2025 , Volume 30, № 6, p.109-121

The article addresses parallel variants of the atax computing kernel obtained using polyhedral optimization. The performance issues of parallel atax variants are discussed in the context of profiling MPI applications using the oprofile tool in Linux. The goal of the study is to identify the inefficiency of using the processor cache when executing parallel programs obtained as a result of polyhedral optimization in a distributed memory environment. The performance of parallel programs is compared, the programs were implemented using the modern parallelizing translator Pluto 0.11.4 and the ilpy translator, which differ in objective functions when finding the optimal alternative. The correlation of the calculated profiling metrics with the execution time of the parallel program is considered. Correlation analysis of the calculated characteristics of cache use in a distributed environment have showed that as the number of processes increases so, the effects of the second- level cache become weaker, but remain for the third-level cache. The effects of the first-level cache are not observed on the considered microarchitecture. The intensity of information exchange affects not only the delays caused by network interaction, but also the delays associated with Linux virtual memory (memory is allocated at the time of use). The latter, in turn, affects the array access patterns, which implies a different operation of caches: atax variants with copied input data show a smaller number of cache misses. The scheme of information exchange, along with the intensity, has a direct effect on the memory access patterns.