Download presentation
Presentation is loading. Please wait.
1
Lottery Meets Wireless
Sharad Saha Shravan Rayanchu Dec 14, 2007
2
The Problem Flexible, fine-grained bandwidth allocation in enterprise wireless LANs Motivation Wireless bandwidth is a scarce resource Emerging applications require QoS guarantees Voice-over-IP, Streaming Multimedia Dec 14, 2007
3
Use case Enterprise WLAN Dec 14, 2007
4
Use case Enterprise WLAN Media contention Dec 14, 2007
5
Current Practice IEEE 802.11e provides service differentiation
Four classes: Data, Voice, Video, Best Effort Not fine grained, (IP, port, protocol), Time Not dynamic, no proportion Dec 14, 2007
6
Current Practice IEEE 802.11e provides service differentiation
Four classes: Data, Voice, Video, Best Effort Limitations Coarse granularity Not fine grained, (IP, port, protocol), Time Not dynamic, no proportion Dec 14, 2007
7
Current Practice IEEE 802.11e provides service differentiation
Four classes: Data, Voice, Video, Best Effort Limitations Coarse granularity Inflexible Not fine grained, (IP, port, protocol), Time Not dynamic, no proportion Dec 14, 2007
8
Current Practice IEEE 802.11e provides service differentiation
Four classes: Data, Voice, Video, Best Effort Limitations Coarse granularity Inflexible No concept of proportions Not fine grained, (IP, port, protocol), Time Not dynamic, no proportion Dec 14, 2007
9
Current Practice IEEE 802.11e provides service differentiation
Four classes: Data, Voice, Video, Best Effort Limitations Coarse granularity Inflexible No concept of proportions Ignorant of traffic characteristics Not fine grained, (IP, port, protocol), Time Not dynamic, no proportion Dec 14, 2007
10
Our Approach Lottery Scheduling at the Access Point
Policy file defines flows (similar to ACLs) and proportions Contending flows participate in lottery with tickets given in proportion Lucky winner gets to send the packet Dec 14, 2007
11
Does it work ? Fine grained allocation : Flows
Flexibility : Dynamic policies Proportions : Tickets Traffic characteristics : Ticket inflation Scalability : Stateless, Simple Dec 14, 2007
12
Outline Introduction Lottery : Design and Implementation Evaluation
Conclusions Dec 14, 2007
13
Lottery Implementation
Define flow:ticket pairs (Port 5001 : 100) (IP : 60) (TCP, Port 80 : 200) On a TX opportunity, draw a lottery and send packet Linux Kernel, Madwifi wireless driver, Proc FS for policy inputs Dec 14, 2007
14
Transmit Path Dec 14, 2007
15
Transmit Path Dec 14, 2007
16
Transmit Path Dec 14, 2007
17
Transmit Path Currently queuing happens in the firmware Dec 14, 2007
18
Transmit Path Currently queuing happens in the firmware
A callback function is invoked after every packet transmission Dec 14, 2007
19
Lottery Implementation
Dec 14, 2007
20
Lottery Implementation
Dec 14, 2007
21
Lottery Implementation
Dec 14, 2007
22
Lottery Implementation
Dec 14, 2007
23
Lottery Implementation
Dec 14, 2007
24
Lottery Implementation
Generic Framework Lottery Scheduling Stride Scheduling Dec 14, 2007
25
Ticket Inflation Packet size affects the throughput
Flow 1 : Flow 2 = 100 : 100 Flow 1 (100 bytes) , Flow (200 bytes) Effective ratio = 1 : 2 Per-packet Ticket Inflation Actual Ticket = Ticket * (Base Size) / (Pkt Size) Flow 1 = 100 * 200/100 = 200 Flow 2 = 200 * 200/200 = 200 Effective ratio = 1 : 1 Dec 14, 2007
26
Outline Introduction Lottery : Design and Implementation Evaluation
Conclusions Dec 14, 2007
27
Exp #1 : Does it work ? Setup One AP, One client
UDP Flows from AP to client Flow 1 : Flow 2 = 1 : 2 Dec 14, 2007
28
Exp #1 : Does it work ? OFF Dec 14, 2007
29
Exp #1 : Does it work ? OFF ON (Desired ratio 2 : 1) Dec 14, 2007
30
Exp #1 : Does it work ? ON (Desired ratio 2: 1) OFF OFF Dec 14, 2007
31
Exp #1 : Does it work ? Yes, it works! ON (Desired ratio 2: 1) OFF OFF
Dec 14, 2007
32
Exp #2 : How well does it work ?
ON (Desired ratio 2: 1) Dec 14, 2007
33
Exp #2 : How well does it work ?
Received Proportion : 1 It works pretty well ! ON (Desired ratio 2: 1) Dec 14, 2007
34
Exp #3 : Multiple clients
Setup One AP, two clients UDP Flows from AP to clients Client 1 : Flow 1, Flow 2 Client 2 : Flow 3 Flow 1: Flow 2: Flow3 = 1 : 2 : 4 Dec 14, 2007
35
Exp #3 : Multiple clients
F1 Dec 14, 2007
36
Exp #3 : Multiple clients
F1 F1, F2 Dec 14, 2007
37
Exp #3 : Multiple clients
F1 F1, F2 F1, F2, F3 Dec 14, 2007
38
Exp #3 : Multiple clients
F1 F1, F2 F1, F2, F3 Received Proportion 1 : 2.01 : 3.98 Works well across clients Dec 14, 2007
39
Exp #4 : Different protocols ?
OFF Dec 14, 2007
40
Exp #4 : Different protocols ?
OFF ON Dec 14, 2007
41
Exp #4 : Different protocols ?
Received proportion 1 : 1.93 Works well across protocols F1, F2 F1, F2, F3 OFF ON Dec 14, 2007
42
Exp #4 : Effect of packet size
OFF Dec 14, 2007
43
Exp #4 : Effect of packet size
ON (NO Inflation) OFF Dec 14, 2007
44
Exp #4 : Effect of packet size
ON (NO Inflation) ON (with Inflation) OFF Dec 14, 2007
45
Exp #4 : Effect of packet size
Agg. Throughput 6.94 Mbps Agg. Throughput 5.83 Mbps ON (NO Inflation) ON (with Inflation) OFF Dec 14, 2007
46
Exp #4 : Effect of packet size
F1, F2, F3 Ticket Inflation makes packet size irrelevant Tradeoff: Efficiency Vs Fairness ON (NO Inflation) ON (with Inflation) OFF Dec 14, 2007
47
Exp #5 : Effect of Bursty Traffic
Dec 14, 2007
48
Exp #5 : Effect of Bursty Traffic
No Contention Contention Dec 14, 2007
49
Exp #5 : Effect of Bursty Traffic
No Contention Contention Lottery makes sense when there is contention Dec 14, 2007
50
Exp #6 : Is it Scalable ? Dec 14, 2007
51
Lottery is scalable, does fine-grained allocation!
Exp #6 : Is it Scalable ? Contention Lottery is scalable, does fine-grained allocation! Dec 14, 2007
52
Exp #7 : Lottery Vs Stride
Dec 14, 2007
53
Exp #7 : Lottery Vs Stride
Stride is more accurate than Lottery But needs more state, processing Dec 14, 2007
54
Exp #8 : TCP and Lossy Links
Bad Links Good Links Dec 14, 2007
55
Conclusions Lottery Scheduling
Flexible, Fine-grained bandwidth allocation Works across different clients, packet sizes, traffic types, protocols Interaction with TCP in presence of lossy links Simple and a powerful concept Dec 14, 2007
Similar presentations
© 2024 SlidePlayer.com. Inc.
All rights reserved.