fizban007 / CoffeeGPU

COmputational Force-FreE Electrodynamics (COFFEE)

How to compile

First clone the directory:

git clone https://github.com/fizban007/CoffeeGPU

Now to build this, you need to have cmake, cuda (optional), mpi, and hdf5 installed. After cloning the repo, do the following:

cd CoffeGPU
mkdir build
cd build
cmake .. -Duse_cuda=1 -Duse_double=0
make

If you instead specify -Duse_cuda=0 then the CPU version will be compiled. -Duse_double controls whether to use double precision. A bunch of executables will be found under CoffeeGPU/bin, corresponding to different setups.

Compilation on tigressdata or tigergpu

To compile this on tigergpu or tigressdata, use the following module load:

module load rh/devtoolset
module load cudatoolkit
module load openmpi/gcc/1.10.2
module load hdf5/gcc
module load boost

Then go into the cloned repo and run this:

cd CoffeeGPU
mkdir build
cd build
cmake3 ..
make

Compilation on Ascent

When on Ascent, use the following module load:

module load cmake/3.14.2
module load gcc/8.1.1
module load cuda/10.1.105
module load spectrum-mpi/10.3.0.0-20190419
module load hdf5/1.10.3
module load  boost/1.66.0

Now go into the cloned repo:

cd CoffeeGPU
mkdir build
cd build
CC="gcc" CXX="g++" cmake ..
make

The default build type is Debug, which is bad for performance. For production one should use cmake .. -DCMAKE_BUILD_TYPE=Release. The CC="gcc" CXX="g++" part seems to be needed on Ascent for mysterious reasons. It is not needed on tigressdata or tigergpu or on my laptop.

To run unit tests in the tests directory, run make check in the build directory. Every time you make, all the tests are compiled but not run. You'll need to manually run them using make check. The first unit test test_stagger.cpp should be a good indication of how to write unit tests. For further information please look at the official documentation at https://github.com/catchorg/Catch2/blob/master/docs/tutorial.md.

Run on Ascent

Note that in order to write data we need to use the scratch (GPFS) directories, e.g.,

/gpfs/wolf/gen127/scratch/userid

After making the executable, go to the bin directory and create the configuration file config.toml (see CoffeeGPU/config.toml.example as an example).

For interactive job submission, first run

bsub -P PROJID -nnodes 1 -W 60 -alloc_flags gpumps -Is $SHELL

PROJID should be GEN127 here, and the number of nodes can be varied (there are 6 GPUs per node). This opens up a new shell. Use export OMP_NUM_THREADS=1 to set OpenMP thread. To run the code, in the folder bin, use the following commands:

mkdir Data
jsrun --smpiargs="-gpu" -n4 -a1 -c1 -g1 ./coffee

Here -n gives the number of resource sets; each resource set includes -a number of MPI tasks, -c number of CPU cores and -g number of GPUs. --smpiargs="-gpu" is needed in order to enable CUDA-Aware MPI.

To submit a batch job, we can use the command

bsub submit.lsf

And an example for the submit file submit.lsf is the following:

#!/bin/bash
# Begin LSF Directives
#BSUB -P GEN127
#BSUB -W 0:30
#BSUB -nnodes 2
#BSUB -alloc_flags gpumps
#BSUB -J emwave
#BSUB -o emwave.%J
#BSUB -e emwave.%J

module load cmake/3.14.2
module load gcc/8.1.1
module load cuda/10.1.105
module load spectrum-mpi/10.3.0.0-20190419
module load hdf5/1.10.3
module load  boost/1.66.0
mkdir Data
jsrun --smpiargs="-gpu" -n8 -a1 -c1 -g1 ./coffee

How the code is structured

At the lowest level we have a multi_array class that abstracts away the memory management from the user (but provides a way to access the underlying raw pointers). The multi_array always allocates a 1D chunk of memory and maps any 3D indices onto the linear memory using a Fortran-like indexing, i.e. array(i, j, k) returns the element idx = i + j * width + k * width * height. Communication between the device and host is done manually using sync_to_device() and sync_to_host() functions.

