We now move to a case where the block group has data in it to begin. This set of experiments varies the fullness of the block group and runs the same small-file creation benchmark (focusing on the single-PID case). Figure 11 plots the results.
From the figure, we can see that by the time a block group is 50% full, range writes improve performance over classic writes by roughly 20%. This improvement stays roughly constant as the block group fills, even as the average write time of both approaches increases. We can also see the effect of fullness on range writes: with fewer options (as the block group fills), it is roughly 70% slower than it was with an empty block group.