Wednesday, June 28, 2017
Argument that low N is best in difficulty algorithms and why dynamic averaging window is not a benefit
I can't recommend a switch from v1 to v2 (static N to dynamic N). The smoothness gained by the higher N is not much: surprisingly, the std dev of solve times increases only 5% from N=30 to N=8. The std dev of D goes from 0.18xD to about 0.45xD for N=30 verses N=8. For N=8 this means 97.5% are less than D=1.96x0.45=2 times more than they should be) . Long story short (due to Poisson median being 0.693 of average): going from N=30 to N=8 means only a 47% increase in 4xT solvetimes. The dynamic window does not capture this benefit: those > 4xT solvetimes are exactly the statistically unlikely times that will trigger the dynamic window back to a lower N, canceling the primary benefit of it rising back up to large N. It looks a lot smoother and nicer _**most**_ of the time when hash rate is constant, but the painful small N events are not reduced.