[coll] Improve column split tests with named threads. (#10735)

include/xgboost/windefs.h

@@ -30,4 +30,4 @@
 
 #endif  // xgboost_IS_MINGW
 
-#endif  // defined(xgboost_IS_WIN)
+#endif  // !defined(xgboost_IS_WIN)

src/collective/loop.cc

@@ -14,6 +14,7 @@
 #include <thread>     // for thread
 #include <utility>    // for move
 
+#include "../common/threading_utils.h"      // for NameThread
 #include "xgboost/collective/poll_utils.h"  // for PollHelper
 #include "xgboost/collective/result.h"      // for Fail, Success
 #include "xgboost/collective/socket.h"      // for FailWithCode
@@ -271,5 +272,6 @@ Loop::Loop(std::chrono::seconds timeout) : timeout_{timeout} {
   worker_ = std::thread{[this] {
     this->Process();
   }};
+  common::NameThread(&worker_, "lw");
 }
 }  // namespace xgboost::collective

src/collective/tracker.cc

@@ -23,6 +23,7 @@
 #include <utility>    // for move, forward
 
 #include "../common/json_utils.h"
+#include "../common/threading_utils.h"  // for NameThread
 #include "comm.h"
 #include "protocol.h"  // for kMagic, PeerInfo
 #include "tracker.h"
@@ -143,6 +144,8 @@ Result RabitTracker::Bootstrap(std::vector<WorkerProxy>* p_workers) {
       Json::Dump(jnext, &str);
       worker.Send(StringView{str});
     });
+    std::string name = "tkbs_t-" + std::to_string(r);
+    common::NameThread(&bootstrap_threads.back(), name.c_str());
   }
 
   for (auto& t : bootstrap_threads) {

src/common/threading_utils.cc

@@ -1,5 +1,5 @@
 /**
- * Copyright 2022-2023 by XGBoost Contributors
+ * Copyright 2022-2024, XGBoost Contributors
  */
 #include "threading_utils.h"
 
@@ -11,6 +11,10 @@
 
 #include "common.h"           // for DivRoundUp
 
+#if defined(__linux__)
+#include <pthread.h>
+#endif
+
 namespace xgboost::common {
 /**
  * Modified from
@@ -113,4 +117,26 @@ std::int32_t OmpGetNumThreads(std::int32_t n_threads) {
   n_threads = std::max(n_threads, 1);
   return n_threads;
 }
+
+void NameThread(std::thread* t, StringView name) {
+#if defined(__linux__)
+  auto handle = t->native_handle();
+  char old[16];
+  auto ret = pthread_getname_np(handle, old, 16);
+  if (ret != 0) {
+    LOG(WARNING) << "Failed to get the name from thread";
+  }
+  auto new_name = std::string{old} + ">" + name.c_str();  // NOLINT
+  if (new_name.size() > 15) {
+    new_name = new_name.substr(new_name.size() - 15);
+  }
+  ret = pthread_setname_np(handle, new_name.c_str());
+  if (ret != 0) {
+    LOG(WARNING) << "Failed to name thread:" << ret << " :" << new_name;
+  }
+#else
+  (void)name;
+  (void)t;
+#endif
+}
 }  // namespace xgboost::common

src/common/threading_utils.h

@@ -1,5 +1,5 @@
 /**
- * Copyright 2019-2023 by XGBoost Contributors
+ * Copyright 2019-2024, XGBoost Contributors
  */
 #ifndef XGBOOST_COMMON_THREADING_UTILS_H_
 #define XGBOOST_COMMON_THREADING_UTILS_H_
@@ -11,12 +11,13 @@
 #include <cstddef>      // for size_t
 #include <cstdint>      // for int32_t
 #include <cstdlib>      // for malloc, free
-#include <functional>   // for function
 #include <new>          // for bad_alloc
+#include <thread>       // for thread
 #include <type_traits>  // for is_signed, conditional_t, is_integral_v, invoke_result_t
 #include <vector>       // for vector
 
 #include "xgboost/logging.h"
+#include "xgboost/string_view.h"  // for StringView
 
