How to Exploit Current and Future HPC Machines Efficiently. Four key areas can be pointed out which must be invested to prepare software for exascale platforms, namely performance profiling, deployment of new programming models, development of new libraries and application co-design. Profiling, using the tools available today on HPC platforms (see above), enables users to locate performance bottlenecks and produces efficiency metrics in terms of both FLOPS and power consumption. Programming models are key to exploiting future exascale platforms. For instance, the most common programming paradigms, namely MPI and OpenMP, have had their standards extended to prepare for exascale. Moreover, fewer common paradigms are becoming more popular, such as Global Address Space (PGAS) Languages, e.g., UPC, Co-Array Fortran, and GASPI, will aid programmers when seeking to exploit future hardware. Given the heterogeneous nature of future exascale systems, which will likely include GPUs and other accelerators such as FPGAs, their associated programming models, e.g., CUDA, OpenACC, OpenMP-4.0 directives (planned), and Hybrid programming, e.g., MPI+X, and OmpSs, are all becoming increasingly important. Employing efficient libraries can typically save the programmer from “reinventing the wheel”, given that the libraries are often tuned for the target platform, and yet the programmer’s code stack can remain portable. Lastly, co-design is where HPC vendors and end users can design the future HPC systems together, thus avoiding HPC systems that grab the headlines with powerful performance figures but are impractical for massive scale scientific computation. Application codes rarely perform and scale well when first parallelised: each doubling of scale typically exposes a new issue. Ensuring your application will scale on HPC systems - both today and on the exascale systems of the future - requires stepwise increasing of scale and validation of correctness, debugging, performance analysis and tuning. Then, the same process is repeated for each significant code extension or optimisation.
