A Simple Linear Model of Demand Paging Performance
Predicting the performance of a proposed automatically
managed multilevel memory system requires 
a model of the patterns by which programs refer to the
information stored in the memory.  Some recent 
experimental measurements on the Multics virtual memory
suggest that, for rough approximations, a remarkably 
simple program reference model will suffice.  The simple
model combines the effect of the information 
reference pattern with the effect of the automatic management
algorithm to produce a single, composite 
statement: the mean number of memory references between
paging exceptions increases linearly with the 
size of the paging memory.  The resulting model is easy
to manipulate, and is applicable to such diverse 
problems as choosing an optimum size for a paging memory,
arranging for reproducible memory usage charges, 
and estimating the amount of core memory sharing.
CACM April, 1974
Saltzer, J. H.
