I used cd-hit-est to remove redundancy. But I found when the threshold was set to 0.99, some sequences with 98% identity were also be clustered in some groups. This seemed not due to short word size (n=8), because longer word size (n=9) gave the same result.