Abstract
|
As we push toward the Exascale time, the quantity of nodes in the high-performance computing and the quantity of cores in every node are quickly expanding. Thus, the physical topology of present day multicore high-performance computing frameworks is winding up increasingly complex both at the inward and inner levels regarding the communication inside and among the nodes. This converts into an expansion in the level of heterogeneity, which results in different purposes of correspondence execution inside the correspondence node. Specifically, between node correspondence is commonly slower than interior node correspondence utilizing shared memory. In this paper, we depict a mapping algorithm that matches parallel application processes as a virtual topology with the physical topology of the target machine to improve the application performance as well as reducing the power consumption. It is also a way to integrate triple models of the parallel programming model as a hybrid tri-model to develop the performance of parallel applications and that combine OpenACC, OpenMP and OpenMPI. We actualized these contributions (the mapping algorithm based on the hybrid model) and accomplished fantastic outcomes to reduce the power consumption and improve the performance of the application.
|