Game-Theoretic Network Bandwidth Distribution for Self- Adaptive - PowerPoint PPT Presentation

Game-Theoretic Network Bandwidth Distribution for Self- Adaptive Cameras Gautham Nayak Seetanadi 1 Martina Maggio, Karl-Erik Å rzen 1 Luis Almeida 2 Luis Oliveira 3 1 Department of Automatic Control, Lund University 2 Universidade do Porto / Faculdade de Engenharia 3 Department of Computer Science, University of Pittsburgh

Introduction • Multiple adaptive camera network • Game-theoretic resource manager • Improved bandwidth utilisation • Local PI Control • Convergence guarantees on bandwidth

Motivation • Problem of Allocation • Finite Resource • Multiple Entities • Fair Bandwidth Distribution • Global Threshold Limit • Successfully applied to Multicore systems [1] [1] M.Maggio et al, “A game-theoretic resource manager for rt applications”, In Euromicro Conference on Real-Time Systems, 2013

Decentralised Resource Allocation • Resource Manager Allocates Resources • Cameras Choose Service Level • Change Quality Data Control Camera 1 Camera 2 Camera n λ 1 λ 2 λ n Resource Manager Ethernet

Decentralised Resource Allocation • Resource Manager Allocates Resources • Cameras Choose Service Level • Change Quality • Determines Responsibility λ i Data Control Camera 1 Camera 2 Camera n λ 1 λ 2 λ n Resource Manager Ethernet

Resource Manager • Game Theory Based • Matching function determines fairness f p,w = B p,w − s p,w B p,w

Resource Manager

Resource Manager t = 0 t = 1 t = 2 Manager 50% Camera 1 50% Camera 2

Resource Manager t = 0 t = 1 t = 2 Manager 50% 75% Camera 1 50% 25% Camera 2

Camera Original Frames Encoded Frames

Camera Can be • Camera generates an image of size -ve or +ve s p,w = 0 . 01 · q p,w · s p, max + δ s p,w , • Adaptable quality • Normalised error

Camera Loop Allocated Frame Bandwidth Error Quality Size ( B p,w ) ( e p,w ) ( q p,w ) ( s p,t w ) Controller P Camera ( k p , k i ) − 1

Network Allocation • SDN using Flexible Time Triggered - SE [2] t = 0 t = 1 t = 2 Manager Camera 1 Camera 2 [2] P .Pedreiras and L.Almeida, “The flexible time triggered paradigm. An approach to QOS management in distributed real-time systems”, 2003

NETWORK ALLOCATION • SDN using Flexible Time Triggered - SE [2] t = 0 t = 1 t = 2 Manager I 1 , 1 I 1 , 2 Camera 1 Camera 2 I 2 , 1 I 2 , 2 [2] P .Pedreiras and L.Almeida, “The flexible time triggered paradigm. An approach to QOS management in distributed real-time systems”, 2003

NETWORK ALLOCATION • SDN using Flexible Time Triggered - SE [2] I 1 , 2 : q 1 , 2 → s 1 , 2 B 1 ,t =0 → B 1 ,w =2 t = 0 t = 1 t = 2 Manager I 1 , 1 I 1 , 2 Camera 1 Camera 2 I 2 , 1 I 2 , 2 [2] P .Pedreiras and L.Almeida, “The flexible time triggered paradigm. An approach to QOS management in distributed real-time systems”, 2003

NETWORK ALLOCATION • SDN using Flexible Time Triggered - SE [2] I 1 , 2 : q 1 , 2 → s 1 , 2 B 1 ,t =0 → B 1 ,w =2 t = 0 t = 1 t = 2 Manager I 1 , 1 I 1 , 2 I 1 , 4 I 1 , 5 Camera 1 Camera 2 I 2 , 1 I 2 , 2 I 2 , 3 I 2 , 4 [2] P .Pedreiras and L.Almeida, “The flexible time triggered paradigm. An approach to QOS management in distributed real-time systems”, 2003

Implementation • Cameras • Logitech c270. COTS • OpenCV • Network • Deadlines enforced using Flexible Time Triggered Switched Ethernet (FTT-SE) • i7-4790 8 core PC running Fedora

Experimental Setup Blue Camera Red Camera

Assessment • Criteria? • Fair Bandwidth Usage (Manager) • Complete Bandwidth Utilisation (Camera) • SSIM (Structural Similarity Index) • SSIM

Experiments Resource Allocator Camera Equal Bandwidth Distribution No Adaptation Equal Bandwidth Distribution DARTES Model [2] Equal Bandwidth Distribution PI Controller Game-Theoretic RA PI Controller [2] J. Silvestre-Blanes, L. Almeida, R. Marau, and P . Pedreiras. “Online qos management for multimedia real-time transmission in industrial networks.” IEEE Transactions on Industrial Electronics, 58(3), March 2011

