reduce benchmarked with ishmem (#722)

* added reduce benchmark with ishmem

* removed suite definitions from cmake as they were added to CI repo

Author: lslusarczyk

CMakeLists.txt

@@ -326,14 +326,7 @@ if(ENABLE_ISHMEM)
   function(target_link_ishmem exec_name)
     target_link_libraries(${exec_name} ze_loader pmi_simple sma)
     target_link_libraries(${exec_name} ${CMAKE_BINARY_DIR}/lib/libishmem.a)
-  endfunction()
-
-  add_compile_definitions(DRISHMEM)
-
-else()
-
-  function(target_link_ishmem exec_name)
-
+    target_compile_definitions(${exec_name} PRIVATE DRISHMEM)
   endfunction()
 
 endif()

benchmarks/gbench/CMakeLists.txt

@@ -29,80 +29,6 @@ if(CMAKE_PROJECT_NAME STREQUAL PROJECT_NAME)
 
   if(ENABLE_SYCL)
     add_subdirectory(shp)
-
     add_custom_target(xhp-bench DEPENDS mhp-bench shp-bench)
-
-    add_custom_target(jfcst-bench DEPENDS jfcst-bench-results)
-    add_custom_command(
-      OUTPUT jfcst-bench-results
-      # github actions log annotation
-      COMMAND echo "::endgroup::"
-      COMMAND dr-bench clean
-      COMMAND dr-bench suite --gpus 4 --ppn 2
-      COMMAND dr-bench plot
-      DEPENDS xhp-bench)
-
-    add_custom_target(aurora-bench-1 DEPENDS aurora-bench-results)
-    add_custom_command(
-      OUTPUT aurora-bench-results
-      # github actions log annotation
-      COMMAND echo "::endgroup::"
-      COMMAND dr-bench clean
-      COMMAND dr-bench suite --gpus 12 --ppn 12
-      COMMAND dr-bench plot
-      DEPENDS xhp-bench)
-
-    # shp fails with PI_OUT_OF_RESOURCES when GPUS>4, 3DFFT fails always on
-    # 2024.1 and sometimes on 2023.2 other shp benchmark also failed in 2024 on
-    # 8 GPUs in initialization
-    add_custom_target(aurora-bench-1-shplimited
-                      DEPENDS aurora-bench-results-shplimited)
-    add_custom_command(
-      OUTPUT aurora-bench-results-shplimited
-      # github actions log annotation
-      COMMAND echo "::endgroup::"
-      COMMAND dr-bench clean
-      COMMAND dr-bench suite --gpus 12 --ppn 12 --mhp-only
-      COMMAND dr-bench suite --gpus 12 --ppn 12 --reference-only
-      COMMAND dr-bench suite --gpus 4 --ppn 12 --shp-only
-      COMMAND dr-bench plot
-      DEPENDS xhp-bench)
-
-    add_custom_target(aurora-bench-2 DEPENDS aurora-bench-results-2)
-    add_custom_command(
-      OUTPUT aurora-bench-results-2
-      # github actions log annotation
-      COMMAND echo "::endgroup::"
-      COMMAND dr-bench clean
-      COMMAND dr-bench suite --min-gpus 13 --gpus 24 --ppn 12 --mhp-only
-      COMMAND dr-bench suite --gpus 12 --ppn 12
-      COMMAND dr-bench plot
-      DEPENDS xhp-bench)
-
-    add_custom_target(aurora-bench-plus-2 DEPENDS aurora-bench-results-plus-2)
-    add_custom_command(
-      OUTPUT aurora-bench-results-plus-2
-      # github actions log annotation
-      COMMAND echo "::endgroup::"
-      COMMAND dr-bench suite --min-gpus 13 --gpus 24 --ppn 12 --mhp-only
-      COMMAND dr-bench plot
-      DEPENDS xhp-bench)
-
-    add_custom_target(quick-bench-gpu DEPENDS quick-bench-gpu-results)
-    add_custom_command(
-      OUTPUT quick-bench-gpu-results
-      COMMAND dr-bench clean
-      COMMAND dr-bench suite --reps 10 --gpus 2
-      COMMAND dr-bench plot
-      DEPENDS xhp-bench)
-
-    add_custom_target(quick-bench-cpu DEPENDS quick-bench-cpu-results)
-    add_custom_command(
-      OUTPUT quick-bench-cpu-results
-      COMMAND dr-bench clean
-      COMMAND dr-bench suite --reps 10 --cores-per-socket 4
-      COMMAND dr-bench plot
-      DEPENDS xhp-bench)
-
   endif()
 endif()

benchmarks/gbench/common/distributed_vector.cpp

@@ -162,9 +162,9 @@ void check_reduce(T actual) {
   }
 }
 
-static void Reduce_DR(benchmark::State &state) {
+template <class... BackendT> static void Reduce_DR(benchmark::State &state) {
   T actual{};
-  xhp::distributed_vector<T> src(default_vector_size, fill);
+  xhp::distributed_vector<T, BackendT...> src(default_vector_size, fill);
   Stats stats(state, sizeof(T) * src.size(), 0);
   for (auto _ : state) {
     for (std::size_t i = 0; i < default_repetitions; i++) {
@@ -176,6 +176,10 @@ static void Reduce_DR(benchmark::State &state) {
 }
 DR_BENCHMARK(Reduce_DR);
 
+#ifdef DRISHMEM
+DR_BENCHMARK(Reduce_DR<dr::mhp::IshmemBackend>)->Name("Reduce_DR_ishmem");
+#endif
+
 static void Reduce_max_DR(benchmark::State &state) {
   T actual{};
   auto max = [](T x, T y) { return std::max(x, y); };

benchmarks/gbench/mhp/CMakeLists.txt

@@ -36,12 +36,14 @@ endif()
 # mhp-quick-bench is for development. By reducing the number of source files, it
 # builds much faster. Change the source files to match what you need to test. It
 # is OK to commit changes to the source file list.
-add_executable(mhp-quick-bench mhp-bench.cpp ../common/sort.cpp)
+add_executable(mhp-quick-bench mhp-bench.cpp ../common/distributed_vector.cpp)
 
 foreach(mhp-bench-exec IN ITEMS mhp-bench mhp-quick-bench)
   target_compile_definitions(${mhp-bench-exec} PRIVATE BENCH_MHP)
   target_link_libraries(${mhp-bench-exec} benchmark::benchmark cxxopts DR::mpi)
-  target_link_ishmem(${mhp-bench-exec})
+  if(ENABLE_ISHMEM)
+    target_link_ishmem(${mhp-bench-exec})
+  endif()
   if(ENABLE_SYCL)
     target_link_libraries(${mhp-bench-exec} MKL::MKL_DPCPP)
   endif()

test/gtest/mhp/CMakeLists.txt

@@ -64,8 +64,10 @@ add_executable(mhp-quick-test
 target_compile_definitions(mhp-quick-test PRIVATE QUICK_TEST)
 
 foreach(test-exec IN ITEMS mhp-tests mhp-tests-3 mhp-quick-test)
+  if(ENABLE_ISHMEM)
+    target_link_ishmem(${test-exec})
+  endif()
   target_link_libraries(${test-exec} GTest::gtest_main cxxopts DR::mpi)
-  target_link_ishmem(${test-exec})
   set_property(TARGET ${test-exec} PROPERTY RULE_LAUNCH_COMPILE
                                             "${CMAKE_COMMAND} -E time")
 endforeach()