 #if !defined(_OPENMP)
 extern "C" {
@@ -308,6 +309,11 @@ class MemStackAllocator {
  * \brief Constant that can be used for initializing static thread local memory.
  */
 std::int32_t constexpr DefaultMaxThreads() { return 128; }
+
+/**
+ * @brief Give the thread a name. Supports only pthread on linux.
+ */
+void NameThread(std::thread* t, StringView name);
 }  // namespace xgboost::common
 
 #endif  // XGBOOST_COMMON_THREADING_UTILS_H_

src/common/threadpool.h

@@ -9,11 +9,15 @@
 #include <memory>              // for make_shared
 #include <mutex>               // for mutex, unique_lock
 #include <queue>               // for queue
+#include <string>              // for string
 #include <thread>              // for thread
 #include <type_traits>         // for invoke_result_t
 #include <utility>             // for move
 #include <vector>              // for vector
 
+#include "threading_utils.h"      // for NameThread
+#include "xgboost/string_view.h"  // for StringView
+
 namespace xgboost::common {
 /**
  * @brief Simple implementation of a thread pool.
@@ -27,11 +31,12 @@ class ThreadPool {
 
  public:
   /**
+   * @param name      Name prefix for threads.
    * @param n_threads The number of threads this pool should hold.
    * @param init_fn   Function called once during thread creation.
    */
   template <typename InitFn>
-  explicit ThreadPool(std::int32_t n_threads, InitFn&& init_fn) {
+  explicit ThreadPool(StringView name, std::int32_t n_threads, InitFn&& init_fn) {
     for (std::int32_t i = 0; i < n_threads; ++i) {
       pool_.emplace_back([&, init_fn = std::forward<InitFn>(init_fn)] {
         init_fn();
@@ -55,6 +60,8 @@ class ThreadPool {
           fn();
         }
       });
+      std::string name_i = name.c_str() + std::string{"-"} + std::to_string(i);  // NOLINT
+      NameThread(&pool_.back(), name_i);
     }
   }
 

src/data/sparse_page_source.h

@@ -336,7 +336,7 @@ class SparsePageSourceImpl : public BatchIteratorImpl<S>, public FormatStreamPol
  public:
   SparsePageSourceImpl(float missing, int nthreads, bst_feature_t n_features, bst_idx_t n_batches,
                        std::shared_ptr<Cache> cache)
-      : workers_{std::max(2, std::min(nthreads, 16)), InitNewThread{}},
+      : workers_{StringView{"ext-mem"}, std::max(2, std::min(nthreads, 16)), InitNewThread{}},
         missing_{missing},
         nthreads_{nthreads},
         n_features_{n_features},

tests/cpp/collective/test_worker.h

@@ -12,10 +12,11 @@
 #include <utility>  // for move
 #include <vector>   // for vector
 
-#include "../../../src/collective/comm.h"
+#include "../../../src/collective/comm.h"              // for RabitComm
 #include "../../../src/collective/communicator-inl.h"  // for Init, Finalize
 #include "../../../src/collective/tracker.h"           // for GetHostAddress
 #include "../../../src/common/cuda_rt_utils.h"         // for AllVisibleGPUs
+#include "../../../src/common/threading_utils.h"       // for NameThread
 #include "../helpers.h"                                // for FileExists
 
 #if defined(XGBOOST_USE_FEDERATED)
@@ -176,6 +177,9 @@ void TestDistributedGlobal(std::int32_t n_workers, WorkerFn worker_fn, bool need
       CHECK(status == std::future_status::ready) << "Test timeout";
       fut.get();
     });
+
+    std::string name = "tw-" + std::to_string(i);
+    common::NameThread(&workers.back(), name.c_str());
   }
 
   for (auto& t : workers) {
@@ -199,7 +203,7 @@ class BaseMGPUTest : public ::testing::Test {
    *                          available.
    */
   template <typename Fn>
-  auto DoTest(Fn&& fn, bool is_federated, bool emulate_if_single = false) const {
+  auto DoTest(Fn&& fn, bool is_federated, [[maybe_unused]] bool emulate_if_single = false) const {
     auto n_gpus = common::AllVisibleGPUs();
     if (is_federated) {
 #if defined(XGBOOST_USE_FEDERATED)

