End-to-End Network Issues

Slides: End-to-End Network Issues

[Start Here (html)] [Slide Sorter (html)]

Download: [Postscript: DoE200306.ps] [PDF: DoE200306.pdf]

Page 1: End-to-End Network Issues

Page 2: What causes end-to-end bottlenecks?

Page 3: The Wizard Gap

Page 4: Imagine closing the wizard gap

Page 5: Moore's law

Page 6: The spoiler

Page 7: Background: What is a protocol?

Page 8: Example: Fat TCP

Page 9: Legacy

Page 10: Background: What is the Internet?

Page 11: Technical Problems

Page 12: "Tuning" requires too much expertise

Page 13: Tuning is really debugging

Page 14: Tuning is painful debugging

Page 15: Web100 Instrumentation

Page 16: False results

Page 17: Congestion Control

Page 18: Reinventing Congestion Control

Page 19: Change global congestion control?

Page 20: A better solution

Page 21: Key point

Page 22: False Results

Page 23: Remote Direct Data Placement (RDDP)

Page 24: Technical pieces

Page 25: RDDP Implementation

Page 26: False Results

Page 27: TCPs 32 bit Sequence Limit

Page 28: What about MTU?

Page 29: Maximum Transmission Unit

Page 30: Possible approaches

Page 31: Apply Moore's law to packet sizes

Page 32: A different view

Page 33: A serious legacy bug

Page 34: Impact of disabling pMTU discovery

Page 35: Two intertwined subtasks

Page 36: The new algorithm

Page 37: Running code

Page 38: LAN vendor resistance

Page 39: Huge payback

Page 40: Maximizing DoE Impact

Page 41: Epilogue

Page 42: TCP is easy to blame, but...

Page 43: Fix the real problems

Page 44: Organizational problems

Page 45:

Generated by MagicPoint