless Local Area Network WirelessDesktop NetworkCard OK Madein Groland w_i ~~ø#||| 1121314156---**788 Access Point WAN PEP 7
Performance Enhancing Proxies allow establishing local control loops even in IP networks Wireless Local Area Network WirelessDesktop NetworkCard OK Madein Groland w_i ~~ø#||| 1121314156---**788 Access Point WAN PEP 7
Background and motivation Measurements Benefits
We studied WiFi DCF behaviour by measuring with real hardware using a simple scenario Wireless Local Area Network Node 3 (measurement node) WirelessDesktop NetworkCard OK Madein Groland w_i ~~ø#||| 1121314156---**788 WirelessDesktop NetworkCard OK Madein Groland w_i ~~ø#||| 1121314156---**788 Node 1 WirelessDesktop NetworkCard OK Madein Groland w_i ~~ø#||| 1121314156---**788 Node 2 WirelessDesktop NetworkCard OK Madein Groland w_i ~~ø#||| 1121314156---**788 Access Point “WAN” 8
When all stations send at the same physical rate, DCF behaves very predictably and fairly 20 25 30 35 40 45 Time (seconds) 0.0 0.5 1.0 1.5 2.0 2.5 3.0 Throughput(bitspersecond) ×107 Total Flow 1 Flow 2 Flow 3 Estimate PHY fixed at 54Mbps 9
DCF behaviour remains highly predictable also when allowing different PHY rates 20 25 30 35 40 45 Time (seconds) 0.00 0.25 0.50 0.75 1.00 1.25 1.50 1.75 Throughput(bitspersecond) ×107 Total Flow 1 Flow 2 Flow 3 PHY fixed at 12Mbps, 24Mbps and 54Mbps, respectively. 10
Normal PHY rate adaptation introduces more noise, yet appears reasonaby predictable 20 25 30 35 40 45 Time (seconds) 0.0 0.5 1.0 1.5 2.0 2.5 3.0 Throughput(bitspersecond) ×107 Total Flow 1 Flow 2 Flow 3 PHY controlled by Minstrel rate adaption algorithm. 11
Design Considerations for RINA Congestion Control over WiFi Links
Relying directly on the DCF enhances performance compared to TCP congestion control 20 25 30 35 40 45 Time (seconds) 0.0 0.5 1.0 1.5 2.0 2.5 3.0 Throughput(bitspersecond) ×107 Total Flow 1 Flow 2 Flow 3 Estimate Pure DCF 20 25 30 35 40 45 Time (seconds) 0.0 0.5 1.0 1.5 2.0 2.5 3.0 Throughput(bitspersecond) ×107 Total Flow 1 Flow 2 Flow 3 TCP Cubic 12
Even TCP BBR is outperformed 20 25 30 35 40 45 Time (seconds) 0.0 0.5 1.0 1.5 2.0 2.5 3.0 Throughput(bitspersecond) ×107 Total Flow 1 Flow 2 Flow 3 Estimate Pure DCF 20 25 30 35 40 45 Time (seconds) 0.0 0.5 1.0 1.5 2.0 2.5 3.0 Throughput(bitspersecond) ×107 Total Flow 1 Flow 2 Flow 3 TCP BBR 13
Background and motivation Measurements Benefits
Even in simple wireless LANs a predictive solution is beneficial compared to plain flow control [Photo: joonas.fi] 14
In RINA there are obvious benefits to knowing the actual attainable link rate Host W iFi shi m App 1st hop AP 2 nd hop AP Host 15
Wireless mesh networks can benefit greatly from the use of a known-rate, hop by hop congestion control [Photo: DeWALT] 16
Unlike many previously proposed cross-layer mechanisms our concept is properly scoped Wireless LAN WirelessDesktop NetworkCard OK Madein Groland w_i ~~ø#||| 1121314156---**788 AP WAN WAN bottleneck! 17
Unlike many previously proposed cross-layer mechanisms our concept is properly scoped Wireless LAN WirelessDesktop NetworkCard OK Madein Groland w_i ~~ø#||| 1121314156---**788 AP WAN WAN bottleneck! 17
In conclusion: Local loop, data-driven WiFi congestion control appears both feasible and superior to end-to-end Predictive, quantified rate + Proper scoping ↓ Optimized WiFi performance Questions? 18
In conclusion: Local loop, data-driven WiFi congestion control appears both feasible and superior to end-to-end Predictive, quantified rate + Proper scoping ↓ Optimized WiFi performance Questions? 18

