Skip to content

Commit

Permalink
Merge pull request #139 from bettiolm/fix_parmetis_bench
Browse files Browse the repository at this point in the history 
ParMETIS benchmark fixed and enabled in the CI
  • Loading branch information
@boeschf
boeschf authored Dec 9, 2021
2 parents c159466 + 1b5065d commit 9a7a401
Show file tree
Hide file tree
Showing 2 changed files with 26 additions and 13 deletions.
6 changes: 3 additions & 3 deletions .github/workflows/CI.yml
View file
Original file line number Diff line number Diff line change
Expand Up @@ -23,7 +23,7 @@ jobs:
df
cat /proc/cpuinfo
mkdir -p build && cd build
eckit_DIR=/opt/eckit Atlas_DIR=/opt/atlas cmake .. -DUCX_DIR=/opt/ucx -DCMAKE_BUILD_TYPE=Debug -DGHEX_ENABLE_ATLAS_BINDINGS=ON -DGHEX_USE_UCP=ON -DGHEX_BUILD_FORTRAN=ON -DCMAKE_Fortran_FLAGS="-cpp -ffree-line-length-none -fcoarray=single" -DGHEX_BUILD_TESTS=ON -DGHEX_BUILD_BENCHMARKS=ON -DMPIEXEC_PREFLAGS=-oversubscribe
eckit_DIR=/opt/eckit Atlas_DIR=/opt/atlas cmake .. -DUCX_DIR=/opt/ucx -DMETIS_INCLUDE_DIR=/opt/metis/include -DMETIS_LIB_DIR=/opt/metis/lib -DPARMETIS_INCLUDE_DIR=/opt/parmetis/include -DPARMETIS_LIB_DIR=/opt/parmetis/lib -DCMAKE_BUILD_TYPE=Debug -DGHEX_USE_UCP=ON -DGHEX_ENABLE_ATLAS_BINDINGS=ON -DGHEX_ENABLE_PARMETIS_BINDINGS=ON -DGHEX_BUILD_FORTRAN=ON -DCMAKE_Fortran_FLAGS="-cpp -ffree-line-length-none -fcoarray=single" -DGHEX_BUILD_TESTS=ON -DGHEX_BUILD_BENCHMARKS=ON -DMPIEXEC_PREFLAGS=-oversubscribe
- name: Build
run: cmake --build build --parallel 2
- name: Execute tests
Expand All @@ -45,7 +45,7 @@ jobs:
df
cat /proc/cpuinfo
mkdir -p build && cd build
CXX=/opt/rocm/llvm/bin/clang++ cmake .. -DUSE_GPU=ON -DUCX_DIR=/opt/ucx -DCMAKE_BUILD_TYPE=Debug -DGHEX_USE_UCP=ON -DGHEX_BUILD_FORTRAN=ON -DCMAKE_Fortran_FLAGS="-cpp -ffree-line-length-none -fcoarray=single" -DGHEX_BUILD_TESTS=ON -DGHEX_BUILD_BENCHMARKS=ON
CXX=/opt/rocm/llvm/bin/clang++ cmake .. -DUSE_GPU=ON -DUCX_DIR=/opt/ucx -DMETIS_INCLUDE_DIR=/opt/metis/include -DMETIS_LIB_DIR=/opt/metis/lib -DPARMETIS_INCLUDE_DIR=/opt/parmetis/include -DPARMETIS_LIB_DIR=/opt/parmetis/lib -DCMAKE_BUILD_TYPE=Debug -DGHEX_USE_UCP=ON -DGHEX_ENABLE_PARMETIS_BINDINGS=ON -DGHEX_BUILD_FORTRAN=ON -DCMAKE_Fortran_FLAGS="-cpp -ffree-line-length-none -fcoarray=single" -DGHEX_BUILD_TESTS=ON -DGHEX_BUILD_BENCHMARKS=ON
- name: Build
run: cmake --build build --parallel 2

