Performance Enhancement of TFRC in Wireless Networks

Performance Enhancement of TFRC in Wireless Networks

Performance Analysis of the Intertwined Effects between Network Layers for 802.11g Transmissions Jon Gretarsson, Feng Li, Mingzhe Li, Ashish Samant, Huahui Wu, Mark Claypool and Robert Kinicki WPI Computer Science Department Worcester, MA, 01609 USA Presenter - Bob Kinicki Wireless Multimedia Networking and Performance Modeling (WMuNeP) Montreal, Canada, October 2005 Outline Motivation Previous Work

Analytic Models Measurement Studies Experiments Tools and Setup Experimental Design Results and Analysis Conclusions and Future Work 2 WMuNeP05 October 13, 2005 Motivation

Previous research has studied WLAN performance through analytic modeling, simulation and measurement. However, the conclusions drawn have not always been precise and the results have focused on one protocol layer (primarily the data link layer). Important Question: Is Quality of Service (QoS) a realistic goal over WLANs? We are interested in refining how one application running over a Wireless LAN (WLAN) can impact another application.

The interaction between protocol layers can yield results that can significantly impact performance when multiple applications run over a WLAN. 3 WMuNeP05 October 13, 2005 Outline Motivation Previous Work Analytic Models Measurement Studies Experiments Tools and Setup

Experimental Design Results and Analysis Conclusions and Future Work 4 WMuNeP05 October 13, 2005 Analytic Models of 802.11 5

[Cali et al. 98] IEEE 802.11 Wireless LAN: Capacity Analysis and Protocol Enhancement Models early 802.11, i.e., no dynamic rate adaptation. Models the ideal channel: no transmission errors, no hidden terminals. [Bianchi 2000] Performance Analysis of the IEEE 802.11 Distributed Coordination Function Uses the same assumptions as Cali, but emphasizes the saturation throughput such that the transmission queue of each wireless node is never empty. [Heuss et al 03] Performance Anomaly of 802.11b Employs analytic equations based on simplified version of Bianchi including no multiple collisions (no retries) and hosts alternate transmissions. Focuses on dynamic rate adaptation effect. Conclusion: A single slow wireless node brings all wireless

nodes down to its throughput level. Simulate and measure, but measurement uses only upstream UDP and TCP. Note: TCP results do not match very well. WMuNeP05 October 13, 2005 Generic WLAN with AP Downstream Client Access Point Server queue

Client Upstream 6 WMuNeP05 October 13, 2005 Measurement Studies of 802.11 [Pilosof et al. 03] Understanding TCP fairness over Wireless LAN Predominantly simulation, but includes one set of measurement results to show that TCP upstream dominates over TCP downstream with background UDP traffic that makes buffers available at the AP the critical

resource. [Aguayo et al. 04] Link-level Measurements from an 802.11b Mesh Network Perform early morning measurements of Roofnet where there is one sender at a time. Conclude there is not a strong correlation with link distance and SNR with link level loss rates and that an important cause of intermediate loss rates is multi-path fading. 7 WMuNeP05 October 13, 2005 Measurement Studies of 802.11

8 [ [Bai and Williamson 04] The Effects of Mobility on Wireless Media Streaming Performance Create their own AP device to vary queue size. Downstream measurements of UDP videos show WLAN supports easily two fixed clients receiving 1Mbps video clips with AP queue < 30 buffers. When one client becomes mobile, it goes through bad locations and frames get discarded, rate adaptation moves to 1 Mbps, AP queue backlogs and overflows!! When one client fixed and one client mobile, mobile client kills

performance of fixed client because the MAC-layer queue fills with frames from poorly-connected client. The AP queue is the bottleneck. [Yarvis et al. 05] Characteristics of 802.11 Wireless Networks Consider: transmission rate, transmission power, node location, house type. Conduct measurements in three homes with link layer retransmissions disabled. Discover: wireless performance can be quite asymmetric, node placement can be a key factor, no correlation with physical distance. WMuNeP05 October 13, 2005 802.11 Physical Layer Adjust transmission rate on the fly [N. Kim]

