dwarf-p-cloudsc

dwarf-p-cloudsc is intended to test the CLOUDSC cloud microphysics scheme of the IFS.

This package is made available to support research collaborations and is not officially supported by ECMWF

Contact

Michael Lange ([email protected]), Willem Deconinck ([email protected]), Balthasar Reuter ([email protected])

Licence

dwarf-p-cloudsc is distributed under the Apache Licence Version 2.0. See LICENSE file for details.

Prototypes available

• dwarf-P-cloudMicrophysics-IFSScheme: The original cloud scheme from IFS that is naturally suited to host-type machines and optimized on the Cray system at ECMWF.
• dwarf-cloudsc-fortran: A cleaned up version of the CLOUDSC prototype that validates runs against platform and language-agnostic off-line reference data via HDF5 or the Serialbox package. The kernel code also is slightly cleaner than the original version.
• dwarf-cloudsc-c: Standalone C version of the kernel that has been generated by ECMWF tools. This relies exclusively on the Serialbox validation mechanism.
• dwarf-cloudsc-gpu-kernels: GPU-enabled version of the CLOUDSC dwarf that uses OpenACC and relies on the !$acc kernels directive to offload the computational kernel.
• dwarf-cloudsc-gpu-claw: GPU-enabled and optimized version of CLOUDSC that is based on an auto-generated version of CLOUDSC based on the CLAW tool. The kernel in this demonstrator has been further optimized with gang-level loop blocking to demonstrate potential performance gains.
• dwarf-cloudsc-gpu-scc: GPU-enabled and optimized version of CLOUDSC that utilises the native blocked IFS memory layout via a "single-column coalesced" (SCC) loop layout. Here the outer NPROMA block loop is mapped to the OpenACC "gang" level and the kernel uses an inverted loop-nest where the outer horizontal loop is mapped to OpenACC " vector" parallelism. This variant lets the CUDA runtime manage temporary arrays and needs a large PGI_ACC_CUDA_HEAPSIZE (eg. PGI_ACC_CUDA_HEAPSIZE=8GB for 160K columns.)
• dwarf-cloudsc-gpu-scc-hoist: GPU-enabled and optimized version of CLOUDSC that also uses the SCC loop layout, but promotes the inner "vector" loop to the driver and declares the kernel as sequential. The block array arguments are fully dimensioned though, and multi-dimensional temporaries have been declared explicitly at the driver level.

Download and Installation

The code is written in Fortran 2003 and it has been tested using the various compilers, including:

GCC 7.3, 9.3
Cray 8.7.7
NVHPC 20.9
Intel

This application does not need MPI nor BLAS libraries for performance. Just a compiler that understands OpenMP directives. Fortran must be at least level F2003.

Inside the dwarf directory you can find some example of outputs inside the example-outputs/ directory.

In addition, to run the dwarf it is necessary to use an input file that can be found inside the config-files/ directory winthin the dwarf folder.

The preferred method to install the CLOUDSC dwarf uses the bundle definition shipped in the main repository. For this please install the bundle via:

./cloudsc-bundle create  # Checks out dependency packages
./cloudsc-bundle build [--build-type=debug|bit|release] [--arch=./arch/ecmwf/machine/compiler/version]

The individual prototype variants of the dwarf are managed as ECBuild features and can be enable or disabled via --cloudsc-<feature>=[ON|OFF] arguments to cloudsc-bundle build.

The use of the boost library or module is required by the Serialbox utility package for filesystem utilities. If boost is not available on a given system, Serialbox's internal "experimental filesystem" can be used via the --serialbox-experimental=ON argument, although this has proven difficult with certain compiler toolchains.

GPU versions of CLOUDSC

The GPU-enabled versions of the dwarf are by default disabled. To enable them use the --with-gpu flag. For example to build on the in-house volta machine:

./cloudsc-bundle create  # Checks out dependency packages
./cloudsc-bundle build --clean --with-gpu --arch=./arch/ecmwf/volta/nvhpc/20.9

MPI-enabled versions of CLOUDSC

Optionally, dwarf-cloudsc-fortran and the GPU versions can be built with MPI support by providing the --with-mpi flag. For example on volta:

./cloudsc-bundle create
./cloudsc-bundle build --clean --with-mpi --with-gpu --arch=./arch/ecmwf/volta/nvhpc/20.9

Running with MPI parallelization distributes the columns of the working set among all ranks. The specified number of OpenMP threads is then spawned on each rank. Results are gathered from all ranks and reported for the global working set. Performance numbers are also gathered and reported per thread, per rank and total.

Important: If the total size of the working set (2nd argument, see "Running and testing") exceeds the number of columns in the input file (the input data in the repository consists of just 100 columns), every rank derives its working set by replicating the columns in the input file, starting with the first column in the file. This means, all ranks effectively work on the same data set. If the total size of the working set is less than or equal to the number of columns in the input file, these are truly distributed and every rank ends up with a different working set.

