Search

Auto-tuning for complex and nontrivial parallel systems is usually time-consuming because of empirical evaluation of huge amount of combinations of parameter values of an initial parallel program in a target execution environment. This paper proposes the improvement of the auto-tuning method using statistical modeling and neural network algorithms that allow to reduce significantly the space of possible combinations of parameters values to analyse. The resulting optimization is illustrated by an example of tuning of parallel sorting program, that combines several sorting methods, by means of automatic training of a neural network model on results of “traditional” tuning cycles with subsequent replacement of some auto-tuner calls with an evaluation from the statistical model. Problems in programming 2018; 2-3: 046-053

Proposed an approach to automatic transformation of the legacy code on Fortran for execution on cloud computing platforms. Proposed architecture of the system based on web-services choreography, which allows infinite scalability of the system and reduces overhead on message passing. Investigated an approach of the approach on example of the program from the quantum chemistry field.