Practical MU-MIMO User Selection on 802.11ac Commodity ...sur/papers/MUSE_MobiCom16_slides.pdf ·...
Transcript of Practical MU-MIMO User Selection on 802.11ac Commodity ...sur/papers/MUSE_MobiCom16_slides.pdf ·...
![Page 1: Practical MU-MIMO User Selection on 802.11ac Commodity ...sur/papers/MUSE_MobiCom16_slides.pdf · suboptimal statistics that lead to poor MU-MIMO performance MUSE leverages existing](https://reader034.fdocuments.in/reader034/viewer/2022042306/5ed1861b0be6a4353031c27c/html5/thumbnails/1.jpg)
Practical MU-MIMO User Selection on 802.11ac Commodity Networks
Ioannis Pefkianakis, Xinyu Zhang and Kyu-Han Kim
Sanjib Sur
![Page 2: Practical MU-MIMO User Selection on 802.11ac Commodity ...sur/papers/MUSE_MobiCom16_slides.pdf · suboptimal statistics that lead to poor MU-MIMO performance MUSE leverages existing](https://reader034.fdocuments.in/reader034/viewer/2022042306/5ed1861b0be6a4353031c27c/html5/thumbnails/2.jpg)
From Legacy to Gbps Wi-Fi
Legacy 802.11a/b/g:Speeds up to 54 Mbps
1999-2003 2009
HT 802.11n:Speeds up to 600 Mbps
2013
VHT 802.11ac:Speeds > 6 Gbps
What is new in 802.11ac?
Channel bandwidth Tx/Rx antennas Modulation & coding Device type Data rate
40 MHz 1x1 256-QAM, 5/6 Smartphone, Tablet 200 Mbps
40 MHz 3x3 256-QAM, 5/6 Laptop 600 Mbps
80 MHz 1x1 256-QAM, 5/6 Smartphone, Tablet 433 Mbps
80 MHz 2x2 256-QAM, 5/6 Laptop, Tablet 867 Mbps
80 MHz 3x3 256-QAM, 5/6 Laptop 1.3 Gbps
160MHz 4-ant AP / 4, 1-ant STAs (MU-MIMO)
256-QAM, 5/6 Smartphone, Tablet 3.39 Gbps
160MHz 8-ant AP / 4, 2-ant STAs (MU-MIMO)
256-QAM, 5/6 Laptop 6.77 Gbps
![Page 3: Practical MU-MIMO User Selection on 802.11ac Commodity ...sur/papers/MUSE_MobiCom16_slides.pdf · suboptimal statistics that lead to poor MU-MIMO performance MUSE leverages existing](https://reader034.fdocuments.in/reader034/viewer/2022042306/5ed1861b0be6a4353031c27c/html5/thumbnails/3.jpg)
Multi-User MIMO Feature on 802.11ac
A MU-MIMO access point transmits multiple data streams concurrently to different receivers
![Page 4: Practical MU-MIMO User Selection on 802.11ac Commodity ...sur/papers/MUSE_MobiCom16_slides.pdf · suboptimal statistics that lead to poor MU-MIMO performance MUSE leverages existing](https://reader034.fdocuments.in/reader034/viewer/2022042306/5ed1861b0be6a4353031c27c/html5/thumbnails/4.jpg)
MU-MIMO User and Rate Selection
User selection determines which users to serve concurrently
×Rate selection determines the best link speed for each users
User and rate selection algorithms are fundamental for MU-MIMO performance
![Page 5: Practical MU-MIMO User Selection on 802.11ac Commodity ...sur/papers/MUSE_MobiCom16_slides.pdf · suboptimal statistics that lead to poor MU-MIMO performance MUSE leverages existing](https://reader034.fdocuments.in/reader034/viewer/2022042306/5ed1861b0be6a4353031c27c/html5/thumbnails/5.jpg)
Practical MU-MIMO Performance
802.11ac UsersCommodity Access Point
~ 50 Mbps
Throughput drops when MU-MIMO is enabled!MU-MIMO gain can be even lower than SU-MIMO in some enterprise settings
![Page 6: Practical MU-MIMO User Selection on 802.11ac Commodity ...sur/papers/MUSE_MobiCom16_slides.pdf · suboptimal statistics that lead to poor MU-MIMO performance MUSE leverages existing](https://reader034.fdocuments.in/reader034/viewer/2022042306/5ed1861b0be6a4353031c27c/html5/thumbnails/6.jpg)
Outline of Today’s Talk
Why commodity APs avoid state-of-the-art solutions?
How can we design a robust and practical MU-MIMO user and rate selection solution?
What is the impact of our design and real-time implementation?
