Enhancing MPI+OpenMP Task Based Applications for Heterogeneous Architectures with GPU Support

Heterogeneous supercomputers are widespread over HPC systems and programming efficient applications on these architectures is a challenge. Task-based models promising way to tackle this Since OpenMP 4.0 4.5, the target directives enable offload pieces of code GPUs express it as tasks with dependencies. Therefore, heterogeneous machines can be programmed using MPI+OpenMP(task+target) exhibit very high level concurrent asynchronous operations for which data transfers, kernel executions, communications CPU computations overlapped. Hence, possible suspend performing CPUs overlap their completion another task execution. Suspended resume once associated event completed in an opportunistic at every scheduling point. We have integrated feature into MPC framework validated AXPY microbenchmark evaluated MPI+OpenMP(tasks) implementation LULESH proxy applications. The results show that we able improve asynchronism overall performance, allowing benefit from execution machines.