Performance models of processors in heterogeneous platforms

Optimal distribution of computations over the processors of a heterogeneous network can be only achieved based on a performance model of the processors which is able to accurately predict the contribution of computations into the total execution time of any particular parallel or distributed algorithm on the network. The project proposes and investigates realistic performance models of processors in heterogeneous networks, taking into account all the features that may have an impact on the performance, such as architectural and platform heterogeneity of the processors, memory heterogeneity, different levels of integration into the network. The project looks at different aspects of the models including accuracy and adequacy, efficiency of their use, building, maintenance and runtime update.

  • PMM: Performance Model Manager
  • FuPerMod: a software tool for optimal data partitioning of parallel scientific applications on dedicated heterogeneous HPC platforms