Hello Jed,
Welcome to the forums! :qiime2:
Before I dive in, I want to suggest that specific, short questions usually get answered faster, because more people read them and it's easier to answer one question than many.
I really liked this part of your analysis:
So, I tried many parameters by changing the value of --p-trunc-len-f and --p-trunc-len-r. The results are as follows:
...
And how to ensure the best cut-off position?
Your method is perfect: try a parameter sweep and pick the setting that works best!
DADA2 first truncates the reads, then joins them, then denoises the full run, so the choice of truncating settings is very powerful, as you have seen.
Second, I want to know the difference between handling all samples at once and only one sample at a time, and which method should I use in different situations.
It is recommended to run DADA2 per sequencing run, through results should be stable as long as the exact same settings are used. See DADA2 workflow for big data and this answer by Justine.
This would be a little surprising if all the same settings were used! If I were investigating this, I would disable chimera checking 'none' and see if the results are consistent before this step.
Let us know what you try next!