Re: Comments on the centroids paper

[Simon E Spero <[e-mail]>]

Chris
  Joann Ordille's work on Nomenklator contained a lot of measurements on the
effect of meta-data caching, as well as ordinary caching. Meta-data caching, 
which records whether or not records of a particular type are held in a certain
subtree can be quite productive.

Full record caching can also be quite productive, but only on a local level; 
imagine the case where somebody consistently uses user-friendly naming for
electronic mail. If no caching is done, then each message will require a 
directory lookup, which is still non-trivial. If even a small cache is used,
then the lookup becomes very cheap. 

Multi-level caching on large data-sets has been shown to yield rapidly 
diminishing returns. Peter Honeyman's "Your cache ain't nothing but trash",
in the Winter '90 usenix (same one as Archie), has some results for 
experiments using the AFS filesystem.

Where MLC becomes really useful is when the cache is placed at a near side of 
a high-latency or low capacity link. The savings from even a small number of 
hits may more than cover the cost of the cache.

Simon