THE FOLD / BOSS / THE GATEKEEPER / THE DELAY BASED
THE DELAY BASED
it does not lose a race, it declines one
1 WHAT IT IS · WHAT IT DOES · FACT OR FICTION
A delay-based flow watches the queue and yields when it starts to grow. A loss-based flow keeps pushing until the buffer overflows. Put them on the same link and the polite one is punished for its manners.
LIT verified live. Capacity 100, buffer 200, 100,000 ticks. The delay-based flow takes 10.24% of the link and the loss-based flow 89.76% — a ratio of 8.77. The delay-based flow is doing the better thing: it keeps the queue short, which is what everybody’s latency depends on. It loses anyway, and it loses because it is the only one looking.
LIT verified live. Capacity 100, buffer 200, 100,000 ticks. The delay-based flow takes 10.24% of the link and the loss-based flow 89.76% — a ratio of 8.77. The delay-based flow is doing the better thing: it keeps the queue short, which is what everybody’s latency depends on. It loses anyway, and it loses because it is the only one looking.
2 HOW IT WAS WEAVED · AI + HUMAN
This is why TCP Vegas never displaced Reno despite being better on an empty network, and the problem every delay-based design since — including BBR — has had to answer.
AVAN (AI) ran both flows on one link rather than measuring each alone, because alone the delay-based flow wins on every metric that matters. The result only appears in competition, and competition is the only condition that has ever existed.
AVAN (AI) ran both flows on one link rather than measuring each alone, because alone the delay-based flow wins on every metric that matters. The result only appears in competition, and competition is the only condition that has ever existed.
3 ONE DIMENSION
Two strategies, one link.
4 TWO DIMENSIONS · INTERACTIVE
Change the buffer and watch who wins.
5 THREE DIMENSIONS + AVAN’S INVERSE
The green forward object: the one that yields.
AVAN’s addition (the inverse-companion): the forward reading is that delay-based control loses to loss-based control. The inverse is that it does not lose a race, it declines one. Yielding when the queue grows is not a weaker strategy for taking bandwidth — it is a refusal to take bandwidth by the only means available, which is to make everyone else wait. Read backwards, the 8.77 is not a measure of which algorithm is better; it is the price of the externality, and it is paid entirely by whoever decided not to impose one.
LIT with a capacity of 100, a buffer of 200 and 100,000 ticks the delay-based flow takes 10.24% of the link and the loss-based flow 89.76%, a ratio of 8.77, and the delay-based flow is doing the better thing - it keeps the queue short, which is what everybody's latency depends on - yet it loses anyway, and it loses because it is the only one looking
FIG This is why TCP Vegas never displaced Reno despite being better on an empty network, and the problem every delay-based design since - including BBR - has had to answer. AVAN ran both flows on one link rather than measuring each alone, because alone the delay-based flow wins on every metric that matters. The result only appears in competition, and competition is the only condition that has ever existed.
FIG This is why TCP Vegas never displaced Reno despite being better on an empty network, and the problem every delay-based design since - including BBR - has had to answer. AVAN ran both flows on one link rather than measuring each alone, because alone the delay-based flow wins on every metric that matters. The result only appears in competition, and competition is the only condition that has ever existed.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of THE GATEKEEPER · David Lee Wise (ROOT0), with AVAN