[Codel] [Bloat] slow start improvement

Jonathan Morton chromatix99 at gmail.com
Fri Dec 29 01:38:06 EST 2023


> On 28 Dec, 2023, at 12:17 pm, Sebastian Moeller via Bloat <bloat at lists.bufferbloat.net> wrote:
> 
> The inherent idea seems to be if one would know the available capacity one could 'jump' the cwnd immediately to that window... (ignoring the fact the rwnd typically takes a while to increase accordingly*). 

Yes, I've just got to the bit about selectively ignoring rwnd - that's a straight violation of TCP.  There may be scope for optimising congestion control in various ways, but rwnd is a fundamental part of the protocol that predates congestion control itself; it implements TCP's original function of "flow control".  Sending data outside the rwnd invites the receiver invoking RST, or even firewall action, which I can guarantee will have a material impact on flow completion time!

Slow-start already increases cwnd to match the BDP in at most 20 RTTs, and that's the extreme condition, starting from an IW of 1 segment and ramping up to the maximum possible window of 2^30 bytes (assuming an MSS of at least 1KB, which is usual).  The more recent standard of having IW=10 already shortens that by 3-4 RTTs.  It's an exponential process, so even quite large changes in available bandwidth don't affect the convergence time very much.  TCP's adaptation to changes in the BDP after slow-start is considerably slower, even with CUBIC.

I also note a lack of appreciation as to how HyStart (and HyStart++) works.  Their delay-sensitive criterion is triggered not when the cwnd exceeds the BDP, but at an earlier point when the packet bursts (issued at double the natural ack-clocked rate) cause a meaningful amount of temporary queue delay.  This queuing is normally drained almost immediately after it occurs, *precisely because* the cwnd has not yet reached the true path BDP.  This allows slow-start to transition to congestion-avoidance smoothly, without a multiplicative-decrease episode.  HyStart++ adds a further phase of exponential growth on a more cautious schedule, but with essentially the same principle in mind.

The irony is that they rely on precisely the same phenomenon of short-term queuing, but observe it in the form of the limited delivery rate of a burst, rather than an increase in delay on the later packets of the burst.

 - Jonathan Morton


More information about the Codel mailing list