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
Michael Lange ([email protected]), Willem Deconinck ([email protected]), Balthasar Reuter ([email protected])
dwarf-p-cloudsc
is distributed under the Apache Licence Version 2.0. See
LICENSE file for details.
- 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 (deprecated!): 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. This variant is defunct
on current Nvidia GPUs and therefore deactivated by default, requiring explicit
--with-claw
flag to build. - 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.
- dwarf-cloudsc-gpu-scc-cuf: GPU-enabled and optimized version of
CLOUDSC that uses the SCC loop layout in combination with CUDA-Fortran
(CUF) to explicitly allocate temporary arrays in device memory and
move parameter structures to constant memory. To enable this variant,
a suitable CUDA installation is required and the
--with-cuda
flag needs to be passed at the build stage. - dwarf-cloudsc-gpu-scc-cuf-k-caching: GPU-enabled and further
optimized version of CLOUDSC that uses the SCC loop layout in
combination with loop fusion and temporary local array demotion, implemented
using CUDA-Fortran (CUF). To enable this variant,
a suitable CUDA installation is required and the
--with-cuda
flag needs to be passed at the build stage. - CUDA C prototypes: To enable these variants, a suitable
CUDA installation is required and the
--with-cuda
flag needs to be pased at the build stage. - dwarf-cloudsc-cuda: GPU-enabled, CUDA C version of CLOUDSC.
- dwarf-cloudsc-cuda-hoist: GPU-enabled, optimized CUDA C version of CLOUDSC including host side hoisted temporary local variables.
- dwarf-cloudsc-cuda-k-caching: GPU-enabled, further optimized CUDA C version of CLOUDSC including loop fusion and temporary local array demotion.
- dwarf-cloudsc-gpu-scc-field: GPU-enabled and optimized version of
CLOUDSC that uses the SCC loop layout, and a dedicated Fortran FIELD
API to manage device offload and copyback. The intent is to demonstrate
the explicit use of pinned host memory to speed-up data transfers, as
provided by the shipped prototype implmentation, and investigate the
effect of different data storage allocation layouts. To enable this
variant, a suitable CUDA installation is required and the
--with-cuda
flag needs to be passed at the build stage.
The code is written in Fortran 2003 and it has been tested using the various compilers, including:
GCC 7.3, 9.3, 11.2
Cray 8.7.7
NVHPC 20.9, 22.1
Intel (classic)
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.
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
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"
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 (default with HDF5 1.10+).
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.
To build on ECMWF's Atos BullSequana XH2000 supercomputer, run the following commands:
./cloudsc-bundle create
./cloudsc-bundle build --arch arch/ecmwf/hpc2020/compiler/version [--single-precision] [--with-mpi]
Currently available compiler/version
selections are:
gnu/9.3.0
andgnu/11.2.0
intel/2021.4.0
nvhpc/22.1
(use with--with-gpu
on AC's GPU partition)
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.
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 double-precision build with the GNU 11.2.0 compiler, performance of ~73 GF/s is achieved.
To run the GPU variant on AC, which includes some GPU nodes, allocate an interactive session on a GPU node and run the binary as usual:
srun -N1 -q ng -p gpu --gres=gpu:4 --mem 200G --pty /bin/bash
bin/dwarf-cloudsc-gpu-scc-hoist 1 262144 128
For a double-precision build with NVHPC 22.1, performance of ~340 GF/s on a single GPU is achieved.
A multi-GPU run requires MPI (build with --with-mpi
) with a dedicated MPI
task for each GPU and (at the moment) manually assigning CUDA devices to each
rank, as Slurm is not yet fully configured for the GPU partition.
To use four GPUs on one node, allocate the relevant resources
salloc -N 1 --tasks-per-node 4 -q ng -p gpu --gres=gpu:4 --mem 200G
and then run the binary like this:
srun bash -c "CUDA_VISIBLE_DEVICES=\$SLURM_LOCALID bin/dwarf-cloudsc-gpu-scc-hoist 1 \$((\$SLURM_NPROCS*262144)) 128"
In principle, the same should work for multi-node execution (-N 2
, -N 4
etc.) once interconnect issues are resolved.
For GPU-enabled runs two internal timer results are reported:
- The isolated compute time of the main compute kernel on device (where
#BLKS == 1
) - The overall time of the execution loop including data offload and copyback
It is important to note that due to the nature of the kernel, data
transfer overheads will dominate timings, and that most supported GPU
variants aim to optimise compute kernel timings only. However, a
dedicated variant dwarf-cloudsc-gpu-scc-field
has been added to
explore host-side memory pinning, which improves data transfer times
and alternative data layout strategies. By default, this will allocate
each array variable individually in pinned memory. A runtime flag
CLOUDSC_PACKED_STORAGE=ON
can be used to enable "packed" storage,
where multiple arrays are stored in a single base allocation, eg.
NV_ACC_CUDA_HEAPSIZE=8G CLOUDSC_PACKED_STORAGE=ON ./bin/dwarf-cloudsc-gpu-scc-field 1 80000 128
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, directly or via downstream tools like CLAW.
The following build flags enable the demonstrator build targets on the ECMWF Atos HPC facility's GPU partition:
./cloudsc-bundle build --clean [--with-gpu] --with-loki --loki-frontend=fp --arch=./arch/ecmwf/hpc2020/nvhpc/22.1
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 (deprecated): 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 (deprecated): 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.
To enable the deprecated and, on GPU, defunct CLAW variants, the build-flag
--with-claw
needs to be specified explicitly.
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
.
To automate parameter space sweeps and ease testing across various platforms, a
JUBE benchmark definition is included in
the directory benchmark
. See the included README for
further details and usage instructions.