[Codel] [Bloat] slow start: small chunks can talk

Sebastian Moeller moeller0 at gmx.de
Mon Aug 7 06:18:14 EDT 2023


Hi Roland.

> On Aug 7, 2023, at 10:48, Bless, Roland (TM) via Bloat <bloat at lists.bufferbloat.net> wrote:
> 
> Hi Dave,
> 
> On 01.08.23 at 00:36 Dave Taht via Bloat wrote:
>> Promising approach:
>> https://ieeexplore.ieee.org/document/10188775
> 
> It's a pity that neither the authors nor the reviewers were aware of earlier related work that you also sent here to the list:
> 
> L. Guo and J. Y. B. Lee, "TCP-FLASH - A Fast Reacting TCP for Modern Networks," in IEEE Access, vol. 9, pp. 68861-68879, 2021, doi: 10.1109/ACCESS.2021.3077612.

	Interesting link. This still uses essentially a packet-pair measuring method (send packet back2back, look at time difference in resulting ACK packets to estimate bottleneck bandwidth in the forward direction). These are still not robust or reliable... (parallel path, reordering, or simply congestion on the reverse path, ACK filtering, ACK compression by GRO on the receiver end, ...). Now, well possible that packet-pair data, while not perfect, might still be good enough for the intended purpose... And even for a traditional slow-start having an educated guess when to leave the exponential growth phase could be helpful...

	More over, let's assume any of these (let's short circuit the probing phase) will actually be deployed at scale; how will the network cope with the much more aggressive ramp-up of such flows (essentially initial-window as one batch the switch to estimated capacity once the bottleneck rate is estimated). It is fun to see a single/few of such flows do the right thing, but what about having the majority of flows use such methods? (Which will likely invalidate the bottleneck rate estimate quickly).

	I guess I might be too cautious here, but I personally see the exponential ramp-up in more traditional slow-start already as pretty aggressive and yet more adaptive to changing capacity than jumping to the estimated capacity essentially cold (I have similar hesitation with the careful resume internet draft).

Regards
	Sebastian

P.S.: "To tackle this problem, FLASH is designed to suppress
the AWnd constraint unless it is zero. This opportunistic
transmission technique, first proposed by Liu and Lee [6],
allows FLASH to send data beyond the AWnd limit up to
CWnd amount of packets inflight. "

This is quite an euphemism "opportunistic transmission technique" for simply ignoring parts of the protocol...

P.P.S.: Dave's link is behind a pay-wall, so I have no idea about that paper beyond the abstract... but if they also relay on packet-pair bandwidth estimates and an faster-than-slow-start ramp-up, I would expect similar issues.



> 
> The fast launch phase and especially the CWnd-Compensated Bandwidth Estimator (CCBE) are quite similar and a comparison of both approaches
> would have been interesting...
> 
> Regards,
> Roland
> 
> _______________________________________________
> Bloat mailing list
> Bloat at lists.bufferbloat.net
> https://lists.bufferbloat.net/listinfo/bloat



More information about the Codel mailing list