第5回 XcalableMPワークショップ

Adapting complex code to take advantage of new computing HPC architectures is a cumbersome task. The development life cycle of weather and climate models does not match the fast pace of the new hardware releases. Therefore, restructuring the code and applying specific architecture optimization is often needed to get optimal performance when porting models to those new architectures. This leads to multiple specific versions of the same code optimized for a single target supercomputer and not performance portable among them. In order to keep a single source code and get performance portability on different hardware, we propose a one-column model abstraction for physical parameterizations. It supports domain scientists by removing the burden of HPC specific optimizations. We are developing a directive language and a tool named CLAW that is able to apply automatic code transformation on abstracted code to produce parallelized versions for dedicated target architecture.

The PGAS programming model aims at reducing communication overhead by decoupling communication and synchronization. The DASH project provides an abstraction of the PGAS programming model that is entirely based on C++11/14 and MPI-3 RMA. At it's core are a set of distributed containers and parallel algorithms operating on them. DASH has so far relied on global synchronization primitives to ensure correct ordering of global memory accesses, which limits its scalability. We try to address this issue by extending the concept of task data dependencies known from OpenMP to describe the data flow between tasks operating in the global memory space across process boundaries. This talk will present our approach together with a prototypical implementation and provide some promising early results.