[SM] It is doe because:
a) TCP needs some capacity estimate
b) preferably quickly
c) in a way gentler than what was used before the congestion collapse.
Right, but we're moving away from capacity shortages to a focus on
better latencies. The speed of distributed compute (or speed of
causality) is now mostly latency constrained.
Also, it's impossible per Jaffe & others for a TCP link to figure out
the on-demand capacity so trying to get one via a "broken control loop"
seems futile. I believe control theory states control loops need to be
an order greater than what they're trying to control. I don't think an
app or transport layer can do more than make educated guesses at for its
control loop. Using a rating might help with that but for sure it's not
accurate in space-time samples. (Note: many APs are rated 60+ Watts.
What's the actual? Has to be sampled and that's only a sample. This
leads to poor PoE designs - but I digress.)
Let's assume the transport layer should be designed to optimize the
speed of causality. This also seems impossible because the e2e jitter is
worse with respect to end host discovery so there seems no way to adapt
from end host only.
If it's true that the end can only guess, maybe the solution domain
comes from incorporating network measurements via telemetry with the ECN
or equivalent? And an app can signal to the network elements to capture
the e2e telemetry. I think this all has to happen within a few RTTs if
the transport or host app is going to adjust.
Bob
_______________________________________________
Bloat mailing list
Bloat@lists.bufferbloat.net
https://lists.bufferbloat.net/listinfo/bloat
