Fig. 2

Running time and accuracy of the tools on Ecoli and HMP datasets. A and E show the total wall-clock time to run the tools for 25-125 randomly selected samples (Simka did not exit after > 48 hours). B and F show the distributions of the percentage errors when estimating cosine and Bray-Curtis, respectively, for 100 samples (ground truth was computed using Simka). C and D show estimated vs. true cosine values for the two datasets (100 samples), and G and H show the same plots for the Bray-Curtis dissimilarity. The plots show that frac-kmc is about 40-50% slower, but is comparatively more accurate in estimating the metrics