Has anyone thought of adding a counter for the number of NVRAM flushes for a given rolling period?
My understanding is that one of the issues in achieving excellent performance on writes is avoiding, or dealing with, back-to-back consistency points.
It strikes me that a simple metric based on the number of NVRAM flushes-to-disk over the last 60-seconds, updated every 5 seconds say, would give people a handle on whether there's a bottleneck happening there.
It might be even more interesting to keep track of in the clustered-filer environment.