[CCWG] ABE and CUBIC parameters beyond beta

[Kazuho Oku <[e-mail]>]

Hi CCWG,

RFC 8511 discusses ABE mostly in terms of changing the multiplicative
decrease factor for ECN, e.g. using beta_ecn = 0.85 for CUBIC instead of
beta_loss = 0.7.

If we make only that change to CUBIC, though, I think two other parameters
may also need to change.

First, consider Fast Convergence. Suppose an ECN congestion event occurs
below the previous W_max and we change only the multiplicative decrease to
beta_ecn = 0.85, while leaving the usual Fast Convergence factor at 0.85.

Then:

cwnd_epoch = 0.85 * cwnd_prior
W_max      = 0.85 * cwnd_prior

and therefore:

K = cbrt((W_max - cwnd_epoch) / C) = 0

So the entire concave part of the CUBIC curve disappears and the new epoch
starts at the inflection point. It seems that the Fast Convergence factor
for an ECN-triggered event should instead be:

(1 + beta_ecn) / 2 = 0.925

Second, there is the Reno-friendly alpha. My understanding is that this is
chosen to match the recovery behavior of Reno toward cwnd_prior. If Reno
itself uses ABE (e.g. beta_ecn = 0.8) while CUBIC uses beta_ecn = 0.85,
retaining the usual CUBIC alpha = 0.529... would no longer preserve that
relationship, so alpha would need to be re-derived for the ECN case as well.

In the case where Reno uses 0.8 and CUBIC using 0.85, it should be (1 +
0.8)(1 - 0.85) / (1 - 0.8)(1 + 0.85) = 0.729.

Thoughts?

-- 
Kazuho Oku