A Grid object keeps track of the underlying grid structuring, including the dimensions of the grid, sizes in numerical units, number of guard cells, etc. This object is copied to the device constant memory at the start of the program.

sim_params holds all of the runtime parameters for the simulation. It is read from config.toml at initialization, and copied to device constant memory as well.

The vector_field class is basically a bundle of 3 multi_arrays, with a stagger structure, and has a pointer to the Grid object it lives on. Every component of the vector_field has a stagger. By default all fields are edge-centered, meaning that E[i] will be staggered in all directions except i. The default convention is that E field will be edge-centered and B field will be face-centered. A vector field can be initialized with a lambda function as follows:

vector_field<double> f(grid);
f.initialize(0, [](double x, double y, double z){
    return x / sqrt(x * x + y * y + z * z);
});
// Now f[0] will be initialized to x/r over the entire domain

The sim_data class is a bundle of all the vector_fields involved in the simulation. The data output routine should take the sim_data and serialize it to HDF5.

The sim_environment class handles most of the initialization, including domain decomposition and managing the parameters. All the other modules in the code should be initialized using only sim_environment only, as it contains all contextual information of any core algorithms.

The modules to be worked on will be the core field update module, the MPI communication module, and the HDF5 output module. The `main.cpp** file will initialize these modules sequentially, while including appropriate user files that define initial and boundary conditions.

In order to have the ability to switch between using float and double, a cmake flag is introduced. The code defines an alias Scalar which by default is equivalent to float. If one adds -Duse_double=1 when running cmake, then Scalar will be mapped to double instead. Remember to rerun make after changing the configuration like this.

Examples for how to write kernels

Here is a simple example for how to write CUDA kernels. A CUDA kernel is a C function which has a decorator __global__, and has to return void. Lets say we defined two vector fields, and want to add them together, here is an example of how to do that:

#include "cuda/constant_mem.h"

__global__ void
add_fields(const Scalar* a, const Scalar* b, Scalar* c) {
  for (int k = threadIdx.z + blockIdx.z * blockDim.z;
       k < dev_grid.dims[2];
       k += blockDim.z * gridDim.z) {
    for (int j = threadIdx.y + blockIdx.y * blockDim.y;
        j < dev_grid.dims[1];
        j += blockDim.y * gridDim.y) {
      for (int i = threadIdx.x + blockIdx.x * blockDim.x;
          i < dev_grid.dims[0];
          i += blockDim.x * gridDim.x) {
        size_t idx = i + j * dev_grid.dims[0]
                     + k * dev_grid.dims[0] * dev_grid.dims[1];
        c[idx] = a[idx] + b[idx];
      }
    }
  }
}

Of course, if you simply want to iterate over the whole array then triple loop is overkill, but in the field solver where indices count, then this is how to do it. We used "grid-stride loops", so that the grid sizes do not need to be exact multiple of the block dimensions that we use to launch the kernel. dev_grid lives in constant memory and holds the grid dimensions and deltas that we specify. In the kernel parameters, notice we use const to specify input arrays and no const for output arrays.

To invoke the kernel, you should use:

vector_field<Scalar> a(grid), b(grid), c(grid);
a.assign(1.0);
b.assign(2.0);

dim3 gridSize(16, 16, 16);
dim3 blockSize(8, 8, 8);
add_fields<<<gridSize, blockSize>>>(a.dev_ptr(0), b.dev_ptr(0), c.dev_ptr(0));
add_fields<<<gridSize, blockSize>>>(a.dev_ptr(1), b.dev_ptr(1), c.dev_ptr(1));
add_fields<<<gridSize, blockSize>>>(a.dev_ptr(2), b.dev_ptr(2), c.dev_ptr(2));
// Now all components of c will be equal to 3