![Page 7: Practical MU-MIMO User Selection on 802.11ac Commodity ...sur/papers/MUSE_MobiCom16_slides.pdf · suboptimal statistics that lead to poor MU-MIMO performance MUSE leverages existing](https://reader034.fdocuments.in/reader034/viewer/2022042306/5ed1861b0be6a4353031c27c/html5/thumbnails/7.jpg)
Outline of Today’s Talk
Why commodity APs avoid state-of-the-art solutions?
How can we design a robust and practical MU-MIMO user and rate selection solution?
What is the impact of our design and real-time implementation?
![Page 8: Practical MU-MIMO User Selection on 802.11ac Commodity ...sur/papers/MUSE_MobiCom16_slides.pdf · suboptimal statistics that lead to poor MU-MIMO performance MUSE leverages existing](https://reader034.fdocuments.in/reader034/viewer/2022042306/5ed1861b0be6a4353031c27c/html5/thumbnails/8.jpg)
Algorithms on Commodity APs
Both user and rate selection solutions are trial-and-error based
Packet error rate increases
Statistics-based Adaptation
Problems:
High convergence time
Poor group formation
![Page 9: Practical MU-MIMO User Selection on 802.11ac Commodity ...sur/papers/MUSE_MobiCom16_slides.pdf · suboptimal statistics that lead to poor MU-MIMO performance MUSE leverages existing](https://reader034.fdocuments.in/reader034/viewer/2022042306/5ed1861b0be6a4353031c27c/html5/thumbnails/9.jpg)
Limitations of State-of-the-Art Solutions
Existing solutions rely on full wireless channel feedback for user and rate selection
Heterogeneous bandwidth users limit grouping opportunities
Limited resources on APs cannot support computationally and memory expensive operations, required by existing solutions
160 MHz 40 MHz 20 MHz
![Page 10: Practical MU-MIMO User Selection on 802.11ac Commodity ...sur/papers/MUSE_MobiCom16_slides.pdf · suboptimal statistics that lead to poor MU-MIMO performance MUSE leverages existing](https://reader034.fdocuments.in/reader034/viewer/2022042306/5ed1861b0be6a4353031c27c/html5/thumbnails/10.jpg)
Outline of Today’s Talk
Why commodity APs avoid state-of-the-art solutions?
How can we design a robust and practical MU-MIMO user and rate selection solution?
What is the impact of our design and real-time implementation?
![Page 11: Practical MU-MIMO User Selection on 802.11ac Commodity ...sur/papers/MUSE_MobiCom16_slides.pdf · suboptimal statistics that lead to poor MU-MIMO performance MUSE leverages existing](https://reader034.fdocuments.in/reader034/viewer/2022042306/5ed1861b0be6a4353031c27c/html5/thumbnails/11.jpg)
Our Solution: MUSE
Fast Configuration Selection
Real-Time Implementation &
Optimization
Users’ Performance Prediction
A practical MU-MIMO user and rate selection for 802.11ac commodity devices
![Page 12: Practical MU-MIMO User Selection on 802.11ac Commodity ...sur/papers/MUSE_MobiCom16_slides.pdf · suboptimal statistics that lead to poor MU-MIMO performance MUSE leverages existing](https://reader034.fdocuments.in/reader034/viewer/2022042306/5ed1861b0be6a4353031c27c/html5/thumbnails/12.jpg)
Our Solution: MUSE
Users’ Performance Prediction
Fast Configuration Selection
Real-Time Implementation &
Optimization
![Page 13: Practical MU-MIMO User Selection on 802.11ac Commodity ...sur/papers/MUSE_MobiCom16_slides.pdf · suboptimal statistics that lead to poor MU-MIMO performance MUSE leverages existing](https://reader034.fdocuments.in/reader034/viewer/2022042306/5ed1861b0be6a4353031c27c/html5/thumbnails/13.jpg)
Our Solution: MUSE
Fast Configuration Selection
160 MHz
40 MHz
20 MHz
20 MHz
80 MHz
Users’ Performance Prediction
Real-Time Implementation &
Optimization
![Page 14: Practical MU-MIMO User Selection on 802.11ac Commodity ...sur/papers/MUSE_MobiCom16_slides.pdf · suboptimal statistics that lead to poor MU-MIMO performance MUSE leverages existing](https://reader034.fdocuments.in/reader034/viewer/2022042306/5ed1861b0be6a4353031c27c/html5/thumbnails/14.jpg)
Our Solution: MUSE
Fast Configuration Selection
Users’ Performance Prediction
Real-Time Implementation &
Optimization
![Page 15: Practical MU-MIMO User Selection on 802.11ac Commodity ...sur/papers/MUSE_MobiCom16_slides.pdf · suboptimal statistics that lead to poor MU-MIMO performance MUSE leverages existing](https://reader034.fdocuments.in/reader034/viewer/2022042306/5ed1861b0be6a4353031c27c/html5/thumbnails/15.jpg)
Users’ Performance Prediction
Signal-to-Interference-Noise Ratio (SINR)
Link Speed
𝜌 =
Signal leakage
Similarity between partial channel reports
{V1, V2}
Inter-user Interference
V1
V2
![Page 16: Practical MU-MIMO User Selection on 802.11ac Commodity ...sur/papers/MUSE_MobiCom16_slides.pdf · suboptimal statistics that lead to poor MU-MIMO performance MUSE leverages existing](https://reader034.fdocuments.in/reader034/viewer/2022042306/5ed1861b0be6a4353031c27c/html5/thumbnails/16.jpg)
MUSE Predictable SINR Accuracy
Comparison between MUSE predictable SINR with full channel feedback SINR
Median prediction error ~ 0.5 dB!
