Handling page faults is the major overhead of page-based software Distributed Shared Memory (sDSM) system. Based on profiled page faults counts and types, a performance model has been developed for two constrasting sDSM, home-less(CLOMP) and home-based sDSM (SCLIB).
To utilize the model, cost of different page faults (write/fetch) need to be obtained on a given hardware platform. Therefore, a micro benchmark to measure these cost has been developed.
The total costs are achieved by measuring elapsed time for fetch/write faults from shared memory and substrate the reference time with same operation on local memory. The number of different types page fault counts are recorded as well for getting per page fault cost purpose.
The following people can be contacted: