I found the kernel thread where the original author of the FS-Cache patches, David Howell, makes it clear that on a quiet network with a quite fast server metadata will take longer from the cache. However, at my work we have very busy large NFS servers connected over the building network which is very busy. When you try to read a large file repeatedly in the middle of the day the traditional NFS caching just doesn't work if the time between reads is more than about 5 minutes. I've resorted to manually copying my datasets to /usr/tmp on the local disk and seen huge performance improvements. (this has other serious issues, like getting confused about which copy you just modified and migrating any changes back to the official NFS copy.) I know this feature makes sense for me and others in similar environments. The problem of course is: (1) it will be years before it makes it into RHEL and (2) it won't be turned on by default, (3) my system admins are weary to trying anything kernel-related that's not stock RHEL. However, if I can show them an order of magnitude improvement in speed, which I think this will do, they might think twice.
