Co Dee N DNS Co DNS Vivek Pai

  • Slides: 16
Download presentation
Co. Dee. N + DNS = Co. DNS Vivek Pai, Larry Peterson, & the

Co. Dee. N + DNS = Co. DNS Vivek Pai, Larry Peterson, & the Co. Dee. N group Princeton University Oct 19, 2003 RPI Planet. Lab Meeting

Origins of Co. DNS Co. Dee. N = Open CDN on Planet. Lab 40

Origins of Co. DNS Co. Dee. N = Open CDN on Planet. Lab 40 nodes, 4+ months, 24/7 operation l Approaching 2 million reqs/day l Extensive node monitoring l Reliability by avoiding bad nodes Reliability at lower layers = delays l Delays = annoyed users l Better to avoid slow peers l Oct 19, 2003 RPI Planet. Lab Meeting 2

Oct 19, 2003 RPI Planet. Lab Meeting 3

Oct 19, 2003 RPI Planet. Lab Meeting 3

Types of Unreliability Dead nodes High application-level packet losses Resource exhaustion (FDs, disk) DNS?

Types of Unreliability Dead nodes High application-level packet losses Resource exhaustion (FDs, disk) DNS? l Yes, DNS l Lookup of origin server on cache misses l Oct 19, 2003 RPI Planet. Lab Meeting 4

DNS Problems Sometimes a major cause of slowdowns May not be node-local failure l

DNS Problems Sometimes a major cause of slowdowns May not be node-local failure l Other resources salvageable Fix may help other non-CDN services l DNS as a “utility” Oct 19, 2003 RPI Planet. Lab Meeting 5

DNS-Related Avoidance Oct 19, 2003 RPI Planet. Lab Meeting 6

DNS-Related Avoidance Oct 19, 2003 RPI Planet. Lab Meeting 6

Insight Use content distribution infrastructure to mask DNS resolver failures If local DNS lookups

Insight Use content distribution infrastructure to mask DNS resolver failures If local DNS lookups failing Send to a nearby, working node l Use locality, load balancing to avoid l Oct 19, 2003 RPI Planet. Lab Meeting 7

Prototype In place on all Co. Dee. N nodes Currently using custom infrastructure l

Prototype In place on all Co. Dee. N nodes Currently using custom infrastructure l Will soon share Co. Dee. N infrastructure l Tries to be intelligent Immediate remote lookup on persistent local failure l Backoff-based delay otherwise l Oct 19, 2003 RPI Planet. Lab Meeting 8

Cornell – Regular Problems Time & “Failures”, regular DNS Oct 19, 2003 RPI Planet.

Cornell – Regular Problems Time & “Failures”, regular DNS Oct 19, 2003 RPI Planet. Lab Meeting 9

Cornell – Regular Problems Lookup times – DNS & Co. DNS Oct 19, 2003

Cornell – Regular Problems Lookup times – DNS & Co. DNS Oct 19, 2003 RPI Planet. Lab Meeting 10

CMU – Cron Jobs? Time & Failures, regular DNS Oct 19, 2003 RPI Planet.

CMU – Cron Jobs? Time & Failures, regular DNS Oct 19, 2003 RPI Planet. Lab Meeting 11

CMU – Cron Jobs? Lookup times – DNS & Co. DNS Oct 19, 2003

CMU – Cron Jobs? Lookup times – DNS & Co. DNS Oct 19, 2003 RPI Planet. Lab Meeting 12

Ga. Tech “Disappearance” Time & Failures, regular DNS Oct 19, 2003 RPI Planet. Lab

Ga. Tech “Disappearance” Time & Failures, regular DNS Oct 19, 2003 RPI Planet. Lab Meeting 13

Ga. Tech “Disappearance” Time & Failures, regular Co. DNS Oct 19, 2003 RPI Planet.

Ga. Tech “Disappearance” Time & Failures, regular Co. DNS Oct 19, 2003 RPI Planet. Lab Meeting 14

UCLA – Failure Burst Lookup times – DNS & Co. DNS Oct 19, 2003

UCLA – Failure Burst Lookup times – DNS & Co. DNS Oct 19, 2003 RPI Planet. Lab Meeting 15

Summary DNS failures fairly widespread Co. DNS prototype very promising Additional overhead very tolerable

Summary DNS failures fairly widespread Co. DNS prototype very promising Additional overhead very tolerable Working on standalone deployment Will listen on port 53, do the right thing l Goal: general Planet. Lab service l Oct 19, 2003 RPI Planet. Lab Meeting 16