Vector Particle-In-Cell GPU version (vpic-GPU)

VPIC modfied version with GPU support for Student Cluster Competition 19, SC19

This repository is fork from https://github.com/lanl/vpic. Please refer to the original project for detail information.

We do NOT guarantee the correctness of the answer.

NTHU, Taiwan

Compile

arch/cuda contains compile arguments designed for Nvidia® Tesla® V100.

module load cuda/10.0
mkdir build
cd build
../arch/cuda
make -j

Test case: lpi_2d_F6_test

cd build/bin
mkdir lpi_2d_F6_test
cd lpi_2d_F6_test
cp ~/vpic-GPU/sample/lpi_2d_F6_test .
../vpic ./lpi_2d_F6_test
./lpi_2d_F6_test.Generic --tpp <thread num>

For multiple GPU environment, you should set CUDA_VISIBLE_DEVICES and excute with MPI, each device would be assigned to a MPI process. If you want to assign more process to a device, you could modify mpiSetDevice() inside cuda_src/gpu.cu.

mpirun -n <device num> ./lpi_2d_F6_test.Generic --tpp <thread num>

Notification

1. The checkpointing should work correctly. However, we don't support checkpointing before the first timestep, you should disable checkpointing before the first timestep. (e.g. test case beam_plas)

2. Some features are not supported in this version:

   • emission
   • injection

Performance

lpi_2d_F6_test 2000 timesteps

• GPU version: eee4017/vpic-GPU 50835bb
• CPU version: lanl/vpic 51d05ac

Hardware Specification:

• GPU: Nvidia® Tesla® V100-PCIE-32GB * 4
• CPU: Intel® Xeon® Silver 4110 * 2 (16 core)

Sofware Specification:

• CUDA 10.0
• Intel® Parallel Studio XE 2018 (include icc and intel mpi)

Results

version	resources	threads	mpi process	total time
CPU	16 cores	16	1	5:26.59
CPU	16 cores	1	16	2:59.22
GPU	1 device	4	1	3:08.45
GPU	4 devices	4	4	1:05.34

Contributers of vpic-GPU

@eee4017 @zlsh80826 @ttpssabc