| Line | % of fetches | Source |
|---|---|---|
| 1 | // Copyright (c) 2007-2015 Hartmut Kaiser | |
| 2 | // | |
| 3 | // Distributed under the Boost Software License, Version 1.0. (See accompanying | |
| 4 | // file LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt) | |
| 5 | ||
| 6 | #include <hpx/runtime/threads/detail/thread_pool.hpp> | |
| 7 | ||
| 8 | #include <hpx/error_code.hpp> | |
| 9 | #include <hpx/exception.hpp> | |
| 10 | #include <hpx/state.hpp> | |
| 11 | #include <hpx/throw_exception.hpp> | |
| 12 | #include <hpx/lcos/local/no_mutex.hpp> | |
| 13 | #include <hpx/runtime/get_worker_thread_num.hpp> | |
| 14 | #include <hpx/runtime/threads/detail/create_thread.hpp> | |
| 15 | #include <hpx/runtime/threads/detail/create_work.hpp> | |
| 16 | #include <hpx/runtime/threads/detail/scheduling_loop.hpp> | |
| 17 | #include <hpx/runtime/threads/detail/set_thread_state.hpp> | |
| 18 | #include <hpx/runtime/threads/detail/thread_num_tss.hpp> | |
| 19 | #include <hpx/runtime/threads/policies/callback_notifier.hpp> | |
| 20 | #include <hpx/runtime/threads/topology.hpp> | |
| 21 | #include <hpx/util/assert.hpp> | |
| 22 | #include <hpx/util/bind.hpp> | |
| 23 | #include <hpx/util/logging.hpp> | |
| 24 | #include <hpx/util/hardware/timestamp.hpp> | |
| 25 | #include <hpx/util/high_resolution_clock.hpp> | |
| 26 | #include <hpx/util/unlock_guard.hpp> | |
| 27 | ||
| 28 | #include <boost/atomic.hpp> | |
| 29 | #include <boost/exception_ptr.hpp> | |
| 30 | #include <boost/system/system_error.hpp> | |
| 31 | #include <boost/thread/barrier.hpp> | |
| 32 | #include <boost/thread/mutex.hpp> | |
| 33 | #include <boost/thread/thread.hpp> | |
| 34 | ||
| 35 | #include <algorithm> | |
| 36 | #include <cstddef> | |
| 37 | #include <cstdint> | |
| 38 | #include <exception> | |
| 39 | #include <functional> | |
| 40 | #include <iomanip> | |
| 41 | #include <mutex> | |
| 42 | #include <numeric> | |
| 43 | ||
| 44 | namespace hpx { namespace threads { namespace detail | |
| 45 | { | |
| 46 | /////////////////////////////////////////////////////////////////////////// | |
| 47 | template <typename Scheduler> | |
| 48 | void thread_pool<Scheduler>::init_tss(std::size_t num) | |
| 49 | { | |
| 50 | thread_num_tss_.init_tss(num); | |
| 51 | } | |
| 52 | ||
| 53 | template <typename Scheduler> | |
| 54 | void thread_pool<Scheduler>::deinit_tss() | |
| 55 | { | |
| 56 | thread_num_tss_.deinit_tss(); | |
| 57 | } | |
| 58 | ||
| 59 | /////////////////////////////////////////////////////////////////////////// | |
| 60 | template <typename Scheduler> | |
| 61 | thread_pool<Scheduler>::thread_pool(Scheduler& sched, | |
| 62 | threads::policies::callback_notifier& notifier, | |
| 63 | char const* pool_name, policies::scheduler_mode m) | |
| 64 | : sched_(sched), | |
| 65 | notifier_(notifier), | |
| 66 | pool_name_(pool_name), | |
| 67 | thread_count_(0), | |
| 68 | used_processing_units_(), | |
| 69 | mode_(m) | |
| 70 | { | |
| 71 | timestamp_scale_ = 1.0; | |
| 72 | } | |
| 73 | ||
| 74 | template <typename Scheduler> | |
| 75 | thread_pool<Scheduler>::~thread_pool() | |
| 76 | { | |
| 77 | if (!threads_.empty()) { | |
| 78 | if (!sched_.has_reached_state(state_suspended)) | |
| 79 | { | |
| 80 | // still running | |
| 81 | lcos::local::no_mutex mtx; | |
| 82 | std::unique_lock<lcos::local::no_mutex> l(mtx); | |
| 83 | stop_locked(l); | |
| 84 | } | |
| 85 | threads_.clear(); | |
| 86 | } | |
| 87 | } | |
| 88 | ||
| 89 | /////////////////////////////////////////////////////////////////////////// | |
| 90 | template <typename Scheduler> | |
| 91 | hpx::state thread_pool<Scheduler>::get_state() const | |
| 92 | { | |
| 93 | // get_worker_thread_num returns the global thread number which might | |
| 94 | // be too large. This function might get called from within | |
| 95 | // background_work inside the os executors | |
| 96 | if (thread_count_ != 0) | |
