tree 453b5fa31565f50dbc29d36560fd438e6fb179ea
parent f6facb67b7444450bdcb2c388b3af7c092fb3210
author Junho Choi <junho@cloudflare.com> 1611906724 -0800
committer Alessandro Ghedini <alessandro@ghedini.me> 1612876676 +0000

cc: fix hystart++ bugs and excessive cwnd growth

Current hystart++ implementation has a few issues:
- hystart.on_packet_acked() returns a new cwnd/ssthresh and
  decide if it need to enter LSS. However during slow start
  and congestion avoidance, it will confuse CC's ABC
  based cwnd calculation, resulting higher cwnd growth
  with HyStart++ during slow start and LSS episode.
  (CUBIC and Reno)
- hystart LSS cwnd calculation is not ABC based,
  can result in higher cwnd growth then expected (no
  2 x ABC_L x MSS upper limit)
- Reno slow start: mixed use of bytes_acked in slow start
  and congestion avoidance. It will result in miscalcuation
  of cwnd.

Changes:
- hystart: remove on_packet_acked() and add try_enter_lss()
  for checking if CC can enter LSS. cwnd/ssthresh will be
  changed from the caller. Also added in_lss() for a
  shortcut of checking the condition whether currently in
  LSS. lss_cwnd() will provide a new cwnd in LSS.
- Separate bytes_acked for slow start (bytes_acked_sl)
  and congestion avoidance (bytes_acked_ca for Reno).
  CUBIC uses cwnd_inc)
- CUBIC: fix a bug where cwnd can grow more than 1 MSS
  in congestion avoidance. cwnd_inc will be reset
  after increasing 1 MSS for preventing cwnd keep increasing
  when cwnd_inc > 2 MSS.
  (This has a future work - need to follow up recent CUBIC
   RFC update which address this problem better)
- Follows HyStart++ draft 03, especially around L based
  cwnd growth limit:
https://tools.ietf.org/html/draft-balasubramanian-tcpm-hystartplusplus-03#section-4.2