THE FOLD / BOSS / THE GAUNTLET / THE RTT BIAS
THE RTT BIAS
a rule written in one frame, judged in another
1 WHAT IT IS · WHAT IT DOES · FACT OR FICTION
Two flows run the identical algorithm on the identical link and do not get the identical share. The one with the shorter round trip increases its window more often, and the rule that is fair in steps is not fair in time.
LIT verified live. Loss rate 0.01, one flow at 10 ms and one at 100 ms — an RTT ratio of 10. The closed form predicts a throughput ratio of 10.00; simulating both flows on a shared link of capacity 200 over 200,000 ticks gives 9.91. The near flow takes 90.83% of the link and the far flow 9.17%, with neither breaking a rule.
LIT verified live. Loss rate 0.01, one flow at 10 ms and one at 100 ms — an RTT ratio of 10. The closed form predicts a throughput ratio of 10.00; simulating both flows on a shared link of capacity 200 over 200,000 ticks gives 9.91. The near flow takes 90.83% of the link and the far flow 9.17%, with neither breaking a rule.
2 HOW IT WAS WEAVED · AI + HUMAN
RTT unfairness follows from the square-root throughput law — rate is inversely proportional to RTT — and it is why a distant user loses to a nearby one on a congested path.
AVAN (AI) ran the simulation and the closed form so that neither has to be taken on faith. 9.91 against 10.00 is the model and the mechanism agreeing, which is worth more than either printed on its own.
AVAN (AI) ran the simulation and the closed form so that neither has to be taken on faith. 9.91 against 10.00 is the model and the mechanism agreeing, which is worth more than either printed on its own.
3 ONE DIMENSION
Same rule, same link, different distance.
4 TWO DIMENSIONS · INTERACTIVE
Move the far flow further away.
5 THREE DIMENSIONS + AVAN’S INVERSE
The green forward object: the same rule, two distances.
AVAN’s addition (the inverse-companion): the forward reading is that RTT bias is an unfairness in the algorithm. The inverse is that the algorithm has no concept of time and cannot be unfair in it. AIMD is defined per round trip, and a round trip is the only clock it has; measured in its own units the two flows are treated identically. Read backwards, the bias is what appears when a rule written in one frame is judged in another, and the near flow is not cheating — it is simply being asked more often.
LIT at a loss rate of 0.01 with one flow at 10 ms and one at 100 ms - an RTT ratio of 10 - the closed form predicts a throughput ratio of 10.00 while simulating both on a shared link of capacity 200 over 200,000 ticks gives 9.91, with the near flow taking 90.83% of the link and the far flow 9.17%, neither having broken a rule
FIG RTT unfairness follows from the square-root throughput law - rate is inversely proportional to RTT - and it is why a distant user loses to a nearby one on a congested path. AVAN ran the simulation AND the closed form so neither has to be taken on faith. 9.91 against 10.00 is the model and the mechanism agreeing, which is worth more than either printed alone.
FIG RTT unfairness follows from the square-root throughput law - rate is inversely proportional to RTT - and it is why a distant user loses to a nearby one on a congested path. AVAN ran the simulation AND the closed form so neither has to be taken on faith. 9.91 against 10.00 is the model and the mechanism agreeing, which is worth more than either printed alone.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of THE GAUNTLET · David Lee Wise (ROOT0), with AVAN