Re: [tcpm] [iccrg] Cwnd growth after fast convergence in CUBIC

[Neal Cardwell <[e-mail]>]

On Fri, Nov 6, 2020 at 3:28 AM Lars Eggert <[e-mail]> wrote:
> That said, the intent was to document what the Linux code was doing
> at the time. Is there a difference between the RFC and what Linux
> has been doing or is doing now? If yes, it might be time to revise
> the RFC.

Yes, it seems that this is a case where the RFC differs from the Linux
TCP CUBIC code.

The RFC says:
      K = cubic_root(W_max*(1-beta_cubic)/C) (Eq. 2)

which is equivalent to:
      K = cubic_root((W_max - W_max*beta_cubic)/C).  [B]

The Linux code does something different: it computes K once per epoch
immediately after the fast recovery or ECN cwnd reduction phase
completes, using:

      K = cubic_root((W_max - cwnd)/C).  [C]

See the computation of bic_K here:
  https://github.com/torvalds/linux/blob/master/net/ipv4/tcp_cubic.c#L254

For the non-fast-convergence case, the result of [B] and [C] are
equivalent, but for the fast convergence case they are not. So it
looks like in the case of fast convergence the RFC formula produces a
result that is incorrect, and does not match the Linux TCP CUBIC
behavior.

AFAICT,  changing Eq 2 to the following might fix the RFC and make it
match the Linux TCP CUBIC behavior:

      K = cubic_root(W_max - W_last_max*beta_cubic)/C) (Eq. 2)

best,
neal