The Integrated Forecasting System (IFS) application software is a state-of-the-art global weather prediction suite developed at ECMWF on behalf of its member states.
The Integrated Forecasting System (IFS) application software is a state-of-the-art global weather prediction suite developed at ECMWF on behalf of its member states. It is used operationally on a daily basis to provide medium-range and extended-range global weather forecasts to member states' national weather services. It is also the basis of the European Copernicus climate monitoring service and is used for seasonal forecasts and climate projections as part of EC-Earth.
IFS is a large and mature software package that has been developed for over 30 years and has been run on a large range of hardware architectures, including shared and distributed memory machines, vector and scalar multi-core processors. More recently it has been run on large parts of Piz-Daint and on Summit. The code has a hybrid parallel design, with MPI- based coarse-grain domain decomposition and OpenMP-based threading. Input and output are managed by dedicated I/O tasks. The code is written in a mixture of Fortran 90 (F90) and F2003/2008, with some C library dependencies.
IFS is packaged as a Real Applications on Parallel Systems (RAPS) benchmarking suite that ECMWF pioneered in 1992. This allows rapid portability between CPU-based machines. In addition, a number of computational kernels or mini-apps (so called weather & climate dwarfs pioneered in the H2020 projects ESCAPE/ESCAPE-2) are available that extract computational motives relevant to Numerical Weather Prediction. Some dwarfs have architecture-specific versions (CPU, GPU, and FPGA) or implement DEEP SEA target programming models (OmpSs, GPI). They will be further developed in order to showcase developments in DEEP-SEA, at the node and system levels. Specifically, there will be a focus on understanding the performance impact of the memory system, and on illustrating DEEP- SEA developments aimed at easing the use of heterogeneous memory resources. Along with that, work distribution paradigms (OpenMP, OmpSs) will be tested. Effort will also be made to ensure that Numerical Weather Prediction (NWP) benchmarks developed in ESCAPE-2 make good use of the DEEP SEA stack. Profiling activities of the full IFS on DEEP-SEA will profit from previous experience with, e.g., Extrae profiling tools in addition to the existing IFS profiling instrumentation with Dr. Hook. Given its hybrid programming model, IFS will also be benchmarked with the MPC runtime.