temporary parameter tweaks for testing

rapidsai · Sep 25, 2024 · 8be2a3f · 8be2a3f
1 parent 3bb6602
commit 8be2a3f
Show file tree

Hide file tree

Showing 3 changed files with 11 additions and 5 deletions.
diff --git a/cpp/tests/utilities/base_fixture.hpp b/cpp/tests/utilities/base_fixture.hpp
@@ -77,10 +77,16 @@ inline auto make_pool(bool use_max = false)
   // effect the maximum amount of parallel tests, and therefore `tests/CMakeLists.txt`
   // `_CUGRAPH_TEST_PERCENT` default value will need to be audited.
   auto const [free, total] = rmm::available_device_memory();
-  auto const min_alloc =
-    use_max ? rmm::align_down(std::min(free, total / 2), rmm::CUDA_ALLOCATION_ALIGNMENT)
+  // EOS: 1 node 0.94 succeeded 0.95 failed, 2+ nodes 0.97 succeeded 0.98 failed
+  auto const init_alloc =
+    use_max ? rmm::align_down(std::min(free, static_cast<size_t>(total * 0.93)), rmm::CUDA_ALLOCATION_ALIGNMENT)
             : rmm::align_down(std::min(free, total / 10), rmm::CUDA_ALLOCATION_ALIGNMENT);
-  return rmm::mr::make_owning_wrapper<rmm::mr::pool_memory_resource>(make_cuda(), min_alloc);
+  std::optional<size_t> max_alloc{};
+  if (use_max) {
+    max_alloc = init_alloc;
+  }
+  std::cout << "init_alloc ratio=" << static_cast<double>(init_alloc) / static_cast<double>(total) << std::endl;
+  return rmm::mr::make_owning_wrapper<rmm::mr::pool_memory_resource>(make_cuda(), init_alloc, max_alloc);
 }
 
 inline auto make_binning()

diff --git a/cpp/tests/utilities/mg_utilities.cpp b/cpp/tests/utilities/mg_utilities.cpp
@@ -60,7 +60,7 @@ std::unique_ptr<raft::handle_t> initialize_mg_handle(size_t pool_size)
     --gpu_row_comm_size;
   }
 
-  cugraph::partition_manager::init_subcomm(*handle, gpu_row_comm_size);
+  cugraph::partition_manager::init_subcomm(*handle, std::max(comm_size / 8, 1));
 
   return std::move(handle);
 }

diff --git a/cpp/tests/utilities/test_graphs.hpp b/cpp/tests/utilities/test_graphs.hpp
@@ -246,7 +246,7 @@ class Rmat_Usecase : public detail::TranslateGraph_Usecase {
     // cuMemAddressReserve
     // (https://developer.nvidia.com/blog/introducing-low-level-gpu-virtual-memory-management), we
     // can reduce the temporary memory requirement to (1 / num_partitions) * (original data size)
-    size_t constexpr num_partitions_per_gpu = 4;
+    size_t constexpr num_partitions_per_gpu = 8;
     size_t num_partitions =
       num_partitions_per_gpu * static_cast<size_t>(multi_gpu ? handle.get_comms().get_size() : 1);