Expand All @@ -65,6 +65,6 @@ jobs:
df
cat /proc/cpuinfo
mkdir -p build && cd build
cmake .. -DUSE_GPU=ON -DUCX_DIR=/opt/ucx -DCMAKE_BUILD_TYPE=Debug -DGHEX_USE_UCP=ON -DGHEX_BUILD_FORTRAN=ON -DCMAKE_Fortran_FLAGS="-cpp -ffree-line-length-none -fcoarray=single" -DGHEX_BUILD_TESTS=ON -DGHEX_BUILD_BENCHMARKS=ON
cmake .. -DUSE_GPU=ON -DUCX_DIR=/opt/ucx -DMETIS_INCLUDE_DIR=/opt/metis/include -DMETIS_LIB_DIR=/opt/metis/lib -DPARMETIS_INCLUDE_DIR=/opt/parmetis/include -DPARMETIS_LIB_DIR=/opt/parmetis/lib -DCMAKE_BUILD_TYPE=Debug -DGHEX_USE_UCP=ON -DGHEX_ENABLE_PARMETIS_BINDINGS=ON -DGHEX_BUILD_FORTRAN=ON -DCMAKE_Fortran_FLAGS="-cpp -ffree-line-length-none -fcoarray=single" -DGHEX_BUILD_TESTS=ON -DGHEX_BUILD_BENCHMARKS=ON
- name: Build
run: cmake --build build --parallel 2
33 changes: 23 additions & 10 deletions benchmarks/unstructured_parmetis.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -44,6 +44,7 @@
#include <ghex/common/defs.hpp>
#ifdef GHEX_CUDACC
#include <ghex/common/cuda_runtime.hpp>
#include <ghex/arch_traits.hpp>
#include <ghex/allocator/cuda_allocator.hpp>
#include <ghex/cuda_utils/error.hpp>
#endif
Expand All @@ -68,6 +69,7 @@ template<typename T>
using gpu_allocator_type = gridtools::ghex::allocator::cuda::allocator<T>;
template<typename T>
using data_descriptor_gpu_type = gridtools::ghex::unstructured::data_descriptor<gridtools::ghex::gpu, domain_id_type, global_index_type, T>;
using device_id_type = gridtools::ghex::arch_traits<gridtools::ghex::gpu>::device_id_type;
#endif


Expand Down Expand Up @@ -169,7 +171,7 @@ struct d_v_pair {

using vertices_dist_type = std::map<int, std::map<d_v_pair<domain_id_type, idx_t>, std::vector<idx_t>>>;
using domain_vertices_dist_type = std::map<domain_id_type, std::map<idx_t, std::vector<idx_t>>>;
domain_vertices_dist_type distribute_parmetis(const vertices_dist_type& vertices_dist, std::size_t n_vertices, MPI_Comm comm) {
domain_vertices_dist_type distribute_parmetis(vertices_dist_type& vertices_dist, std::size_t n_vertices, MPI_Comm comm) {

int size;
MPI_Comm_size(comm, &size);
Expand Down Expand Up @@ -374,7 +376,7 @@ TEST(unstructured_parmetis, receive_type) {
vertices_dist_type vertices_dist{};
for (idx_t v_id = vtxdist_v[rank], i = 0; i < static_cast<idx_t>(ap_n.size() - 1); ++v_id, ++i) {
vertices_dist[domain_to_rank(part_v[i], num_threads)]
.insert(std::make_pair(d_v_pair<domain_id_type, idx_t>{part_v[i], v_id}, std::vector<idx_t>{ai.begin() + ap_n[i], ai.begin() + ap_n[i+1]}));
.insert(std::make_pair(d_v_pair<domain_id_type, idx_t>{static_cast<domain_id_type>(part_v[i]), v_id}, std::vector<idx_t>{ai.begin() + ap_n[i], ai.begin() + ap_n[i+1]}));
}
auto domain_vertices_dist = distribute_parmetis(vertices_dist, ap_n.size() - 1, comm);

Expand All @@ -392,7 +394,7 @@ TEST(unstructured_parmetis, receive_type) {
int gh_rank = context.rank();

// barrier
gridtools::ghex::tl::barrier_t gh_barrier{num_threads};
gridtools::ghex::tl::barrier_t gh_barrier{static_cast<std::size_t>(num_threads)};

// timers (local = rank local)
#ifdef GHEX_PARMETIS_BENCHMARK_UNORDERED
Expand Down Expand Up @@ -432,7 +434,7 @@ TEST(unstructured_parmetis, receive_type) {
for (auto d_id : rank_to_domains<domain_id_type>(gh_rank, num_threads)) {
std::vector<idx_t> vertices{};
vertices.reserve(domain_vertices_dist[d_id].size()); // any missing domain gets actually inserted into the map here
std::vector<idx_t> adjncy{}; // size may be computed in advance, not preformance critical anyway // HERE
std::vector<idx_t> adjncy{}; // size may be computed in advance, not preformance critical anyway
for (const auto& v_a_pair : domain_vertices_dist[d_id]) {
vertices.push_back(v_a_pair.first);
adjncy.insert(adjncy.end(), v_a_pair.second.begin(), v_a_pair.second.end());
Expand Down Expand Up @@ -639,6 +641,10 @@ TEST(unstructured_parmetis, receive_type) {
auto& context = *context_ptr;
int gh_rank = context.rank();
auto gh_comm = context.get_communicator();
int num_devices;
GHEX_CHECK_CUDA_RESULT(cudaGetDeviceCount(&num_devices));
device_id_type device_id = gh_rank % num_devices;
GHEX_CHECK_CUDA_RESULT(cudaSetDevice(device_id));

