Execution performance of a parallel process in general-purpose 
NUMA(Non-Uniform Memory Acess) systems
is greatly affected by how resources are allocated to it through its lifetime.
Concurrently running multiple parallel processes will exhaust physical memory.
We propose two resource management mechanisms.
One is a scheduling policy that reflects resource consumption states.
A process is scheduled to clusters where it has physical pages.
The other is a memory-replacement strategy based on page classification
under distributed shared memory system.
Shared copy pages of currently not running processes are first victimized.
The performances of the two mechanisms are evaluated by a probabilistic simulation.
It allows to simulate a variety of process sets and finite resources are manipulated
with concrete management methods.
The results show the superiority of our resource management mechanisms.