There is nothing wrong with having different 16S rRNA lengths, assuming you are predicting from a metagenomic sample. That simply means that some of your contigs contain incomplete rRNA sequences - no program can predict a complete rRNA if it isn't there. As long as you have the latest version,
barrnap is a safe choice.
If you want to convince yourself that
barrnap predictions are solid, another program to try is RNAmmer. It has similar speed and sensitivity to
barrnap, though the latter is more likely to be up to date. Lastly, if you want to go with heavy artillery, your choice would be SSU-ALIGN. It will have the highest sensitivity because it uses covariance models for scoring, but that's also why it will be slowest by couple orders of magnitude. If you have say 100+ Mb of sequence data, it will take many hours (even a whole day) to complete the prediction, and that's assuming the job is spread over many CPUs. You can give it a try if your dataset is relatively small or if you have time to spare, but keep in mind that it won't give you longer sequences if they are not there. The most you could expect is that it may detect some borderline matches, but usually the result will end up being the same as with