Active Queue Management - What it is, and why you should care
summary
Explains concrete outcomes: an outage cascades from one node restart into a worldwide meltdown that takes 3 hours to recover, after growth of about four orders of magnitude to roughly 3,000 simultaneous calls, so capacity planning alone fails. Describes specific AQM mechanisms—RED, FRED, CoDel (Kodell), preferential and weighted drop—points to an Erlang implementation in Jesper’s u safety valve, and gives actionable tactics like programmatic retry-after and timing-based dropping instead of naive tail drop.
AI-written summary · about
chapters (44)
- 0:00Intro
- 3:31My company
- 4:16The numbers
- 5:28The infrastructure
- 6:06The Taj Mahal
- 6:54What happened
- 7:06What do you do
- 7:30One hour before
- 7:43Worldwide outages are bad
- 8:18What happens when you dont think about emergent properties
- 10:25The modern age
- 11:06Total system meltdown
- 11:39Black swan
- 12:37Election 2008
- 14:08Romney 2012
- 16:03What if you dont have enough
- 18:26Buffer Bloat
- 18:43Put everything in the queue
- 19:12Back pressure
- 19:33Possible problems
- 20:03Possible solutions
- 20:18Programmable
- 20:45Behavioral techniques
- 21:21Selfmanaged techniques
- 22:00Best practitioners
- 22:14Airlines
- 22:28Cancel
- 23:05Not making everybody happy
- 24:05Air airlines do this
- 24:30Tail drop
- 25:40Van Jacobson
- 26:12TCPIP
- 26:37Random Early Drop
- 27:57Its an idea
- 28:41Thread
- 29:24Preferential Drop
- 31:13Adaptive Red
- 34:28CodeL
- 34:47Read the paper
- 36:04What is slow start
- 37:30Edge cases
- 37:54Testing
- 38:18Black swans
- 39:49Problem you should know
related talks
22mQoE Impact from Router Buffer sizing and Active Queue Management
20mMulti-Queue Fair Queuing
46mQueueing Theory in Practice: Performance Modeling for the Working Engineer
45mEverything Will Flow
30mQueueing Theory, In Practice: Performance Modelling in Cloud-Native Territory [I]
30mWhat's the Cost of a Millisecond?
43mSleeping on the job
06mPractical performance theory