Demo application: Shallow-water model
To show you what mpi4jax
is capable of, we include a full implementation of a physical nonlinear shallow-water model.
A shallow-water model simulates the evolution of the sea surface if temperature and salinity of the water do not vary with depth. Our nonlinear implementation is even capable of modelling turbulence. A possible solution looks like this:
The demo script is too long to include here, but you can
download it
or see the source here.
Running the demo
Apart from mpi4jax
, you will need some additional requirements to run the demo:
$ pip install matploblib tqdm
Then, you can run it like this:
$ mpirun -n 4 python shallow_water.py
WARNING:absl:No GPU/TPU found, falling back to CPU. (Set TF_CPP_MIN_LOG_LEVEL=0 and rerun for more info.)
WARNING:absl:No GPU/TPU found, falling back to CPU. (Set TF_CPP_MIN_LOG_LEVEL=0 and rerun for more info.)
WARNING:absl:No GPU/TPU found, falling back to CPU. (Set TF_CPP_MIN_LOG_LEVEL=0 and rerun for more info.)
WARNING:absl:No GPU/TPU found, falling back to CPU. (Set TF_CPP_MIN_LOG_LEVEL=0 and rerun for more info.)
100%|█████████▉| 9.98/10.00 [00:28<00:00, 2.90s/model day]
Solution took 25.79s
This will execute the demo on 4 processes and show you the results in a matplotlib
animation.
Benchmarks
Using the shallow water solver, we can observe how the performance behaves when we increase the number of MPI processes or switch to GPUs. Here we show some benchmark results on a machine with 2x Intel Xeon E5-2650 v4 CPUs and 2x NVIDIA Tesla P100 GPUs.
Note
To amortize the constant computational cost of using JAX and MPI, we used a 100x bigger domain for the following benchmarks (array shape (3600, 1800)
).
# CPU
$ JAX_PLATFORM_NAME=cpu mpirun -n 1 -- python examples/shallow_water.py --benchmark
92%|█████████▏| 0.09/0.10 [01:55<00:09, 1248.13s/model day]
Solution took 111.95s
$ JAX_PLATFORM_NAME=cpu mpirun -n 2 -- python examples/shallow_water.py --benchmark
92%|█████████▏| 0.09/0.10 [01:33<00:07, 1010.01s/model day]
Solution took 89.67s
$ JAX_PLATFORM_NAME=cpu mpirun -n 4 -- python examples/shallow_water.py --benchmark
92%|█████████▏| 0.09/0.10 [00:41<00:03, 451.75s/model day]
Solution took 38.57s
$ JAX_PLATFORM_NAME=cpu mpirun -n 6 -- python examples/shallow_water.py --benchmark
92%|█████████▏| 0.09/0.10 [00:31<00:02, 345.56s/model day]
Solution took 28.70s
$ JAX_PLATFORM_NAME=cpu mpirun -n 8 -- python examples/shallow_water.py --benchmark
92%|█████████▏| 0.09/0.10 [00:23<00:01, 260.17s/model day]
Solution took 20.62s
$ JAX_PLATFORM_NAME=cpu mpirun -n 16 -- python examples/shallow_water.py --benchmark
92%|█████████▏| 0.09/0.10 [00:19<00:01, 208.55s/model day]
Solution took 15.73s
# GPU
$ JAX_PLATFORM_NAME=gpu mpirun -n 1 -- python examples/shallow_water.py --benchmark
92%|█████████▏| 0.09/0.10 [00:09<00:00, 103.18s/model day]
Solution took 6.28s
$ JAX_PLATFORM_NAME=gpu MPI4JAX_USE_CUDA_MPI=0 mpirun -n 2 -- python examples/shallow_water.py --benchmark
92%|█████████▏| 0.09/0.10 [00:07<00:00, 76.42s/model day]
Solution took 3.87s
$ JAX_PLATFORM_NAME=gpu MPI4JAX_USE_CUDA_MPI=1 mpirun -n 2 -- python examples/shallow_water.py --benchmark
92%|█████████▏| 0.09/0.10 [00:07<00:00, 76.28s/model day]
Solution took 3.89s
Using pure NumPy, the same model takes over 12 minutes (770s) to execute on 1 process, which goes to show how efficient JAX is for this kind of workload.