More Related Content

Design Considerations for RINA Congestion Control over WiFi Links

  • 1. Design Considerations for RINA Congestion Control over WiFi Links Kristian A. Hiorth1, Michael Welzl1 University of Oslo, Norway1 February 18, 2019
  • 2. Design Considerations for RINA Congestion Control over WiFi Links Kristian A. Hiorth Dept. of Informatics University of Oslo [Photo: Andre Douque]
  • 3. Overview: we will describe the background for our work, examine measurement results and discuss benefits of our proposed concept Background 20 25 30 35 40 45 Time (seconds) 0.0 0.5 1.0 1.5 2.0 2.5 3.0 Throughput(bitspersecond) ×107 Total Flow 1 Flow 2 Flow 3 Estimate Measurements Benefits 1
  • 5. Wireless networks break the assumptions baked into traditional Internet congestion control 2
  • 6. Wireless networks break the assumptions baked into traditional Internet congestion control 2
  • 7. Wireless networks break the assumptions baked into traditional Internet congestion control [Photos: “Pete” and Andre Douque] 2
  • 8. The IEEE 802.11 Distributed Coordination Function ultimately determines sending rates over WiFi [Diagram: Wikipedia] 3
  • 9. The DCF has been extensively studied, however models assume numerous constraints Example assumptions: Saturation Fixed frame size Fixed MAC settings 4
  • 10. The DCF has been extensively studied, however models assume numerous constraints Example assumptions: Saturation Fixed frame size Fixed MAC settings 4
  • 11. The DCF has been extensively studied, however models assume numerous constraints Example assumptions: Saturation Fixed frame size Fixed MAC settings 4
  • 12. To overcome model limitations, we envision a measurement driven machine learning solution Measure → Predict → Cross-check with buffer drainage 5
  • 13. To overcome model limitations, we envision a measurement driven machine learning solution Measure → Predict → Cross-check with buffer drainage 5
  • 14. To overcome model limitations, we envision a measurement driven machine learning solution Measure → Predict → Cross-check with buffer drainage 5
  • 15. RINA is scope-aware and naturally enables localized congestion control loops Application Link Link Link Link Routing 6
  • 16. Performance Enhancing Proxies allow establishing local control loops even in IP networks Wireless Local Area Network WirelessDesktop NetworkCard OK Madein Groland w_i ~~ø#||| 1121314156---**788 Access Point WAN PEP 7
  • 17. Performance Enhancing Proxies allow establishing local control loops even in IP networks Wireless Local Area Network WirelessDesktop NetworkCard OK Madein Groland w_i ~~ø#||| 1121314156---**788 Access Point WAN PEP 7
  • 19. We studied WiFi DCF behaviour by measuring with real hardware using a simple scenario Wireless Local Area Network Node 3 (measurement node) WirelessDesktop NetworkCard OK Madein Groland w_i ~~ø#||| 1121314156---**788 WirelessDesktop NetworkCard OK Madein Groland w_i ~~ø#||| 1121314156---**788 Node 1 WirelessDesktop NetworkCard OK Madein Groland w_i ~~ø#||| 1121314156---**788 Node 2 WirelessDesktop NetworkCard OK Madein Groland w_i ~~ø#||| 1121314156---**788 Access Point “WAN” 8
  • 20. When all stations send at the same physical rate, DCF behaves very predictably and fairly 20 25 30 35 40 45 Time (seconds) 0.0 0.5 1.0 1.5 2.0 2.5 3.0 Throughput(bitspersecond) ×107 Total Flow 1 Flow 2 Flow 3 Estimate PHY fixed at 54Mbps 9
  • 21. DCF behaviour remains highly predictable also when allowing different PHY rates 20 25 30 35 40 45 Time (seconds) 0.00 0.25 0.50 0.75 1.00 1.25 1.50 1.75 Throughput(bitspersecond) ×107 Total Flow 1 Flow 2 Flow 3 PHY fixed at 12Mbps, 24Mbps and 54Mbps, respectively. 10
  • 22. Normal PHY rate adaptation introduces more noise, yet appears reasonaby predictable 20 25 30 35 40 45 Time (seconds) 0.0 0.5 1.0 1.5 2.0 2.5 3.0 Throughput(bitspersecond) ×107 Total Flow 1 Flow 2 Flow 3 PHY controlled by Minstrel rate adaption algorithm. 11
  • 24. Relying directly on the DCF enhances performance compared to TCP congestion control 20 25 30 35 40 45 Time (seconds) 0.0 0.5 1.0 1.5 2.0 2.5 3.0 Throughput(bitspersecond) ×107 Total Flow 1 Flow 2 Flow 3 Estimate Pure DCF 20 25 30 35 40 45 Time (seconds) 0.0 0.5 1.0 1.5 2.0 2.5 3.0 Throughput(bitspersecond) ×107 Total Flow 1 Flow 2 Flow 3 TCP Cubic 12
  • 25. Even TCP BBR is outperformed 20 25 30 35 40 45 Time (seconds) 0.0 0.5 1.0 1.5 2.0 2.5 3.0 Throughput(bitspersecond) ×107 Total Flow 1 Flow 2 Flow 3 Estimate Pure DCF 20 25 30 35 40 45 Time (seconds) 0.0 0.5 1.0 1.5 2.0 2.5 3.0 Throughput(bitspersecond) ×107 Total Flow 1 Flow 2 Flow 3 TCP BBR 13
  • 27. Even in simple wireless LANs a predictive solution is beneficial compared to plain flow control [Photo: joonas.fi] 14
  • 28. In RINA there are obvious benefits to knowing the actual attainable link rate Host W iFi shi m App 1st hop AP 2 nd hop AP Host 15
  • 29. Wireless mesh networks can benefit greatly from the use of a known-rate, hop by hop congestion control [Photo: DeWALT] 16
  • 30. Unlike many previously proposed cross-layer mechanisms our concept is properly scoped Wireless LAN WirelessDesktop NetworkCard OK Madein Groland w_i ~~ø#||| 1121314156---**788 AP WAN WAN bottleneck! 17
  • 31. Unlike many previously proposed cross-layer mechanisms our concept is properly scoped Wireless LAN WirelessDesktop NetworkCard OK Madein Groland w_i ~~ø#||| 1121314156---**788 AP WAN WAN bottleneck! 17
  • 32. In conclusion: Local loop, data-driven WiFi congestion control appears both feasible and superior to end-to-end Predictive, quantified rate + Proper scoping ↓ Optimized WiFi performance Questions? 18
  • 33. In conclusion: Local loop, data-driven WiFi congestion control appears both feasible and superior to end-to-end Predictive, quantified rate + Proper scoping ↓ Optimized WiFi performance Questions? 18