I found it completely depends on the expected yield of small RNA in your samples. I have had seq data from samples which had large RNA degradation issues, so we saw lots of rRNA in the preps and total miRNA levels < 5% and thus distinguishing noise from signal was more difficult. But also had samples where the yield was much higher (low background rRNA); I think these were exosomal preps. In both cases the reads were between 5-20M and still useful, so it's probably a safe bet to use 10M if your confident in your sample quality.