tests/cpp/common/test_threadpool.cc

@@ -21,7 +21,7 @@ TEST(ThreadPool, Basic) {
   // 4 is an invalid value, it's only possible to set it by bypassing the parameter
   // validation.
   ASSERT_NE(orig, GlobalConfigThreadLocalStore::Get()->verbosity);
-  ThreadPool pool{n_threads, [config = *GlobalConfigThreadLocalStore::Get()] {
+  ThreadPool pool{StringView{"test"}, n_threads, [config = *GlobalConfigThreadLocalStore::Get()] {
                     *GlobalConfigThreadLocalStore::Get() = config;
                   }};
   GlobalConfigThreadLocalStore::Get()->verbosity = orig;  // restore

tests/cpp/test_learner.cc

@@ -745,8 +745,7 @@ void VerifyColumnSplitWithArgs(std::string const& tree_method, bool use_gpu, Arg
   std::shared_ptr<DMatrix> sliced{p_fmat->SliceCol(world_size, rank)};
   std::string device = "cpu";
   if (use_gpu) {
-    auto gpu_id = common::AllVisibleGPUs() == 1 ? 0 : rank;
-    device = "cuda:" + std::to_string(gpu_id);
+    device = MakeCUDACtx(DistGpuIdx()).DeviceName();
   }
   auto model = GetModelWithArgs(sliced, tree_method, device, args);
   ASSERT_EQ(model, expected_model);
@@ -807,44 +806,32 @@ class ColumnSplitTrainingTest
   }
 };
 
-auto MakeParamsForTest() {
-  std::vector<std::tuple<std::string, bool, bool>> configs;
-  for (auto tm : {"hist", "approx"}) {
-#if defined(XGBOOST_USE_CUDA)
-    std::array<bool, 2> use_gpu{true, false};
-#else
-    std::array<bool, 1> use_gpu{false};
-#endif
-    for (auto i : use_gpu) {
+auto WithFed() {
 #if defined(XGBOOST_USE_FEDERATED)
-      std::array<bool, 2> fed{true, false};
+  return ::testing::Bool();
 #else
-      std::array<bool, 1> fed{false};
+  return ::testing::Values(false);
 #endif
-      for (auto j : fed) {
-        configs.emplace_back(tm, i, j);
-      }
-    }
-  }
-  return configs;
 }
 }  // anonymous namespace
 
 TEST_P(ColumnSplitTrainingTest, ColumnSampler) {
-  auto param = GetParam();
-  std::apply(TestColumnSplitColumnSampler, param);
+  std::apply(TestColumnSplitColumnSampler, GetParam());
 }
 
 TEST_P(ColumnSplitTrainingTest, InteractionConstraints) {
-  auto param = GetParam();
-  std::apply(TestColumnSplitInteractionConstraints, param);
+  std::apply(TestColumnSplitInteractionConstraints, GetParam());
 }
 
 TEST_P(ColumnSplitTrainingTest, MonotoneConstraints) {
-  auto param = GetParam();
-  std::apply(TestColumnSplitMonotoneConstraints, param);
+  std::apply(TestColumnSplitMonotoneConstraints, GetParam());
 }
 
-INSTANTIATE_TEST_SUITE_P(ColumnSplit, ColumnSplitTrainingTest,
-                         ::testing::ValuesIn(MakeParamsForTest()));
+INSTANTIATE_TEST_SUITE_P(Cpu, ColumnSplitTrainingTest,
+                         ::testing::Combine(::testing::Values("hist", "approx"),
+                                            ::testing::Values(false), WithFed()));
+
+INSTANTIATE_TEST_SUITE_P(MGPU, ColumnSplitTrainingTest,
+                         ::testing::Combine(::testing::Values("hist", "approx"),
+                                            ::testing::Values(true), WithFed()));
 }  // namespace xgboost