1 / 15

Follow-up to Observations on Anycast Topology and Performance

This paper presents observations on anycast topology and performance, including query distribution from several anycast systems and the selection of servers based on customer relationships. It also discusses anomalies in the data and local node performance. The paper concludes that fixing peering issues improves performance and suggests the need for a global node closer to South America.

hudsond
Download Presentation

Follow-up to Observations on Anycast Topology and Performance

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. Follow-up to Observations on Anycast Topology and Performance Steve Gibbard Packet Clearing House

  2. Original paper • Observations on Anycast Topology and Performance • Presented at last OARC/DNS Ops meeting • Looked at query distribution from several anycast systems • J Root data from Verisign • K Root data from RIPE • C, F, K Root data from CAIDA

  3. Original paper (cont.) • Observed server selection following customer relationships rather than geography for J and K roots • Transit from different providers at different nodes • ISPs prefer to send traffic to customers, regardless of geography

  4. Original paper (cont.) • Attempted to show that an anycast system being consistent about transit and peering policies avoided those issues • Did data collection from PCH anycast system • Four global nodes, transit from NTT/Teleglobe • Some peering on the global nodes • Traffic followed geography, except when it didn’t • Cases where it didn’t were anomalous

  5. Original paper -- questions • Methodology issues • Anomalies in the data • Local node performance

  6. Methodology issues • Used unique query sources (actually source /24s) instead of hit counts. • Did that distort results, or miss important networks? • Reran analysis, using hit counts, for answers: • Results mostly similar. • Weighting by query sources did have interesting effects. Half of “Belarus” queries from one network in the US.

  7. Local nodes • Left out of earlier analysis: • Initially considered out of scope. • Added in later tests. • Results more or less as expected: • Isolated regions -- Nepal, Bangladesh, Kenya, mostly self-contained. • More major nodes -- Amsterdam, Stockholm, Singapore, drew from wider, but geographically contiguous, areas.

  8. Data Anomalies • Traffic anomalies, blamed on inconsistent peering: • Southeast Europe/North Africa to Ashburn • Indian sources to Ashburn • Spanish sources to Hong Kong • Lots of Asian traffic to Palo Alto rather than Hong Kong

  9. Chasing anomalies • Southeast Europe/North Africa • 210k queries from Telecom Italia • 62k queries from UPC • We peered with both in US but not Europe. • Turned up peering, and traffic shifted.

  10. Peering shift Ashburn, before shift Ashburn, after shift Amsterdam, before shift Amsterdam, after shift

  11. Asia to Palo Alto • Much of Asia hot-potatoing to Palo Alto in January • Still doing so in March • Going to Hong Kong in May • Spot check shows these mostly seen through transit, so shift presumably in transit provider networks Palo Alto, January Hong Kong, September

  12. New issues • Turk Telecom -- Uses circuit to Eastern US. Not much we can do in short term. • Planned node in DE-CIX might fix this. • Things that changed on their own: • No longer seeing Jazz Telecom in Hong Kong -- not sure why. • Indian traffic heading East rather than West.

  13. Latin America • Going to right global node. Local nodes are closer. • Could be fixed by peering in Miami or Sao Paulo • Ending up in Ashburn: • Region-wide: 474k qpd from Telefonica. • Caribbean: 142k qpd from Columbus/New World • Brazil: • 150k qpd from Bahia • 80k qpd from Telefonica • 80k qpd from Embratel • 60k qpd from Brazil Telecom.

  14. Conclusions • Fixing peering issues does fix performance issues. • Some networks cooperate with this, while others make it difficult. • A global node closer to South America might be useful.

  15. Thanks! Paper at: http://www.pch.net/resources/papers/anycast-performance/ Steve Gibbard Packet Clearing House www.pch.net scg@pch.net

More Related