TR3801 gives an excellent introduction (Paul's TR). We typically let pcs spin up for a day, then run the stats command to get output (and perfstat too) then turn it off. So pick a time when you have representatve workload and let it run for a day then run the output for the counters. I would make separate runs for each of the modes (default metadata/data, metadata only, lopri) if you need that but for most cases the default works...but I could see testing the other two modes if you don't see that the ec counters are getting good hit rates or if you know that metadata is better or that data written is read back after.