1. 1 Latency Equalization: A Programmable Routing Service Primitive Minlan Yu minlanyu@cs.princeton.edu Joint work with Marina Thottan, Li Li at Bell Labs PRESTO’08

2. 2 Motivation Latency EQualization (LEQ) service Online interactive applications require equalized delay among multiple users Online gaming Players vote to exclude players with higher lag Distributed online music concert Delay difference among musicians at different places degrade music quality Online trading Unfair advantage to shopping agents with lower delay

3. 3 Latency Compensation Techniques Server side solutions Buffer packets till all clients response arrive Expensive due to processing overhead Limits on number of clients Client side solutions Buffer the packets to wait for other clients Requires coordination among clients Easy to cheat We need network service for LEQ

4. 4 LEQ Service Reducing latency difference Difference of Maximum delay and minimum delay client1 client2 server client3

5. 5 LEQ Architecture - Basic Idea Avoid changing every router Place a few hub routers in the network Add customized logic to hub routers Redirect traffic through hub routers server client1 client2client3 hub

6. Easy to Deploy Require only a few programmable routers Allow incremental deployment With one hub in the network, we can reduce delay difference by 40% on average compared with OSPF No modification of underlying routing protocols Can be implemented as an overlay

7. LEQ Architecture Architecture Select a set of hubs for each client Set up tunnels between clients and hubs, hubs and servers. MPLS tunnels or packet encapsulation Hub routers redirect packets to servers

8. 8 Hub Router Implementation Packet classifier Identify application from src, dst, port Identify class of packets within an application Initial game setup packets: shortest path routing Interactive event packets: LEQ routing FIB Packet classifier

9. 9 Hub Router Implementation Customize routing for each application Different applications may have different max delay bound; different client, server location LEQ routing for gaming, live concert, trading OSPF routing for other applications Control Plane LEQ routing for gaming OSPF LEQ routing For concert LEQ routing For trading FIB

10. Other benefits of hubs Add service-specific logic to a few enhanced routers in the network Application level packet processing Gaming: Update aggregation, packet inspection Concert: Echo cancellation Multicast Server multicast to hubs Hubs then multicast to their clients Load balancing among servers Hubs can select the least-loaded server for the client 10

11. 11 Hub Placement Problem Input Location of client/server edge routers Maximum number of hub routers (M) Number of hub routers per client (m) Output: A set of m hubs for each client Goal: Minimize delay difference among clients

12. 12 Problem Complexity We proved it is NP-hard and inapproximable Reduced to set cover problem Greedy hub placement algorithm Based on multi-set cover algorithm See details in the paper

13. 13 Evaluation Static analysis Use Rocketfuel Data with different ISPs Focus on lightly loaded network propagation delay Result LEQ achieve 80% reduction of latency difference compared with shortest path routing Reduce from 35ms to 5ms Only need to place 5 hubs in the network Similar maximum delay of LEQ and OSPF

14. 14 Evaluation (cont.) Dynamic Analysis Under dynamic traffic condition Focus on congested network Consider both propagation delay and queuing delay Result LEQ routing can get around congestion

15. 15 Alternative Network-based Solutions OSPF No delay difference consideration Tune weights for each application Computationally hard Source routing Clients need to know the global network condition Require collaboration among clients Hard to compute and implement

16. 16 Conclusion Interactive online applications New requirement on delay difference Latency Equalization service Place a few hubs to provide latency-equalized, reliable paths Deployment on programmable routers Easy to implement and deploy

17. 17 Acknowledgement Grenville Armitage, Swinburne University of Technology Wu-Chang Feng, Washington State University Jennifer Rexford, Princeton University Thomas Woo, Bell Labs

18. 18 Thanks! Q&A?

19. 19 Static Analysis Telstra network 80% 5 hubs are sufficient

20. 20 Static Analysis AT&T network

21. 21 Premise of LEQ: Trading delay for delay difference