build_dir=build
mkdir -p ${build_dir}
pushd ${build_dir}
cmake ../src
cmake --build .CMake will detect the compute capabilities of the GPUs in the node where CMake is run. If building on a different node than the one that will be used for benchmarking, specify the compute capabilities by passing `-DCUDA_ARCHS=70;80" (for 7.0 and 8.0) to CMake.
$ ./TP06 -h
Usage: cell_model_cuda_example [-h|--help] [-V|--version]
[-NDOUBLE|--num_cells=DOUBLE] [-tINT|--num_timesteps=INT]
[--dt=DOUBLE] [-iINT|--device_id=INT] [-sENUM|--scheme=ENUM]
-h, --help Print help and exit
-V, --version Print version and exit
-N, --num_cells=DOUBLE number of cells (independent instances of the ODE
system) (default=`1E7')
-t, --num_timesteps=INT number of time steps (default=`100')
--dt=DOUBLE time step (in ms) used to solve the cell model
(default=`0.001')
-i, --device_id=INT CUDA device to use (default=`0')
-s, --scheme=ENUM numerical scheme used to solve the cell model
(possible values="FE", "RL", "GRL1"
default=`FE')
These performance results were obtained using a single NVIDIA A100-SXM4-80GB GPU from a DGX A100 system.
[kghustad@g002 23:38:18 build_a100q (master)*]$ nvidia-smi
Fri Jul 29 23:40:15 2022
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 515.48.07 Driver Version: 515.48.07 CUDA Version: 11.7 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|===============================+======================+======================|
| 0 NVIDIA A100-SXM... On | 00000000:4C:00.0 Off | 0 |
| N/A 32C P0 64W / 400W | 0MiB / 81920MiB | 0% Default |
| | | Disabled |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=============================================================================|
| No running processes found |
+-----------------------------------------------------------------------------+
Note that cell steps / second is a throughput metric defined as number_of_cells * number_of_time_steps / time_elapsed.
[kghustad@g002 23:37:54 build_a100q (master)*]$ ./TP06 -s FE
Solving with the FE scheme
Solved 100 timesteps with 10000000 cells in 0.584167 seconds.
Throughput: 1.71184e+09 cell steps / second.
Final V values are in the interval [-8.54229956918900655e+01, -8.54229956918900655e+01]
[kghustad@g002 23:38:10 build_a100q (master)*]$ ./TP06 -s RL
Solving with the RL scheme
Solved 100 timesteps with 10000000 cells in 0.710556 seconds.
Throughput: 1.40735e+09 cell steps / second.
Final V values are in the interval [-8.54229956918546520e+01, -8.54229956918546520e+01]
[kghustad@g002 23:38:14 build_a100q (master)*]$ ./TP06 -s GRL1
Solving with the GRL1 scheme
Solved 100 timesteps with 10000000 cells in 0.945808 seconds.
Throughput: 1.0573e+09 cell steps / second.
Final V values are in the interval [-8.54229956921292626e+01, -8.54229956921292626e+01]