1. The Delta Routing Project Low-loss Routing for Hybrid Private Networks George Porter (UCB) Minwen Ji, Ph.D. (SRC - HP Labs)

2. Outline Motivation/overview of corporate networks Problem Statement Architecture –Two layers: Physical and Overlay –The Delta Protocol –The Delta+TM Protocol Evaluation Conclusions

3. Corporate Network Construction Network Layer Distributed Locations connected by leased lines due to: –Need for predictable performance –Security –Management and control Fixed initial cost, incremental additional cost due to traffic volume Not necessarily overprovisioned Reprovisioning on the timescale of days (or weeks) Expensive (compared to ISP connectivity SF NYC LON LA DC DFW ALX SEA HOU OSPF

4. Corporate Network Construction Overlay Layer ISP Connectivity alreay at selected nodes to provide: –Web/Email access –VPN access to at-home or distance workers –Business services Per-byte, ISP much cheaper than “Intranet” –But no QoS Intranet corporate network with ISP links is called a ‘Hybrid Private Network’ SF NYC LON LA DC DFW ALX SEA HOU ISP Connectivity

5. Problem of Congestion Flash traffic (video, backup, data transfer) or steady corporate growth can lead to periodic congestion Problem Statement: –Reduce congestion and packet loss on the Intranet by utilizing ISP connectivity while providing good end- to-end performance SF NYC LON LA DC DFW ALX SEA HOU

6. Architecture

7. Overlay Layer: - Need to forward traffic around congested portions of the Intranet - Measurement-based path construction - Intermediate point may be better than “last hop” selection - Metric include measured latency and local queuing delay - Paths are selected on order of seconds or minutes Physical (Intranet) Layer: - Single-domain routing protocol (OSPF) - Dijkstra - Forwarding decision: which packets go to Intranet and which go to the preselected overlay paths? (per packet)

8. Physical Forwarding Algorithm Ji, Minwen. Dial-controlled Hash: Reducing Path Oscillation in Multipath Networks. Proceedings of the International Conference on Computer Communications and Networks (ICCCN). Oct 2003. Current Algorithm: –Prefer physical path, but if physical queue full send to overlay layer.

9. Overlay Path Selection Algorithms Static –Lasthop –Nexthop –Random Dynamic –Delta Minimize end-to-end delay –Delta+TM Predict and avoid congestion by inferring global traffic matrix

10. Delta Path Selection Find path to minimize the sum of: –Local Queue delay + WAN delay + Intranet delay Key feature is the use of locally obtained information

11. Limitation of Delta Algorithm Since Delta uses local information, it might send traffic to an overloaded link: congested Can we avoid this?

12. Delta+TM (Traffic Matrix) Key idea: –Don’t choose paths that will subject the traffic to congestion Use the original Delta algorithm (minimize end- to-end delay) but throw out paths that will subject packets to congestion But how do we find out about remote congestion? –Given that message flooding will likely be inaccurate and might make the problem worse

13. Traffic Matrix Estimation 1234 10.004.541.921.14 21.040.001.571.78 32.523.040.004.08 43.644.304.130.00 + Topology Information =.3.6 1.2.03 1.3.89 Each node measures flows that transit through it Long-term averages are flooded to fill in the entries of the table that a node can’t directly measure

14. Evaluation 1)Simple Example 2)Algorithm-antagonistic Topologies 3)Large-scale Topology (PlanetLab- based)

15. Linear Topology

16. Congestion Event

18. Evaluation 1)Simple Example 2)Algorithm-antagonistic Topologies 3)Large-scale Topology (PlanetLab- based)

19. Algorithm-antagonistic Topology Simple topology with traffic flows that should expose a weakness to each topology

20. Algorithm-antagonistic Topology

21. Evaluation 1)Simple Example 2)Algorithm-antagonistic Topologies 3)Large-scale Topology (PlanetLab- based)

22. Planetlab as VPN-network source Large, distributed testbed We modelled the Overlay part of a fictional 43-node corporate network using traces taken over planetlab The Intranet link topology was obtained from 2-level clustering and eyeballing Traffic flows include a “measured flow” and a set of background and disruptive flows

23. PlanetLab (UCLA->ac.uk) ppnonly4,438 nexthop2,492 lasthop3,469 DeltaTM3,523 Delta4,443 Packet Losses

24. Conclusions Utilizing ISP connectivity enables balancing packet loss rate –vs- end-to-end delays Dynamic algorithms can adapt to a variety of wide-area conditions Congestion prediction can help in certain environments, however local-only decision making works well Certain “choke points” must be identified so that synchronization effects will not occur Making better use of bandwidth can lower cost of deploying distributed corporate networks