Peering vs. Transit Adnan Ahmed University of Iowa University of - PowerPoint PPT Presentation

Peering vs. Transit Adnan Ahmed University of Iowa University of Iowa

Introduction • Transit • Provides connectivity to the Internet • Traffic volume based fees • Peering • Bilateral exchange • Settlement-free (no fee) University of Iowa 1. Introduction

Related work Interconnection strategies in peering ecosystem • Agent-based analysis [Lodhi-Dhamdhere, SIGMETRICS ’12] • Open-peering [Lodhi et al., Infocom ’14] • Game-theoretic models [Accongiagioco et al., IFIP ’14][Badasyan-Chakrabarti, Telecommunications Policy ’08] • Complexities in decision making [Lodhi et al., Infocom ‘15] Evolution of peering and topological Impact • Network model [Dhamdhere-Dovrolis, CoNEXT ’10] • Remote peering [Castro et al., CoNEXT ‘14] • IXP study [Ager et al., SIGCOMM ‘12] • PeeringDB analysis [Lodhi et al., SIGCOMM CCR ’14] Our focus • Large-scale measurement based performance comparison University of Iowa 2. Related work

Methodology • Throughput measurements • Strain the network • Delay measurements using ICMP packet probing • Rate limiting at ISPs • Our approach • HTTP based end-to-end delay measurements University of Iowa 3. Methodology

The big picture University of Iowa 4. The big picture

The big picture University of Iowa 4. The big picture

The big picture University of Iowa 4. The big picture

Our approach University of Iowa 5. Our approach

Our approach University of Iowa 5. Our approach

Our approach University of Iowa 5. Our approach

Our approach University of Iowa 5. Our approach

Our approach University of Iowa 5. Our approach

Data collection • A commercial CDN • Collected across PoPs at 19 IXPs • 1M measurements • ~350K clients • 360 Autonomous systems University of Iowa 6. Data collection

Peering vs Transit University of Iowa 7. Peering vs Transit

Peering vs Transit University of Iowa 7. Peering vs Transit

Peering vs Transit University of Iowa 7. Peering vs Transit

Peering vs Transit University of Iowa 7. Peering vs Transit

RTT components University of Iowa 8. RTT components

RTT components Transmission ~ 0 • small size of pixel tag • University of Iowa 8. RTT components

RTT components Transmission ~ 0 • small size of pixel tag • RTT ~ Propagation delay + Queueing delay • University of Iowa 8. RTT components

RTT components RTT measurements over time University of Iowa 8. RTT components

RTT components Diurnal pattern • RTT measurements over time University of Iowa 8. RTT components

RTT components Diurnal pattern • Time series latency pings • University of Iowa 8. RTT components

RTT components Diurnal pattern • Time series latency pings • University of Iowa 8. RTT components

RTT components Diurnal pattern • Time series latency pings • Propagation delay < RTT min • University of Iowa 8. RTT components

RTT components Diurnal pattern • Time series latency pings • Propagation delay < RTT min • University of Iowa 8. RTT components

RTT components Diurnal pattern • Time series latency pings • Propagation delay < RTT min • University of Iowa 8. RTT components

RTT components Diurnal pattern • Time series latency pings • Propagation delay < RTT min • Maximum queueing delay • ~ RTT max - RTT min University of Iowa 8. RTT components

Propagation and Queueing delays University of Iowa 9. Propagation and Queueing delays

Propagation and Queueing delays Peering is better University of Iowa 9. Propagation and Queueing delays

Propagation and Queueing delays Transit is better University of Iowa 9. Propagation and Queueing delays

Propagation and Queueing delays II I III IV University of Iowa 9. Propagation and Queueing delays

Propagation and Queueing delays II I III IV University of Iowa 9. Propagation and Queueing delays

Propagation and Queueing delays II I III IV University of Iowa 9. Propagation and Queueing delays

Path length comparison University of Iowa 10. Path length comparison

Path length comparison University of Iowa 10. Path length comparison

Path length comparison Shorter path length via peering University of Iowa 10. Path length comparison

Conclusion • Peering generally outperforms transit for a majority of clients • Peering almost always has better propagation delays • Shorter path lengths for peering • Transit sometimes has better queueing delays • Under-provisioned peering paths University of Iowa 11. Conclusion