Performance prediction of scientific applications on heterogeneous HPC architectures

At universities are often several High Performance Computing (HPC)-systems provided.
Either their own or through partner universities. Thus, there is an interest to find out on
which system an application can be executed best. These platforms can consist of different
generations of a vendor or a completely different CPU architecture.
In this thesis, six different Intel processor generations are compared and the theoretical
performance is calculated. Afterwards, many different benchmarks are compared with the
theoretical computing performance to see if the measured computing performance has
achieved the same level as the theoretical computing performance.
Unfortunately, this did not work out very well. There was a strong increase in the
newer generations in various benchmarks, but not to the level that was expected with the
theoretical computing performance. This is probably due to the fact that the theoretical
computing performance was determined with the CPU’s base frequency. However, when
using vector extensions, such as SSE, AVX or AVX512, the CPU clock frequency are lower
than with calculations without vector expansion.
... mehr
In performance measurements with only one CPU core, a precise result could be achieved.
By a parallel measurement of the CPU frequency, a much more realistic theoretical com-
puting performance could be determined. When comparing the compute performance
determined in this way with the measurement results, a deviation of partly less than 10%
could be achieved. It allows a reasonable prediction for a different platform based on a
known CPU frequency.
Furthermore, the influence of compilers were inspected on the basis of the Intel Compiler
and the GNU Compiler. It turned out that the O3 flag should always be used for a productive
application. The execution time could be reduced by over 50%. By using AVX or AVX512,
it was also found out that the GNU compiler produced better compilations for modern Ice
Lake CPUs, which achieved a shorter execution time compared to the compilations of the
Intel compiler.