Preventing parallel processes from unexpected inefficiencies is a
major concern for constructing multiple user/multiple job environment
in NUMA systems.  Systems can achieve higher performance by using
scheduling policies which reflects resource consumption states.  For
a general environment, which must support concurrent execution of
multiple processes, a way is needed to keep systems' effectiveness
when physical memories are full.  In NUMA systems, memory pages can be
classified by access frequencies and required costs for accesses when
target pages are not found locally.
Selecting victim pages according to the classification enhances system
performance.  We built a probabilistic model with a concrete memory
management scheme and differentiated memory access costs, and simulated
processes sets with given access frequencies.  The paper describes an
evaluation of scheduling policies using resource informations for each
process and of page replacement policies based on page coloring under
the model.