1 / 25

Routing Dynamics in Simultaneous Overlay Networks

Routing Dynamics in Simultaneous Overlay Networks. Mukund Seshadri Randy Katz ( mukunds@cs.berkeley.edu randy@cs.berkeley.edu ) Berkeley-Helsinki Short Course Aug. 2003. Problem. Consider overlay routing when multiple independent overlay networks/flows interact:

ingrid
Télécharger la présentation

Routing Dynamics in Simultaneous Overlay Networks

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Routing Dynamics in Simultaneous Overlay Networks Mukund Seshadri Randy Katz (mukunds@cs.berkeley.edurandy@cs.berkeley.edu) Berkeley-Helsinki Short Course Aug. 2003

  2. Problem • Consider overlay routing when multiple independent overlay networks/flows interact: • Can this be unstable/inefficient? • Identify such scenarios. • Suggest improvements. • Identify scope for reduction of measurement overhead.

  3. General Motivation • End-host controlled routing can become significant • Pure Overlay Network protocols (RON[3], Detour[4], ESM[5]) • Overlay primitives (“Path reflection”[1], i3-based [2]) • Better routing than Internet/BGP (resilience/performance/multicast/etc.) • What if several entities set up their own overlays? • Companies setting up distribution overlay networks… • Or, more ad-hoc users setting up overlay networks… • Flows within a single overlay… • Consider overlay networks/flows which have some physical links in common, but don’t explicitly coordinate with each other.

  4. Unstable Routing Example 1+ Mbps (L2) Primary Paths Alternate Paths 2 Mbps L1 Sources Bottleneck Phy. Link Destinations 1 Mbps (L3) Ov.Nw. Nodes (2 Ovns) • L1 failure can cause synchronized oscillation of both flows between the two alternate paths

  5. Focus • Main application – multimedia streams • Long-lived (medium) flows : ~ 1hr (5min) . • Flows require specified bandwidth levels • Flows require route stability (Packet-reordering, jitter undesirable) • Secondary app – long high volume transfers/sessions • Problem considered: selection of best routes (not location/DHTs) • Size: 50-500 overlay flows; 10-50 nodes each. • Independent decision makers - no explicit info. sharing • Unlike PlanetLab[6], underlay[7] model, i3-based soln.[2] • Independent administration might be desirable. • Don’t have to wait for infrastructure nodes to come up. • Most protocols like ESM can’t scale to thousands of nodes.

  6. Overlay Network Model • Given M overlay networks/flows with N nodes each • Probing of all potential paths is done (O(N) cost). • Path characteristics are inferred from probes in some time window • With some error factor • We consider only bandwidth • Best path is selected to send traffic on (GREEDY) • Route change based on bandwidth improvement threshold (H) • Path-level simulator • Characterizes shared bottleneck links. • The level of sharing is characterized by “path density” • Unicast CBR flows with bandwidth requirement. • Metrics of interest • Loss Rate (related to bandwidth) • Stabilization time

  7. Contribution • Study the need for “restraint” in route selection • Randomness in selection selection • Hysteresis • Time between re-route decisions

  8. Hysteresis Required • No hysteresis threshold (H) for route change => unstable. • We will use 99% stabilization time.

  9. H affects loss rate… • Will explore more later in the talk…

  10. When does Greedy “fail”? • Defaults: • 500 overlay flows, • 50 bottleneck links • link capacities ~ flow requirements • ~50% cross-traffic • 10% measurement error. • 4x variation in link b/w. • ~25 links/flow (density) • Optimal Threshold Assumed • Large flows => more effect when re-routed => lower stability

  11. When does Greedy “fail”? • High sharing=>many route-changes • Flows within a single overlay. • when overlay nodes are skewed towards certain ASes, like univ.s. • if several overlay flows independently use a medium size shared infrastructure.

  12. Cross-Traffic • High Cross-Traffic causes the effect of overlay flows on available bandwidths to be lower, so greedy is more stable. • Other factors investigated: routing window variation, measurement error, excess capacity, bandwidth distribution.

  13. Summary of “Greedy” • The following factors contribute to poor stability and performance of “Greedy” overlay path selection • Several flows’ paths share a large number of bottleneck links. • There is not much spare capacity in paths used. • There is a large variation in link and flow bandwidths. • The overlay traffic is a high fraction of traffic on the bottleneck links • Each flow’s bandwidth is significant compared to bottleneck link bandwidth.

  14. Improvements to Greedy • Randomly select path to be chosen • ARAND: In proportion to available bandwidths • SRAND: Best of randomly selected subset of size S • …in proportion to capacity • Reduces measurement overhead • Works well for server load balancing [8] • (but different work model: jobs arrive and leave, and are assigned to only one server for their lifetime) • GRAND: Randomly select from the best S paths

  15. Does Randomizing Help? • Randomization more useful at high densities. • More stable, lower loss, less sensitive to threshold setting.

  16. Hysteresis Threshold • Optimal value of H very sensitive to parameters. • Flows can automatically discover the values of H. • Flows can independently “probe” values of H • No route change => decrease H • Route change => increase H • Try AIAD, MIMD, etc. • Can perform even better than with fixed H…

  17. Exploring “H” • Very similar, MIMD stabilizes slightly quicker… • I/D pmtrs. not as sensitive to simulated network pmtrs. as H.

  18. Exploring “H” (Contd.) • Performs much better than with fixed threshold, loss rates close to 0 • Stabilization times similar to fixed case.

  19. Summary • SRAND is as good as or better than GREEDY in most cases • Measurement costs lowered, with performance similar to the proportional randomization method. • Automatic discovery of H works better than fixed H (and is more feasible). • Increasing time windows can help, particularly when flows arrive/depart.

  20. Future Work • Define a general method that combines randomization, hysteresis estimation, and time variation (like simulated annealing) • Explore dynamic scenarios (flows arrive/depart). • Explore 2nd level control loop for MIMD pmtrs. • Implement/simulate using real topologies. • Can we define a general notion of “friendliness” pertaining to both route selection and traffic distribution over different routes?

  21. References • Network layer Support for Overlay Networks – John Jannotti – OpenArch 2002. • Infrastructure Primitives for Overlay Networks – Karthik Lakshminarayanan et al. – under submission. • Resilient Overlay Networks – Andersen et al – SOSP 2001 • Detour: a Case for Informed Routing and Transport – Savage et al. – IEEE Micro Jan 1999. • A Case for End System Multicast – Yang-hua Chu et al. – JSAC 2002. • PlanetLab – http://www.planet-lab.org • A Routing Underlay for Overlay Networks – Nakao et al. – Sigcomm 2003. • How Useful is Old Information – M.Mitzenmacher – PODC 1997 • An Analysis of Internet Content Delivery Systems – Saroiu et al. – OSDI 2002.

  22. …Backup Slides…

  23. Stabilization Times of the *RANDs • Generally SRAND and ARAND stabilize quickly and have a very low loss rate. • Also investigated the effect of subset size on SRAND

  24. Other Factors • Small amount of cheating doesn’t hurt the good flows, large amount does. • If link bandwidths are much higher than flow bandwidths, Greedy is more stable and performs better. • If link and flow BW are similar, then a high variation in the same causes Greedy to be fairly unstable.

  25. Extra Slide2-Flow Illustration • We can randomize • Route selection • Proportional to Available BW • Time intervals • Of assessment and rerouting.

More Related