9 WMuNeP05 October 13, 2005 Outline Motivation Previous Work Analytic Models Measurement Studies Experiments Tools and Setup Experimental Design Results and Analysis

Conclusions and Future Work 10 WMuNeP05 October 13, 2005 Tools & Experimental Setup Wireless Signal Strength Good Location: >= -70dBm Bad Location: <= - 75dBm 11 WMuNeP05 October 13, 2005

Signal Strength Analysis Fig.2 . Fig.3. 12 12 WMuNeP05 October 13, 2005 Experiments Streaming Video Characteristics Length : 2 minutes

Encoding bit rate: 5Mbps Resolution: 352 x 288 Frame Rate : 24 fps 13 WMuNeP05 October 13, 2005 Video Frames 14 WMuNeP05 October 13, 2005 MPEG Group Of Pictures (GOP)

IBBPBBPBBPBBIBBPBBPBBPBBI Frame types are of different sizes This creates VBR video transmissions I0 15

B 00 B 01 P 1 B 10

P 2 Euro IMSA February 14, 2006 I0 Experimental Design The two laptops are positioned in exactly the same location with the same physical orientation and at locations known for little wireless traffic. All experiments were conducted at night to minimize motion (from people).

Although the videos stream for two minutes, our analysis uses data between the 50-100 second interval. Each experiment was repeated three times. 16 WMuNeP05 October 13, 2005 Consistency Test dynamic rate adaptation 17

Figure 2: Wireless Signal Strength and Channel Capacity for Three Separate Runs WMuNeP05 October 13, 2005 Outline Motivation Previous Work Analytic Models Measurement Studies Experiments Tools and Setup Experimental Design Results and Analysis Conclusions and Future Work

18 WMuNeP05 October 13, 2005 A Single TCP Download Average throughput 18.8 Mbps Fig 3a TCP Download to a Good Location 19 WMuNeP05 October 13, 2005 802.11 Performance

Anomaly A: 9.3 Mbps B: 9.6 Mbps A: 2.8 Mbps B: 2.1 Mbps Fig 3 TCP Throughput Comparison Fig 4 Wireless Received Signal Strength Indicator Comparison 20 WMuNeP05 October 13, 2005

TCP Download Channel Capacities Figure 5: Channel Capacity Impacted by Location 21 WMuNeP05 October 13, 2005 Bad UDP Stream A: 2.8 Mbps B: 2.1 Mbps A: 0.3 Mbps B: 2.5 Mbps

UDP stream kills TCP download!! Figure 6: Throughput Impacted by Location and Application 22 WMuNeP05 October 13, 2005 Frame Retries and Packet Loss A: 0.05 B: 0.2

AP queue causes high packet loss Figure 7: Wireless Layer Retry Fraction 23 Figure 8: UDP Ping Loss Fraction WMuNeP05 October 13, 2005 Round Trip Times

large UDP RTTs 24 WMuNeP05 October 13, 2005 Round Trip Times TCP does not fill the AP Queue. 25 WMuNeP05 October 13, 2005

Good Streaming 26 WMuNeP05 October 13, 2005 Bad Streaming TCP streaming adjusts to bad location 27 WMuNeP05 October 13, 2005 Single Bad Streams

28 WMuNeP05 October 13, 2005 Conclusions and Future Work Application behavior impacts WLAN performance of concurrent applications. The choice of Transport Protocol impacts performance over a WLAN. Just modeling the data link channel misses interwined effects of the AP network layer queuing. We need to get inside the AP to understand the

queuing in both the upstream and downstream direction. Is there a way for streaming application to get hints about the wireless data link layer? 29 WMuNeP05 October 13, 2005 Thank You! Performance Analysis of the Intertwined Effects between Network Layers for 802.11g Transmissions Jon Gretarsson, Feng Li, Mingzhe Li, Ashish Samant, Huahui Wu, Mark Claypool and Robert Kinicki

WPI Computer Science Department Worcester, Massachusetts 01609 [email protected] Wireless Multimedia Networking and Performance Modeling (WMuNeP) Montreal, Canada, October 2005

Recently Viewed Presentations