Add groupby_max multi-threaded benchmark (#16154)

This PR adds **groupby_max** multi-threaded benchmark. The benchmark runs multiple **max groupby aggregations** concurrently using one CUDA stream per host thread.

Closes #16134

Authors:
  - Srinivas Yadav (https://github.com/srinivasyadav18)

Approvers:
  - Yunsong Wang (https://github.com/PointKernel)
  - Vukasin Milovanovic (https://github.com/vuule)

URL: #16154

cpp/benchmarks/CMakeLists.txt

@@ -231,8 +231,8 @@ ConfigureBench(
 )
 
 ConfigureNVBench(
-  GROUPBY_NVBENCH groupby/group_max.cpp groupby/group_nunique.cpp groupby/group_rank.cpp
-  groupby/group_struct_keys.cpp
+  GROUPBY_NVBENCH groupby/group_max.cpp groupby/group_max_multithreaded.cpp
+  groupby/group_nunique.cpp groupby/group_rank.cpp groupby/group_struct_keys.cpp
 )
 
 # ##################################################################################################

cpp/benchmarks/groupby/group_max.cpp

@@ -48,20 +48,25 @@ void groupby_max_helper(nvbench::state& state,
       cudf::type_to_id<Type>(), row_count{num_rows}, data_profile{builder});
   }();
 
+  auto const num_aggregations = state.get_int64("num_aggregations");
+
   auto keys_view = keys->view();
   auto gb_obj    = cudf::groupby::groupby(cudf::table_view({keys_view, keys_view, keys_view}));
 
   std::vector<cudf::groupby::aggregation_request> requests;
-  requests.emplace_back(cudf::groupby::aggregation_request());
-  requests[0].values = vals->view();
-  requests[0].aggregations.push_back(cudf::make_max_aggregation<cudf::groupby_aggregation>());
+  for (int64_t i = 0; i < num_aggregations; i++) {
+    requests.emplace_back(cudf::groupby::aggregation_request());
+    requests[i].values = vals->view();
+    requests[i].aggregations.push_back(cudf::make_max_aggregation<cudf::groupby_aggregation>());
+  }
 
   auto const mem_stats_logger = cudf::memory_stats_logger();
   state.set_cuda_stream(nvbench::make_cuda_stream_view(cudf::get_default_stream().value()));
   state.exec(nvbench::exec_tag::sync,
              [&](nvbench::launch& launch) { auto const result = gb_obj.aggregate(requests); });
   auto const elapsed_time = state.get_summary("nv/cold/time/gpu/mean").get_float64("value");
-  state.add_element_count(static_cast<double>(num_rows) / elapsed_time / 1'000'000., "Mrows/s");
+  state.add_element_count(
+    static_cast<double>(num_rows * num_aggregations) / elapsed_time / 1'000'000., "Mrows/s");
   state.add_buffer_size(
     mem_stats_logger.peak_memory_usage(), "peak_memory_usage", "peak_memory_usage");
 }
@@ -91,7 +96,8 @@ NVBENCH_BENCH_TYPES(bench_groupby_max,
   .set_name("groupby_max")
   .add_int64_axis("cardinality", {0})
   .add_int64_power_of_two_axis("num_rows", {12, 18, 24})
-  .add_float64_axis("null_probability", {0, 0.1, 0.9});
+  .add_float64_axis("null_probability", {0, 0.1, 0.9})
+  .add_int64_axis("num_aggregations", {1, 2, 4, 8, 16, 32});
 
