Leveraging the Task-Aware GASPI's One-Sided Communications in Saiph

Saiph is a Domain-Specific Language (DSL) that eases the simulation of physical phenomena from the Computational Fluid Dynamics (CFD) domain in HPC environments. Saiph offers a high level of abstraction and a numerical library implementing parallel Finite Difference Methods (FDM) and explicit time solvers. Internally, Saiph features multiple backends to generate different parallel versions of high-level applications. There is one back-end implemented only with MPI and another back-end implemented with MPI and OmpSs-2 tasks that leverages the Task-Aware MPI library. The MPI-only approach manually overlaps computations and communications using non-blocking MPI primitives. In contrast, the hybrid TAMPI approach fully taskifies both computations and communications using task data dependencies and the TAMPI library, allowing the natural and automatic overlap of both phases.

In this technical report, we extend Saiph with a new back-end that generates a hybrid task-based GASPI+OmpSs-2 variant that leverages the one-sided communications provided by the Task- Aware GASPI library. The one-sided (or RMA) communications allow the user applications to reduce the communication granularity and scale better to a high number of computing nodes.