![Page 17: Practical MU-MIMO User Selection on 802.11ac Commodity ...sur/papers/MUSE_MobiCom16_slides.pdf · suboptimal statistics that lead to poor MU-MIMO performance MUSE leverages existing](https://reader034.fdocuments.in/reader034/viewer/2022042306/5ed1861b0be6a4353031c27c/html5/thumbnails/17.jpg)
Fast Configuration Selection
160 MHz
40 MHz
20 MHz
20 MHz
80 MHz
160 MHz 160 MHz 160 MHz 80 MHz 80 MHz 80 MHz 20 MHz 20 MHz 20 MHz
160 MHz 80 MHz 80 MHz 80 MHz 20 MHz×××
![Page 18: Practical MU-MIMO User Selection on 802.11ac Commodity ...sur/papers/MUSE_MobiCom16_slides.pdf · suboptimal statistics that lead to poor MU-MIMO performance MUSE leverages existing](https://reader034.fdocuments.in/reader034/viewer/2022042306/5ed1861b0be6a4353031c27c/html5/thumbnails/18.jpg)
Fast Configuration Selection
Partial channel report correlation remains unaffected irrespective of the channel bandwidth
Correlation between two users
20 MHz
20 MHz
20 MHz
![Page 19: Practical MU-MIMO User Selection on 802.11ac Commodity ...sur/papers/MUSE_MobiCom16_slides.pdf · suboptimal statistics that lead to poor MU-MIMO performance MUSE leverages existing](https://reader034.fdocuments.in/reader034/viewer/2022042306/5ed1861b0be6a4353031c27c/html5/thumbnails/19.jpg)
Fast Configuration Selection
Partial channel report correlation remains unaffected irrespective of the channel bandwidth
Correlation between two users
40 MHz
40 MHz
40 MHz
![Page 20: Practical MU-MIMO User Selection on 802.11ac Commodity ...sur/papers/MUSE_MobiCom16_slides.pdf · suboptimal statistics that lead to poor MU-MIMO performance MUSE leverages existing](https://reader034.fdocuments.in/reader034/viewer/2022042306/5ed1861b0be6a4353031c27c/html5/thumbnails/20.jpg)
Fast Configuration Selection
Partial channel report correlation remains unaffected irrespective of the channel bandwidth
Correlation between two users
80 MHz
80 MHz
80 MHz
![Page 21: Practical MU-MIMO User Selection on 802.11ac Commodity ...sur/papers/MUSE_MobiCom16_slides.pdf · suboptimal statistics that lead to poor MU-MIMO performance MUSE leverages existing](https://reader034.fdocuments.in/reader034/viewer/2022042306/5ed1861b0be6a4353031c27c/html5/thumbnails/21.jpg)
Fast Configuration Selection
Strict rate increase for higher channel bandwidth beyond deterministic signal strength
80 MHz PHY rate > 40 MHz PHY rate beyond 22 dB
![Page 22: Practical MU-MIMO User Selection on 802.11ac Commodity ...sur/papers/MUSE_MobiCom16_slides.pdf · suboptimal statistics that lead to poor MU-MIMO performance MUSE leverages existing](https://reader034.fdocuments.in/reader034/viewer/2022042306/5ed1861b0be6a4353031c27c/html5/thumbnails/22.jpg)
Fast Configuration Selection
160 MHz 160 MHz 160 MHz 80 MHz 80 MHz 80 MHz 20 MHz 20 MHz 20 MHz
160 MHz 80 MHz 80 MHz 80 MHz 20 MHz×
All SINR > Threshold
![Page 23: Practical MU-MIMO User Selection on 802.11ac Commodity ...sur/papers/MUSE_MobiCom16_slides.pdf · suboptimal statistics that lead to poor MU-MIMO performance MUSE leverages existing](https://reader034.fdocuments.in/reader034/viewer/2022042306/5ed1861b0be6a4353031c27c/html5/thumbnails/23.jpg)
Practical Implementation Challenge
WLAN Chip
Target CPU
On-Chip Memory
Host System
Firmware Host CPU
DDR
Linux Kernel
32-bit PCIe bus
Core PHY/MAC functionalities, radio management including
legacy user and rate selection
Upper network layers, process Ethernet packets
Access Point
To Ethernet backhaul
![Page 24: Practical MU-MIMO User Selection on 802.11ac Commodity ...sur/papers/MUSE_MobiCom16_slides.pdf · suboptimal statistics that lead to poor MU-MIMO performance MUSE leverages existing](https://reader034.fdocuments.in/reader034/viewer/2022042306/5ed1861b0be6a4353031c27c/html5/thumbnails/24.jpg)