| 97 | { | |
| 98 | std::size_t num_thread = get_worker_thread_num() % thread_count_; | |
| 99 | if (num_thread != std::size_t(-1)) | |
| 100 | return get_state(num_thread); | |
| 101 | } | |
| 102 | return sched_.get_minmax_state().second; | |
| 103 | } | |
| 104 | ||
| 105 | template <typename Scheduler> | |
| 106 | hpx::state thread_pool<Scheduler>::get_state(std::size_t num_thread) const | |
| 107 | { | |
| 108 | HPX_ASSERT(num_thread != std::size_t(-1)); | |
| 109 | return sched_.get_state(num_thread).load(); | |
| 110 | } | |
| 111 | ||
| 112 | template <typename Scheduler> | |
| 113 | bool thread_pool<Scheduler>::has_reached_state(hpx::state s) const | |
| 114 | { | |
| 115 | return sched_.has_reached_state(s); | |
| 116 | } | |
| 117 | ||
| 118 | /////////////////////////////////////////////////////////////////////////// | |
| 119 | template <typename Scheduler> | |
| 120 | std::size_t thread_pool<Scheduler>::init(std::size_t num_threads, | |
| 121 | policies::init_affinity_data const& data) | |
| 122 | { | |
| 123 | topology const& topology_ = get_topology(); | |
| 124 | std::size_t cores_used = sched_.Scheduler::init(data, topology_); | |
| 125 | ||
| 126 | resize(used_processing_units_, threads::hardware_concurrency()); | |
| 127 | for (std::size_t i = 0; i != num_threads; ++i) | |
| 128 | used_processing_units_ |= sched_.Scheduler::get_pu_mask(topology_, i); | |
| 129 | ||
| 130 | return cores_used; | |
| 131 | } | |
| 132 | ||
| 133 | /////////////////////////////////////////////////////////////////////////// | |
| 134 | template <typename Scheduler> | |
| 135 | std::size_t thread_pool<Scheduler>::get_pu_num(std::size_t num_thread) const | |
| 136 | { | |
| 137 | return sched_.Scheduler::get_pu_num(num_thread); | |
| 138 | } | |
| 139 | ||
| 140 | template <typename Scheduler> | |
| 141 | mask_cref_type thread_pool<Scheduler>::get_pu_mask( | |
| 142 | topology const& topology, std::size_t num_thread) const | |
| 143 | { | |
| 144 | return sched_.Scheduler::get_pu_mask(topology, num_thread); | |
| 145 | } | |
| 146 | ||
| 147 | template <typename Scheduler> | |
| 148 | mask_cref_type thread_pool<Scheduler>::get_used_processing_units() const | |
| 149 | { | |
| 150 | return used_processing_units_; | |
| 151 | } | |
| 152 | ||
| 153 | template <typename Scheduler> | |
| 154 | void thread_pool<Scheduler>::do_some_work(std::size_t num_thread) | |
| 155 | { | |
| 156 | sched_.Scheduler::do_some_work(num_thread); | |
| 157 | } | |
| 158 | ||
| 159 | template <typename Scheduler> | |
| 160 | void thread_pool<Scheduler>::report_error(std::size_t num, | |
| 161 | boost::exception_ptr const& e) | |
| 162 | { | |
| 163 | sched_.set_all_states(state_terminating); | |
| 164 | notifier_.on_error(num, e); | |
| 165 | sched_.Scheduler::on_error(num, e); | |
| 166 | } | |
| 167 | ||
| 168 | /////////////////////////////////////////////////////////////////////////// | |
| 169 | template <typename Scheduler> | |
| 170 | void thread_pool<Scheduler>::create_thread(thread_init_data& data, | |
| 171 | thread_id_type& id, thread_state_enum initial_state, bool run_now, | |
| 172 | error_code& ec) | |
| 173 | { | |
| 174 | // verify state | |
| 175 | if (thread_count_ == 0 && !sched_.is_state(state_running)) | |
| 176 | { | |
| 177 | // thread-manager is not currently running | |
| 178 | HPX_THROWS_IF(ec, invalid_status, | |
| 179 | "thread_pool<Scheduler>::create_thread", | |
| 180 | "invalid state: thread pool is not running"); | |
| 181 | return; | |
| 182 | } | |
| 183 | ||
| 184 | detail::create_thread(&sched_, data, id, initial_state, run_now, ec); //-V601 | |
| 185 | } | |
| 186 | ||
| 187 | template <typename Scheduler> | |
| 188 | void thread_pool<Scheduler>::create_work(thread_init_data& data, | |
| 189 | thread_state_enum initial_state, error_code& ec) | |
| 190 | { | |
| 191 | // verify state | |
