The model further supports transparent execution of designated tasks on heterogeneous platforms, including clusters of GPUs.
SIMPLY FORTRAN APGRAPHS CODE
The experimental resuls validate the theoretical model of data ow performance and show that the functionalities introduced in this thesis present good performance.ĪBSTRACT StarSs is a task-based programming model that allows to parallelize sequential applications by means of annotating the code with compiler directives. We implemented all the work introduced in our TALM data ow model and executed experiments with such implementations. In this thesis we introduce the following contributions to data ow execution: (i) novel static/dynamic scheduling algorithms for dataflow runtimes (ii) theoretical tools that allow us to model the performance of dynamic dataflow programs (iii) an algorithm for error detection and recovery in data ow execution and (iv) a model for GPU+CPU execution that incorporates GPU functionalities to the dataflow graph.
Besides that, there has been no work in the area of Error Detection and Recovery that targets data ow execution. Prior work on bounds for the performance of parallel programs mostly focus on DAGs (Directed Acyclic Graphs), a model that can not directly represent data ow programs. The shift of focus toward data ow calls for research that expands the knowledge about dataflow execution and that adds functionalities to the spectrum of what can be done with data ow. Recently, dataflow execution has regained traction as a tool for programming in the multicore and manycore era. Data ow execution, where instructions can start executing as soon as their input operands are ready, is a natural way to obtain parallelism.
0 kommentar(er)
