Four different benchmarking suites for testing the performance of MPI routines have been considered on an SGI Origin2000.
Special properties of these benchmarking suites which are mostly hidden to the user turned out as being of considerable influence
on the benchmarking results for ccNUMA systems such as number and location of buffers, warm-up of the cache before running
the benchmark, or procedure of measuring the time. In addition, we consider interpretation of results and their approximation
by piecewise linear curves.
Key Words message passing - performance analysis - benchmarking - MPI - ccNUMA architectures