Metadata operations power lots of day-to-day activities on your typical storage array: make/rename/remove directories, timestamps, file permissions, locks, etc.
In this January 2009 dissertation by Christoph Biardzki, you'll find all sorts of references to NetApp, WAFL, GX, FAS3050, Consistency Points, NVRAM, RAID-DP, FlexShare, etc. as part of their HLRB 2 Supercomputer:
Analyzing Metadata Performance in Distributed File Systems
There's some very interesting observations in the above paper ... for example:
"Because there are no recognizable patterns of WAFL consistency points, it can be concluded that performance is not limited by saturation of the NFS server. When using Lustre and the underlying ext3 file system it is, unfortunately, not possible to reason in the same manner: Flushing the ext3 journal seems to briefly stall all operations, independent of the load on the server. In the case of WAFL processing, a consistency point uses up resources but does not interrupt processing."
The other important point is the need for retooling mainstream applications, introducing more parallelism in the years to come -- especially considering future deployments of Parallel NFS (pNFS) in the storage stack.
The future of file storage is parallel. Even your metadata.