When running with multiple GPUs each rank needs to be assigned a different device. This can be achieved using the CUDA_VISIBLE_DEVICES environment variable:

mpirun -np 2 bash -c "CUDA_VISIBLE_DEVICES=\${OMPI_COMM_WORLD_RANK} bin/dwarf-cloudsc-gpu-claw 1 163840 8192"

Choosing between HDF5 and Serialbox input file format

The default build configuration relies on HDF5 input and reference data for dwarf-cloudsc-fortran as well as GPU and Loki versions. The original dwarf-P-cloudMicrophysics-IFSScheme always uses raw Fortran binary format.

Please note: The HDF55 installation needs to have the f03 interfaces installed.

As an alternative to HDF5, the Serialbox library can be used to load input and reference data. This, however, requires certain boost libraries or its own internal experimental filesystem, both of which proved difficult on certain compiler toolchains or more exotic hardware architectures.

The original input is provided as raw Fortran binary in prototype1, but input and reference data can be regenerated from this variant by running

CLOUDSC_WRITE_INPUT=1 ./bin/dwarf-P-cloudMicrophysics-IFSScheme 1 100 100
CLOUDSC_WRITE_REFERENCE=1 ./bin/dwarf-P-cloudMicrophysics-IFSScheme 1 100 100

Note that this is only available via Serialbox at the moment. Updates to HDF5 input or reference data have to be done via manual conversion. A small Python script for this with usage instructions can be found in the serialbox2hdf5 directory.

A64FX version of CLOUDSC

Preliminary results for CLOUDSC have been generated for A64FX CPUs on Isambard. A set of arch and toolchain files and detailed installation and run instructions are provided here.

Running and testing

The different prototype variants of the dwarf create different binaries that all behave similarly. The basic three arguments define (in this order):

• Number of OpenMP threads
• Size of overall working set in columns
• Block size (NPROMA) in columns

An example:

cd build
./bin/dwarf-P-cloudMicrophysics-IFSScheme 4 16384 32  # The original
./bin/dwarf-cloudsc-fortran 4 16384 32   # The cleaned-up Fortran
./bin/dwarf-cloudsc-c 4 16384 32   # The standalone C version

On the Atos system, a high-watermark run on a single socket can be performed as follows:

export OMP_NUM_THREADS=64
OMP_PLACES="{$(seq -s '},{' 0 $(($OMP_NUM_THREADS-1)) )}" srun -q np --ntasks=1 --hint=nomultithread --cpus-per-task=$OMP_NUM_THREADS ./bin/dwarf-cloudsc-fortran $OMP_NUM_THREADS 163840 32

For a build with the Intel 2021.1.1 compiler, performance of ~74 GF is achieved.

Loki transformations for CLOUDSC

Loki is an in-house developed source-to-source translation tool that allows us to create bespoke transformations for the IFS to target and experiment with emerging HPC architectures and programming models. We use the CLOUDSC dwarf as a demonstrator for targeted transformation capabilities of physics and grid point computations kernels, including conversion to C and GPU via downstream tools like CLAW.

To use the Loki demonstrators, Loki and CLAW need to be installed as described in the Loki install instructions. Please note that the in-house "volta" machine needs some manual workarounds for this atm.

Once Loki and CLAW are installed and activated via source loki-activate, the following build flags enable the demonstrator build targets:

# For general use on workstations with GNU
# Please note that OpenACC needs to be disable with GNU,
# since CLAW-generated code currently does not comply with GNU.
./cloudsc-bundle build --clean --with-loki --loki-frontend=fp --arch=./arch/ecmwf/leap42/gnu/7.3.0

# For GPU exploration on volta
./cloudsc-bundle build --clean [--with-gpu]--with-loki --loki-frontend=fp --arch=./arch/ecmwf/volta/nvhpc/20.9

The following Loki modes are included in the dwarf, each with a bespoke demonstrator build:

• cloudsc-loki-idem: "Idempotence" mode that performs a full parse-unparse cycle of the kernel and performs various housekeeping transformations, including the driver-level source injection mechanism currently facilitated by Loki.
• cloudsc-loki-sca: Pure single-column mode that strips all horizontal vector loops from the kernel and introduces an outer "column-loop" at the driver level.
• cloudsc-loki-claw-cpu: Same as SCA, but also adds the necessary CLAW annotations. The resulting cloudsc.claw.F90 file is then processed by CLAW to re-insert vector loops for optimal CPU execution.
• cloudsc-loki-claw-gpu: Creates the same CLAW-ready kernel file, but triggers the GPU-specific optimizations in the CLAW compiler to insert OpenACC-offload instructions in the driver and an OpenACC parallel loop inside the kernel for each block. This needs to be run with large block sizes (eg. NPROMA=1024-8192).
• cloudsc-loki-c: A prototype C transpilation pipeline that converts the kernel to C and calls it via iso_c_bindings interfaces from the driver.

A note on frontends

Loki currently supports three frontends to parse the Fortran source code:

• FParser (loki-frontend=fp): The preferred default; developed by STFC for PsyClone.
• OMNI frontend (loki-frontend=omni): Generates the same AST as used by CLAW.
• OFP, a Python wrapper around the ROSE frontend (loki-frontend=ofp): Supported, but bugged in some places and slow; use with care.

For completeness, all three frontends are tested in our CI, which means we require the .xmod module description files for utility routines in src/common for processing the CLOUDSC source files with the OMNI frontend. These are stored in the source under src/cloudsc_loki/xmod.