 NVBENCH_BENCH_TYPES(bench_groupby_max_cardinality, NVBENCH_TYPE_AXES(nvbench::type_list<int32_t>))
   .set_name("groupby_max_cardinality")

cpp/benchmarks/groupby/group_max_multithreaded.cpp

@@ -0,0 +1,102 @@
+/*
+ * Copyright (c) 2024, NVIDIA CORPORATION.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <benchmarks/common/generate_input.hpp>
+#include <benchmarks/fixture/benchmark_fixture.hpp>
+
+#include <cudf/detail/utilities/stream_pool.hpp>
+#include <cudf/groupby.hpp>
+#include <cudf/utilities/default_stream.hpp>
+#include <cudf/utilities/thread_pool.hpp>
+
+#include <nvbench/nvbench.cuh>
+
+template <typename Type>
+void bench_groupby_max_multithreaded(nvbench::state& state, nvbench::type_list<Type>)
+{
+  auto const cardinality      = static_cast<cudf::size_type>(state.get_int64("cardinality"));
+  auto const num_rows         = static_cast<cudf::size_type>(state.get_int64("num_rows"));
+  auto const null_probability = state.get_float64("null_probability");
+  auto const num_threads      = state.get_int64("num_threads");
+  auto const num_aggregations = state.get_int64("num_aggregations");
+
+  auto const keys = [&] {
+    data_profile const profile =
+      data_profile_builder()
+        .cardinality(cardinality)
+        .no_validity()
+        .distribution(cudf::type_to_id<int32_t>(), distribution_id::UNIFORM, 0, num_rows);
+    return create_random_column(cudf::type_to_id<int32_t>(), row_count{num_rows}, profile);
+  }();
+
+  auto const vals = [&] {
+    auto builder = data_profile_builder().cardinality(0).distribution(
+      cudf::type_to_id<Type>(), distribution_id::UNIFORM, 0, num_rows);
+    if (null_probability > 0) {
+      builder.null_probability(null_probability);
+    } else {
+      builder.no_validity();
+    }
+    return create_random_column(
+      cudf::type_to_id<Type>(), row_count{num_rows}, data_profile{builder});
+  }();
+
+  auto keys_view = keys->view();
+  auto gb_obj    = cudf::groupby::groupby(cudf::table_view({keys_view, keys_view, keys_view}));
+
+  auto streams = cudf::detail::fork_streams(cudf::get_default_stream(), num_threads);
+  cudf::detail::thread_pool threads(num_threads);
+
+  std::vector<std::vector<cudf::groupby::aggregation_request>> requests(num_threads);
+  for (auto& thread_requests : requests) {
+    for (int64_t j = 0; j < num_aggregations; j++) {
+      thread_requests.emplace_back();
+      thread_requests.back().values = vals->view();
+      thread_requests.back().aggregations.push_back(
+        cudf::make_max_aggregation<cudf::groupby_aggregation>());
+    }
+  }
+
+  auto const mem_stats_logger = cudf::memory_stats_logger();
+  state.exec(
+    nvbench::exec_tag::sync | nvbench::exec_tag::timer, [&](nvbench::launch& launch, auto& timer) {
+      auto perform_agg = [&](int64_t index) { gb_obj.aggregate(requests[index], streams[index]); };
+      timer.start();
+      for (int64_t i = 0; i < num_threads; ++i) {
+        threads.submit(perform_agg, i);
+      }
+      threads.wait_for_tasks();
+      cudf::detail::join_streams(streams, cudf::get_default_stream());
+      cudf::get_default_stream().synchronize();
+      timer.stop();
+    });
+
+  auto const elapsed_time = state.get_summary("nv/cold/time/gpu/mean").get_float64("value");
+  state.add_element_count(
+    static_cast<double>(num_rows * num_threads * num_aggregations) / elapsed_time / 1'000'000.,
+    "Mrows/s");
+  state.add_buffer_size(
+    mem_stats_logger.peak_memory_usage(), "peak_memory_usage", "peak_memory_usage");
+}
+
+NVBENCH_BENCH_TYPES(bench_groupby_max_multithreaded,
+                    NVBENCH_TYPE_AXES(nvbench::type_list<int32_t, int64_t, float, double>))
+  .set_name("groupby_max_multithreaded")
+  .add_int64_axis("cardinality", {0})
+  .add_int64_power_of_two_axis("num_rows", {12, 18})
+  .add_float64_axis("null_probability", {0, 0.1, 0.9})
+  .add_int64_axis("num_aggregations", {1})
+  .add_int64_axis("num_threads", {1, 2, 4, 8});