In section 5 in Atkin sieve paper. I don't really understand his analysis of algorithm complexity.
He started with defining $\displaystyle W=12\left(\prod_{3<p\le \sqrt{\log N}}p\right)$ which is roughly $\exp(\sqrt{\log N})$, then he selected $B$ close to $W\sqrt{N}$ (How can we select this step if $WB$ is roughly $N$?, I am not sure about this but the algorithm find primes between $LW$ and $LW+WB$ and the argument he suggested that the number of operations overall $\delta$ is $\mathcal O\left(\dfrac{WB}{\log\log N}\right)$)