ndt-server icon indicating copy to clipboard operation
ndt-server copied to clipboard

Published 20 hours ago • verified publisher icon m-lab

Implement BBR holdback experiment to help diagnose unexpected ndt7 results

Open mattmathis opened this issue 5 years ago • 4 comments
trafficstars

For some small fraction (1%?) of the ndt7 tests use CUBIC instead of BBR, to help identify the root causes of unexpected low performance.

mattmathis avatar Jul 27 '20 15:07 mattmathis

See also #314

mattmathis avatar Jul 27 '20 15:07 mattmathis

Good candidate for future OKR for long term data calibration and A/B testing

laiyi-ohlsen avatar Aug 10 '20 16:08 laiyi-ohlsen

I encountered another WiFi issue that might be due to an interaction with BBR. The only robust experiment to test this would be to see if CUBIC gets significantly better performance under similar scenarios.

mattmathis avatar Jan 14 '22 22:01 mattmathis

Tentative requirements (combining #351)

  • Enable the client to force CC from a list of valid names
  • If no forced CC from client, Use weighted random selection.
  • parameters for random selection should be set on then command line, such that they can be set in the container config.

mattmathis avatar Mar 27 '22 04:03 mattmathis