Aim: WP1 will identify and define Weather & Climate Dwarfs,
a prerequisite for other work-packages and an important facilitator for
co-design and future code development. Moreover, WP1 in collaboration
with other work packages will provide prototypical implementations,
delivering a toolbox with NWP specific algorithms that can be classified
by their data flow and data locality, their parallel communication
pattern, their energy efficiency, their need for numerical precision and
accuracy, and their floating point rate on various HPC architectures.
WP1 will make available and further develop new data structures,
mathematical algorithms and numerical methods including time-stepping
strategies that allow global and regional forecast models to employ
structured and unstructured meshes as well as multi-scale numerical
techniques for enhancing parallelism and scientific flexibility. Approach and methodology:
A significant contribution to this work package will be drawn from the
research supported by the frontier research project PantaRhei as well as
the ECMWF scalability programme. As part of these projects, substantial
advances have been made and software developed at ECMWF that implements
alternative options and generalized formulations for geophysical fluid
dynamics in support of future operational NWP. Through ESCAPE, these
developments will be shared with the European NWP community, and
rigorously tested and optimized for future HPC architectures. Specific
developments will focus on the topic of horizontal and vertical
discretization and on scalable solvers, with both topics using
hierarchical meshes and multi-level themes. A key focus area of
this work package will be an adaptation of the “Berkeley Dwarfs”
(Asanovic et al. 2006) and the “mini-apps” (Heroux et al., 2009) themes.
Their translation to NWP modelling (Weather & Climate Dwarfs)
necessitates work on the following topics leading to Weather & Climate Dwarfs: - Spectral transforms,
involving the Fast Fourier Transform (FFT) and the (Fast) Legendre
Transform (FLT) kernels. The spectral transforms constitute a
substantial part of the compute-intensive part, and thus the energy
consumption in global, high-resolution simulations with IFS. But despite
their cost, their application provides the tool for a very fast direct
solver, which in combination with large time steps still gives the
shortest time-to-solution in operational NWP today. This dwarf thus
serves both as an important part of the reference solution as well as a
tool for exploring the acceleration and energy-efficiency of this
technique on emerging hardware architectures, e.g. optical processors
which promise to substantially accelerate FFTs and matrix-matrix
multiplications (with the latter being the most costly part of the
Legendre transform computations).
- Structured grids,
or block-structured grids, the former oriented along latitudes and
longitudes of the spherical Earth still forming the basis of most
operational NWP model computations today. Supporting grids with some
structure can have advantages in both scalability as well as accuracy.
The vertical column level structure of essentially all NWP models today
is not least a reflection of the dominant hydrostatic balance of the
atmosphere, and the columnar structure is an essential part of all
physical parameterization and observation operator packages that
substantially contribute to the accuracy of NWP models.
- Unstructured meshes,
facilitating locally compact stencils (thus global communication
avoiding) and facilitating refinement. The use of unstructured meshes is
new in operational NWP (with the exception of the semi-structured
icosahedral grid models) and there are numerous developments in this
area to exploit the flexibility of different meshes for future NWP and
climate simulations.
- Advective transport strategies,
important for both the evolution of the moist-dynamic atmosphere as
well as for chemical tracer advection, as needed for ECMWF’s atmospheric
composition services (Copernicus). Semi-Lagrangian transport techniques
have been used so far with great success due to the robustness and the
large time steps that can be afforded. On the other hand, Eulerian
flux-form schemes have been successfully applied in cloud-resolving
simulations, where monotone, sign-preserving transport is important. The
implications for energy-aware metrics has yet to be established when
these different techniques are applied in the context of NWP on emerging
hardware architectures.
- Time-stepping strategies,
important for achieving minimal time-to-solution in the “critical path”
for operational NWP applications. Typically, the best possible solution
that can be afforded must be completed within less than one hour, or
equivalently, 200-300 times faster than real time. The time-stepping
strategy is not necessarily independent from the transport choices.
Semi-implicit time stepping is the standard reference solution applied
in IFS, but several horizontally-explicit and vertically implicit
solution procedures emerge for applications in NWP.
- Bespoke scalable, preconditioned, non-symmetric Krylov solvers,
arising in large time-step, semi-implicit problems in compressible or
sound-proof atmospheric models. These solvers can be self-adaptive, and
may provide robust solution procedures for NWP applications, but their
scalability and competitiveness must be explored in the context of NWP.
An important aspect is successful pre-conditioning of the global NWP
problem.
- Hierarchies, including multi-scale,
multi-level and multi-grid techniques for use in process-dependent and
scale-selective algorithms and data layouts. There is also substantial
scope for extracting parallelism and energy-efficiency in using multiple
grid options for specific processes, e.g. radiation, physical and
chemical parameterizations, dynamic or static mesh adaptation, and
employing spatial and temporal filtering. Some of these options will be
encapsulated in Weather & Climate Dwarfs.
|