Practical Implementation Challenge
WLAN Chip
Target CPU
On-Chip Memory
Host System
Firmware Host CPU
DDR
Linux Kernel
32-bit PCIe bus
Core MAC-layer functionalities, radio management including
legacy user and rate selection
Upper network layers, process Ethernet packets
Access Point
To Ethernet backhaul
WLAN chip has only 1 MB on-chip memory and a 350 MHz CPU
Approximately 98% memory is used by existing functionalities
![Page 25: Practical MU-MIMO User Selection on 802.11ac Commodity ...sur/papers/MUSE_MobiCom16_slides.pdf · suboptimal statistics that lead to poor MU-MIMO performance MUSE leverages existing](https://reader034.fdocuments.in/reader034/viewer/2022042306/5ed1861b0be6a4353031c27c/html5/thumbnails/25.jpg)
Outline of Today’s Talk
Why commodity APs avoid existing solutions?
How can we design a robust and practical MU-MIMO user and rate selection solution?
What is the impact of our design and real-time implementation?
![Page 26: Practical MU-MIMO User Selection on 802.11ac Commodity ...sur/papers/MUSE_MobiCom16_slides.pdf · suboptimal statistics that lead to poor MU-MIMO performance MUSE leverages existing](https://reader034.fdocuments.in/reader034/viewer/2022042306/5ed1861b0be6a4353031c27c/html5/thumbnails/26.jpg)
Experimental Evaluation
We compare MUSE with Legacy-US and existing solution PUMA*
* Mode and User Selection for Multi-User MIMO WLANs without CSI, N. Anand et. al., INFOCOM’15
PUMA uses a SINR metric where interference between users are fixed
UDP saturated traffics in static and dynamic settings
Per-user median gain > 70%Median performance gain ~50%,
with up to 4x per-user gain!
Static cases Mobile cases
![Page 27: Practical MU-MIMO User Selection on 802.11ac Commodity ...sur/papers/MUSE_MobiCom16_slides.pdf · suboptimal statistics that lead to poor MU-MIMO performance MUSE leverages existing](https://reader034.fdocuments.in/reader034/viewer/2022042306/5ed1861b0be6a4353031c27c/html5/thumbnails/27.jpg)
Analysis of MUSE in High-Gain Scenario
Side-benefit of MUSE from correct user and rate selection
PER reduction implicitly improves frame aggregation level
Super-frame
Sub-frames
Median frame-aggregation level
improves by 4.5x!
![Page 28: Practical MU-MIMO User Selection on 802.11ac Commodity ...sur/papers/MUSE_MobiCom16_slides.pdf · suboptimal statistics that lead to poor MU-MIMO performance MUSE leverages existing](https://reader034.fdocuments.in/reader034/viewer/2022042306/5ed1861b0be6a4353031c27c/html5/thumbnails/28.jpg)
Field-Trial of MUSE
Realistic field-trial with 4 APs placed in an enterprise setting
15 smartphones connected to the AP, uncontrolled environmental activities, interference from external access points
Aggregate throughput gain of 30-45% and 20-30% compared to Legacy-US and PUMA. Per-user throughput gain can be up to 3x!
![Page 29: Practical MU-MIMO User Selection on 802.11ac Commodity ...sur/papers/MUSE_MobiCom16_slides.pdf · suboptimal statistics that lead to poor MU-MIMO performance MUSE leverages existing](https://reader034.fdocuments.in/reader034/viewer/2022042306/5ed1861b0be6a4353031c27c/html5/thumbnails/29.jpg)
Summary
Commodity APs avoid state-of-the-art solutions and rely on suboptimal statistics that lead to poor MU-MIMO performance
MUSE leverages existing low-layer feedback to design and implement a practical user and rate selection solution
Our work is the first to optimize MU-MIMO performance on 802.11ac commodity access points