Task-based parallel programming for scalable matrix product algorithms

Task-based programming models have succeeded in gaining the interest of high-performance mathematical software community because they relieve part burden developing and implementing distributed-memory parallel algorithms an efficient portable way.In increasingly larger, more heterogeneous clusters computers, these appear as a way to maintain enhance complex algorithms. However, task-based lack flexibility features that are necessary express elegant compact scalable rely on advanced communication patterns. We show Sequential Task Flow paradigm can be extended write yet routines for linear algebra computations. Although, this work focuses dense General Matrix Multiplication, proposed enable implementation describe resulting GEMM operation. Finally, we present experimental analysis two homogeneous supercomputers showing our approach is competitive up 32,768 CPU cores with state-of-the-art libraries may outperform them some problem dimensions. Although code use GPUs straightforwardly, do not deal case it implies other issues which out scope work.