Skip to content

Instantly share code, notes, and snippets.

@finchd

finchd/seagl2018-tcp-perf.md

Last active November 13, 2018 06:46
Show Gist options
  • Star 0 You must be signed in to star a gist
  • Fork 0 You must be signed in to fork a gist
  • Save finchd/dcc37801129306bb0043a4faa46f73b5 to your computer and use it in GitHub Desktop.
Save finchd/dcc37801129306bb0043a4faa46f73b5 to your computer and use it in GitHub Desktop.
Download ZIP
SeaGl 2018 - The impact of lossy networks on TCP performance
Raw
seagl2018-tcp-perf.md

The impact of lossy networks on TCP performance

Jeff Silverman (jeffsilverm at- gmail.com)

questions to answer

How well does TCP cope with loss?

How does TCP on IPv4 vs. IPv6 behave?

tools

Using tc to create latency & loss.

post-processing in jupyter, pandas, numpy, scipy, matplotlib (graphs)

Findings

  • Quite resilient even in 3/5 60% packet loss

  • IPv4/IPv6

  • no real difference on LAN
  • WAN IPv4 worked better, but CenturyLink IPv6 was in encapsulation
  • 80% loss & 500ms delay - still worked
  • failures are not deterministic - delay not constinent with what tc set
  • some uncaught failures were actually DNS

Data

504 runs

ipv4 failed 9 times (3.6%) IPv6 failed 19 times (7.5%) failures non-deterministic - not correlated to conditions

Size wasn't a predictable indicator of failures

  • global tcp retransmits from netstat -s from all connections
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment