Extended Data Fig. 2: Runtime simulation of FRACTAL.

a, Conceptual diagram of the runtime simulation of FRACTAL. Under a distributed computing environment with a fixed number of available nodes d, each FRACTAL job of input sequence size n starts if there is one or more available free computing node; otherwise, it is stalled until a node is released from one of the ongoing jobs. Each FRACTAL job process is modeled to occupy one computing node for a runtime f(n) and produce two new child jobs for the next job cycles each with an input sequence size of ⌊n/2⌋. When the input sequence size is under a certain threshold (n ≤ t), the job terminates after occupying one node for a runtime f(n) assuming the terminal lineage reconstruction process. b, Two implementation models of FRACTAL. In FRACTAL, each job cycle contains three steps that require high computing loads: sequence subsampling from the entire input sequences, phylogenetic placement, and sorting of sequences mapped on each of the sample lineage clades for the next job cycles. In model A, none of the three steps is parallelized where a runtime of each cycle is f(n) for the input sequence size of n. In model B, all of the three steps are perfectly parallelized where a runtime of each cycle is linearly reduced by the number of available computing nodes. The denominator of the formula represents the number of available computing nodes, assuming that the smaller the input sequences for a FRACTAL iteration cycle, the higher the likelihood of computing nodes being occupied by the other job processes at the same period. c, Linear regression f(n) using runtime log data of independent FRACTAL iteration cycles any of whose major three steps was not parallelized for the lineage reconstruction of the 235 million sequences.