Project

General

Profile

Bug #3660

tcp_slow_start_* tunables should allow increasing the initial congestion window

Added by Theo Schlossnagle over 7 years ago. Updated over 7 years ago.

Status:
Resolved
Priority:
Normal
Category:
networking
Start date:
2013-03-27
Due date:
% Done:

100%

Estimated time:
2.00 h
Difficulty:
Bite-size
Tags:
needs-triage
Gerrit CR:

Description

The current implementation of initial congestion window calculation uses RFC 3390 correctly, but in the event the operator wants to violate the RFC for a fast TCP start (initial congestion window of 10) re: http://research.google.com/pubs/pub36640.html it cannot be done except for programmatically via setsockopt.

Setting the two tunables should allow the operator to increase the initial congestion window up within reason. The implementation of TCP_INIT_CWND via setsockopt has a "sane" maximum of 16, that should be adopted.

The tunables: tcp_slow_start_initial and tcp_slow_start_after_idle should change in behaviour. They should default to 0, which takes on the new meaning of strict adherence to RFC 3390 and if there are set to a number larger than 0, then the initial congestion window (or restart after idle) is set directly to that specified value. The allowable ranges of the two tunables should remain unchanged (sans allowing 0 for "default" behavior).

History

#1

Updated by Theo Schlossnagle over 7 years ago

  • % Done changed from 0 to 90

RTI'd webrev

#2

Updated by Theo Schlossnagle over 7 years ago

  • % Done changed from 90 to 100
#3

Updated by Theo Schlossnagle over 7 years ago

  • Status changed from New to Resolved

Also available in: Atom PDF