The Case for Resilient Overlay Networks. MIT Laboratory for Computer Science

Size: px

Start display at page:

Download "The Case for Resilient Overlay Networks. MIT Laboratory for Computer Science"

Milton Gibbs
6 years ago
Views:

1 The Case for Resilient Overlay Networks David G. Andersen Hari Balakrishnan, M. Frans Kaashoek, Robert Morris MIT Laboratory for Computer Science May

2 Idea Scalability may get in the way of deploying services and protocols that may not scale So do cool things in small overlays More aggressive Things that re less efficient Resilient Overlay Networks Slide 1

3 Routing around Internet Failures Utah MIT "The Internet" ArosNet MA Cable Modem People expect all-to-all communication... Resilient Overlay Networks Slide 2

4 Routing around Internet Failures Utah MIT ArosNet MA Cable Modem Which the Internet can t always provide. Resilient Overlay Networks Slide 3

5 Internet Failures Physical link failures (backhoes) Excess Traffic (14-year-olds...) Router misconfiguration The list goes on... Resilient Overlay Networks Slide 4

6 Routing around Internet Failures Utah MIT ArosNet MA Cable Modem But we think cooperating hosts can do better... Multiparty videoconferences Overlay Internet Service, Companies with VPNs, etc. Do we need this? Resilient Overlay Networks Slide 5

7 The Internet Recovers Slowly [Labovitz 00]: Internet... routing convergence is an order of magnitude slower than previously thought. 3 minute average recovery time 15 minute max for simple failures Our tests: Indirect routing had 5x-10x fewer outages [Sneak preview] Resilient Overlay Networks Slide 6

8 Internet Trade-offs Scalability and heterogeneity: Slow Recovery (Is this a fundamental trade-off?) RON takes a different approach: Fast recovery for small groups in an overlay Exploit redundancy in the Internet Resilient Overlay Networks Slide 7

9 A More Realistic Picture vbns / Internet 2 MIT Utah BBN ArosNet Cable Modem Resilient Overlay Networks Slide 8

10 Hidden Links vbns / Internet 2 MIT Utah BBN ArosNet Cable Modem Resilient Overlay Networks Slide 9

11 Policy and AUPs WAN routing policy expression is a sledgehammer But we need policy control (Internet2, etc) RON could violate AUPs But RON can provide flexible policies More complex routing decisions Multiple routing tables Deeper packet inspection Resilient Overlay Networks Slide 10

12 The Approach Utah ArosNet Probes Traffic Forwarding MIT Measure all inter-node paths Exchange routing information Route along best path Cable Modem Resilient Overlay Networks Slide 11

13 Architecture Node 1 Node 2 Node 3 Probes Probes Probes Probe between nodes to establish best route Active, application probing N of paths 2 Passive measurements Resilient Overlay Networks Slide 12

14 Architecture Node 1 Node 2 Node 3 Probes Forwarder Router Probes Forwarder Router Probes Forwarder Router Performance Database Probe between nodes to establish best route Link-state routing protocol between nodes Resilient Overlay Networks Slide 13

15 Architecture Data Node 1 Conduit Forwarder Node 2 Node 3 Conduit Forwarder Conduit Forwarder Probes Router Probes Router Probes Router Performance Database Probe between nodes to establish best route Link-state routing protocol between nodes Data handled by application-specific conduit Forwarded in UDP Resilient Overlay Networks Slide 14

16 Conduits: Gateways into the RON IP off the wire conduit (Used for evaluation) Emulates sendto and recvfrom The application itself Interface: send, register, callback Resilient Overlay Networks Slide 15

17 Preliminary Investigation Tested between 4 hosts 70 hours of ping-style measurements Looks promising, but we suspected it would. Resilient Overlay Networks Slide 16

18 From MIT to ArosNet on the Internet MIT to ArosNet % Packet Loss Rate Time (30 minute blocks) The direct Internet path is quite bad. Resilient Overlay Networks Slide 17

19 From MIT to Utah on the Internet 100 % Packet Loss Rate MIT to Utah Time (30 minute blocks) Internet loss rate But the path from MIT to Utah looks good... Resilient Overlay Networks Slide 18

20 From Utah to ArosNet on the Internet 100 % Packet Loss Rate Utah to ArosNet (Very low loss, except for three small bumps) Time (30 minute blocks) As does the path from Utah to ArosNet... Resilient Overlay Networks Slide 19

21 From MIT to ArosNet with RON 100 % Packet Loss Rate MIT to ArosNet Direct Internet Path RON Path Time (30 minute blocks) Putting them together... Resilient Overlay Networks Slide 20

22 Other results Big latency reduction between MIT - ArosNet Big latency reduction between Cable Modem - Utah Real results are hiding in a thesis Resilient Overlay Networks Slide 21

23 Some Research Questions Is this a stupid idea? How many intermediate hops? How do we best choose routes? How frequently do we probe? What routing policies can we express? How do RONs interact? Resilient Overlay Networks Slide 22

24 Status To vu.nl Lulea.se OR DSL CMU MIT MA Cable Cisco CA T1 CCI Aros Utah Cornell NYU Resilient Overlay Networks Slide 23

Understanding Internet Path Failures: Location, Characterization, Correlation

Understanding Internet Path Failures: Location, Characterization, Correlation Nick Feamster, David Andersen, Hari Balakrishnan M.I.T. Laboratory for Computer Science {feamster,dga,hari}@lcs.mit.edu Big