// timers
timer_type t_buf_local_gpu, t_buf_global_gpu; // 1 - unordered halos - buffered receive
Expand All @@ -650,7 +656,7 @@ TEST(unstructured_parmetis, receive_type) {
ss_file << gh_rank;
std::string filename = "unstructured_parmetis_receive_type_gpu_" + ss_file.str() + ".txt";
std::ofstream file(filename.c_str());
file << "Unstructured ParMETIS receive type benchmark\n\n";
file << "Unstructured ParMETIS receive type benchmark; DEBUG: GPU device id = " << device_id << "\n\n";

// GPU allocator
gpu_allocator_type<idx_t> gpu_alloc{};
Expand All @@ -659,7 +665,14 @@ TEST(unstructured_parmetis, receive_type) {

// setup
domain_id_type d_id{gh_rank}; // 1 domain per rank
domain_descriptor_type d{d_id, r_v_ids_rank, r_adjncy_rank, levels}; // CSR constructor
std::vector<idx_t> vertices{};
vertices.reserve(domain_vertices_dist[d_id].size()); // any missing domain gets actually inserted into the map here
std::vector<idx_t> adjncy{}; // size may be computed in advance, not preformance critical anyway
for (const auto& v_a_pair : domain_vertices_dist[d_id]) {
vertices.push_back(v_a_pair.first);
adjncy.insert(adjncy.end(), v_a_pair.second.begin(), v_a_pair.second.end());
}
domain_descriptor_type d{d_id, vertices, adjncy, levels}; // CSR constructor
std::vector<domain_descriptor_type> local_domains{d};
halo_generator_type hg{};
auto p = gridtools::ghex::make_pattern<grid_type>(context, hg, local_domains);
Expand All @@ -669,7 +682,7 @@ TEST(unstructured_parmetis, receive_type) {
initialize_field(d, f_cpu, d_id_offset);
idx_t* f_gpu = gpu_alloc.allocate(d.size() * d.levels());
GHEX_CHECK_CUDA_RESULT(cudaMemcpy(f_gpu, f_cpu.data(), d.size() * d.levels() * sizeof(idx_t), cudaMemcpyHostToDevice));
data_descriptor_gpu_type<idx_t> data_gpu{d, f_gpu, 0};
data_descriptor_gpu_type<idx_t> data_gpu{d, f_gpu, device_id};

// exchange
for (int i = 0; i < n_iters_warm_up; ++i) { // warm-up
Expand Down Expand Up @@ -707,7 +720,7 @@ TEST(unstructured_parmetis, receive_type) {
initialize_field(d_ord, f_ord_cpu, d_id_offset);
idx_t* f_ord_gpu = gpu_alloc.allocate(d_ord.size() * d_ord.levels());
GHEX_CHECK_CUDA_RESULT(cudaMemcpy(f_ord_gpu, f_ord_cpu.data(), d_ord.size() * d_ord.levels() * sizeof(idx_t), cudaMemcpyHostToDevice));
data_descriptor_gpu_type<idx_t> data_ord_gpu{d_ord, f_ord_gpu, 0};
data_descriptor_gpu_type<idx_t> data_ord_gpu{d_ord, f_ord_gpu, device_id};

// exchange
for (int i = 0; i < n_iters_warm_up; ++i) { // warm-up
Expand Down Expand Up @@ -742,7 +755,7 @@ TEST(unstructured_parmetis, receive_type) {
initialize_field(d_ord, f_ipr_cpu, d_id_offset);
idx_t* f_ipr_gpu = gpu_alloc.allocate(d_ord.size() * d_ord.levels());
GHEX_CHECK_CUDA_RESULT(cudaMemcpy(f_ipr_gpu, f_ipr_cpu.data(), d_ord.size() * d_ord.levels() * sizeof(idx_t), cudaMemcpyHostToDevice));
data_descriptor_gpu_type<idx_t> data_ipr_gpu{d_ord, f_ipr_gpu, 0};
data_descriptor_gpu_type<idx_t> data_ipr_gpu{d_ord, f_ipr_gpu, device_id};

// exchange
for (int i = 0; i < n_iters_warm_up; ++i) { // warm-up
Expand Down Expand Up @@ -771,7 +784,7 @@ TEST(unstructured_parmetis, receive_type) {

// ======== output =======================================================

idx_t n_halo_vertices_local{d.size() - d.inner_size()}, n_halo_vertices_global;
idx_t n_halo_vertices_local{static_cast<idx_t>(d.size() - d.inner_size())}, n_halo_vertices_global;
MPI_Allreduce(&n_halo_vertices_local, &n_halo_vertices_global, 1, MPI_INT64_T, MPI_SUM, context.mpi_comm()); // MPI type set according to parmetis idx type
file << "total exchanged size in GB (assuming value type = idx_t): "
<< static_cast<double>(n_halo_vertices_global * levels * sizeof(idx_t) * 2) / (1024.0 * 1024.0 * 1024.0) << "\n\n";
Expand Down

0 comments on commit 9a7a401

Please sign in to comment.