Experiment 1 AllocBW Camera 1 AllocBW Camera 2 InstBW Camera 1 InstBW Camera 2 30 . 0 BW [Mbps] 20 . 0 10 . 0 0 . 0 0 10 20 30 40 50 60 70 80 Time [s] Resource Allocator Camera Equal Bandwidth Distribution No Adaptation

Experiment 1 AllocBW Camera 1 AllocBW Camera 2 InstBW Camera 1 InstBW Camera 2 30 . 0 BW [Mbps] 20 . 0 10 . 0 0 . 0 0 10 20 30 40 50 60 70 80 Time [s] Resource Allocator Camera Equal Bandwidth Distribution No Adaptation

Experiment 1 AllocBW Camera 1 AllocBW Camera 2 InstBW Camera 1 InstBW Camera 2 30 . 0 BW [Mbps] 20 . 0 10 . 0 0 . 0 0 10 20 30 40 50 60 70 80 Time [s] Resource Allocator Camera Equal Bandwidth Distribution No Adaptation

Experiment 2 AllocBW Camera 1 AllocBW Camera 2 InstBW Camera 1 InstBW Camera 2 4 . 0 3 . 0 BW [Mbps] 2 . 0 1 . 0 0 . 0 0 10 20 30 40 50 60 70 80 90 100 110 120 Time [s] Resource Allocator Camera Equal Bandwidth Distribution DARTES Model

Experiment 2 AllocBW Camera 1 AllocBW Camera 2 InstBW Camera 1 InstBW Camera 2 4 . 0 3 . 0 BW [Mbps] 2 . 0 1 . 0 0 . 0 0 10 20 30 40 50 60 70 80 90 100 110 120 Time [s] Resource Allocator Camera Equal Bandwidth Distribution DARTES Model

Experiment 3 AllocBW Camera 1 AllocBW Camera 2 InstBW Camera 1 InstBW Camera 2 4 . 0 3 . 0 BW [Mbps] 2 . 0 1 . 0 0 . 0 0 10 20 30 40 50 60 70 80 90 100 110 120 Time [s] Resource Allocator Camera Equal Bandwidth Distribution PI Controller

Experiment 3 AllocBW Camera 1 AllocBW Camera 2 InstBW Camera 1 InstBW Camera 2 4 . 0 3 . 0 BW [Mbps] 2 . 0 1 . 0 0 . 0 0 10 20 30 40 50 60 70 80 90 100 110 120 Time [s] Resource Allocator Camera Equal Bandwidth Distribution PI Controller

Experiment 4 AllocBW Camera 1 AllocBW Camera 2 InstBW Camera 1 InstBW Camera 2 4 . 0 3 . 0 BW [Mbps] 2 . 0 1 . 0 0 . 0 0 5 10 15 20 25 30 35 40 45 50 55 60 Time [s] Resource Allocator Camera Game-Theoretic RA PI Controller

Experiment 4 AllocBW Camera 1 AllocBW Camera 2 InstBW Camera 1 InstBW Camera 2 4 . 0 3 . 0 BW [Mbps] 2 . 0 1 . 0 0 . 0 0 5 10 15 20 25 30 35 40 45 50 55 60 Time [s] Resource Allocator Camera Game-Theoretic RA PI Controller

SSIM SSIM with PI controller − SSIM with DARTES model 0 , 1 Camera 1 Camera 2 0 , 05 0 0 200 400 600 800 1000 1200 1400 1600 Frame Number

SSIM SSIM with PI controller − SSIM with DARTES model 0 , 1 Camera 1 Camera 2 0 , 05 0 0 200 400 600 800 1000 1200 1400 1600 Frame Number

Conclusions • CPU allocation strategy • Need for adaptation • PI control advantages • Game-Theoretic resource manager e ffi ciency • Convergence guarantees

Future Work • Time triggered network manager decisions have drawbacks • Moving to event triggered

gautham@control.lth.se Resource Allocator Camera AllocBW Camera 1 AllocBW Camera 2 AllocBW Camera 1 AllocBW Camera 2 InstBW Camera 1 InstBW Camera 2 InstBW Camera 1 InstBW Camera 2 4 . 0 4 . 0 3 . 0 3 . 0 BW [Mbps] BW [Mbps] 2 . 0 2 . 0 1 . 0 1 . 0 0 . 0 0 . 0 0 10 20 30 40 50 60 70 80 90 100 110 120 0 10 20 30 40 50 60 70 80 90 100 110 120 Time [s] Time [s] AllocBW Camera 1 AllocBW Camera 2 InstBW Camera 1 InstBW Camera 2 4 . 0 3 . 0 BW [Mbps] 2 . 0 1 . 0 0 . 0 0 5 10 15 20 25 30 35 40 45 50 55 60 Time [s]