1. Multi-Layer Traffic Engineering in IP over Optical Networks October 20, 2004 Hung-Ying Tyan Department of Electrical Engineering National Sun Yat-sen University

2. 2 Outline IP Network Transport Network Traffic Engineering Some Observations Multi-Layer Traffic Engineering Our ML-TE Framework Our ML-TE Algorithms Evaluation

3. 3 IP Network LAN MAN, WAN Company School Enterprise Internet Service Providers (ISP) Carriers router

4. 4 IP Network OXC (Optical) Transport Network

5. 5 IP Network over OTN OXC Conceptual view Actual IP link = circuit Data center

6. 6 Transport Network Evolved from traditional telecommunications networks Good at long distance transmission of digital signal Technologies  Synchronous Optical Network (SONET)  Wavelength Division Multiplexing (WDM)  Providing long-term circuits between end points Separation from application networks  Network on network; “overlay network”  Business tiers: carriers vs ISPs

7. 7 Traffic Engineering (TE) Mechanisms to allocate network resources according to traffic demand  ISP: Make better use of resources ($$$) Static: Network planning/provisioning/optimization Dynamic: Resources allocation adapts to traffic change

8. 8 Dynamic TE Basic idea: Move traffic around to alleviate congestion Why is it effective?  Data traffic can be bursty  Special events occur more frequently in data networks

9. 9 Observations ISPs and carriers want to provide better service with less cost Over-provisioning because of slow response to adding capacity and large variation in traffic demand  Utilization < 25% Current dynamic TE is still limited  Only deal with congestion

10. 10 Large Daily Traffic Variation OC-48 link between Dallas and Washington DC

11. 11 More Observations “Information Super Highway”? Distribution channel of electronic information products Electronic post office

12. 12 Technology Advances Control Plane Technology  A separate network dedicated to resources control  Allows resources to be added or released quickly Optical devices and equipments  Optical laser, receiver, filter etc  Wavelength conversion  Optical add-drop multiplexer (OADM)  Optical cross connect (OXC)

13. 13 New TE Paradigm – Multi-Layer TE For ISP, IP links can be leased or released on demand  IP network topology can be changed on demand  Let IP network topology adapt to actual traffic demand 4 3 4 3 2 2 2 Peak Hours Off-Peak Hours

14. 14 Value proposition For ISP  OPEX reduction  Simplified network planning For Carrier  New applications/customers for Carrier  Increased (overall) revenues  Improved resource efficiency  More revenue from the same resources

15. 15 Network Model Two-layer overlay  IP/MPLS network  Optical network Assume that Optical TE is already available OXC

16. 16 Our MLTE Framework Input -Traffic matrix -Physical topology -etc Initial provisioning MPLS-TE Hybrid path routing Activate new IP links Remove idle IP links under-utilization congestion Network monitoring no yes Cost down?

17. 17 Network Monitoring & MPLS-TE IP/MPLS Network 1 2 1. Monitor outgoing IP links Detect congestion (if utilization > TH_high) Detect underutilization ( if utilization < TH_low) 2. Select target LSPs and notify ingress nodes 3. Ingress node attempts to re-route LSPs 3

18. 18 OXC Optical Network IP/MPLS Network OXC Optical fiber Hybrid Path Routing Augmented topology information from optical layer: candidate links Hybrid path consisting of  Existing IP links  Candidate links Special cost function for both congestion and under-utilization

19. 19 Define Network_Cost = sum( Link_Cost i ) Link_Cost i = F(Link_Utilization i ) x real_link_cost i Algorithm:  Triggered by congested or under- utilized links  Dijkstra’s shortest path  d(link_cost i )= F(expected_link_utilization i ) – F(link_utilization i )  Granting a new route only if it decreases the real network cost Hybrid Path Routing F link_utilization UH

20. 20 North America Model SF LA Dallas Atlanta Miami DC NYC Boston Cleveland Detroit Chicago Denver Kansas City Seattle OXC 14 Nodes 24 Links (fiber) 193 LSPs LSP Demand Time (hr) High Low 816 0

21. 21 Experiment Results Simulation tool: J-Sim (www.j-sim.org)www.j-sim.org North America Model  14 nodes, 24 links, 193 LSPs  Average # of IP links ~ 21  # of IP links at peak demand = 33  Cost saving ~ 36% v.s. over-provisioning  Tradeoff between cost and number of LSP reroutes

22. 22 Visualization Tool

23. 23 Research Topics ML-TE framework TE operations  MPLS-TE procedure and Optical-TE Topology transformation algorithm Hybrid path routing algorithm  Suitable for both congestion and under-utilization

24. 24 Thank you! Question?

25. 25 Basic workflow