| 192 | if (thread_count_ == 0 && !sched_.is_state(state_running)) | |
| 193 | { | |
| 194 | // thread-manager is not currently running | |
| 195 | HPX_THROWS_IF(ec, invalid_status, | |
| 196 | "thread_pool<Scheduler>::create_work", | |
| 197 | "invalid state: thread pool is not running"); | |
| 198 | return; | |
| 199 | } | |
| 200 | ||
| 201 | detail::create_work(&sched_, data, initial_state, ec); //-V601 | |
| 202 | } | |
| 203 | ||
| 204 | template <typename Scheduler> | |
| 205 | thread_state thread_pool<Scheduler>::set_state( | |
| 206 | thread_id_type const& id, thread_state_enum new_state, | |
| 207 | thread_state_ex_enum new_state_ex, thread_priority priority, | |
| 208 | error_code& ec) | |
| 209 | { | |
| 210 | return detail::set_thread_state(id, new_state, //-V107 | |
| 211 | new_state_ex, priority, get_worker_thread_num(), ec); | |
| 212 | } | |
| 213 | ||
| 214 | template <typename Scheduler> | |
| 215 | thread_id_type thread_pool<Scheduler>::set_state( | |
| 216 | util::steady_time_point const& abs_time, | |
| 217 | thread_id_type const& id, thread_state_enum newstate, | |
| 218 | thread_state_ex_enum newstate_ex, thread_priority priority, | |
| 219 | error_code& ec) | |
| 220 | { | |
| 221 | return detail::set_thread_state_timed(sched_, abs_time, id, | |
| 222 | newstate, newstate_ex, priority, get_worker_thread_num(), ec); | |
| 223 | } | |
| 224 | ||
| 225 | template <typename Scheduler> | |
| 226 | void thread_pool<Scheduler>::abort_all_suspended_threads() | |
| 227 | { | |
| 228 | sched_.Scheduler::abort_all_suspended_threads(); | |
| 229 | } | |
| 230 | ||
| 231 | template <typename Scheduler> | |
| 232 | bool thread_pool<Scheduler>::cleanup_terminated(bool delete_all) | |
| 233 | { | |
| 234 | return sched_.Scheduler::cleanup_terminated(delete_all); | |
| 235 | } | |
| 236 | ||
| 237 | /////////////////////////////////////////////////////////////////////////// | |
| 238 | template <typename Scheduler> | |
| 239 | std::size_t thread_pool<Scheduler>::get_worker_thread_num() const | |
| 240 | { | |
| 241 | return thread_num_tss_.get_worker_thread_num(); | |
| 242 | } | |
| 243 | ||
| 244 | template <typename Scheduler> | |
| 245 | boost::thread& thread_pool<Scheduler>::get_os_thread_handle( | |
| 246 | std::size_t num_thread) | |
| 247 | { | |
| 248 | HPX_ASSERT(num_thread < threads_.size()); | |
| 249 | return threads_[threads_.size() - num_thread - 1]; | |
| 250 | } | |
| 251 | ||
| 252 | template <typename Scheduler> | |
| 253 | std::int64_t thread_pool<Scheduler>::get_thread_count( | |
| 254 | thread_state_enum state, thread_priority priority, | |
| 255 | std::size_t num, bool reset) const | |
| 256 | { | |
| 257 | return sched_.Scheduler::get_thread_count(state, priority, num, reset); | |
| 258 | } | |
| 259 | ||
| 260 | template <typename Scheduler> | |
| 261 | bool thread_pool<Scheduler>::enumerate_threads( | |
| 262 | util::function_nonser<bool(thread_id_type)> const& f, | |
| 263 | thread_state_enum state) const | |
| 264 | { | |
| 265 | return sched_.Scheduler::enumerate_threads(f, state); | |
| 266 | } | |
| 267 | ||
| 268 | template <typename Scheduler> | |
| 269 | void thread_pool<Scheduler>::reset_thread_distribution() | |
| 270 | { | |
| 271 | return sched_.Scheduler::reset_thread_distribution(); | |
| 272 | } | |
| 273 | ||
| 274 | template <typename Scheduler> | |
| 275 | void thread_pool<Scheduler>::set_scheduler_mode( | |
| 276 | threads::policies::scheduler_mode mode) | |
| 277 | { | |
| 278 | return sched_.set_scheduler_mode(mode); | |
| 279 | } | |
| 280 | ||
| 281 | /////////////////////////////////////////////////////////////////////////// | |
| 282 | template <typename Scheduler> | |
| 283 | bool thread_pool<Scheduler>::run(std::unique_lock<boost::mutex>& l, | |
| 284 | std::size_t num_threads) | |
| 285 | { | |
| 286 | HPX_ASSERT(l.owns_lock()); | |
| 287 | ||
| 288 | LTM_(info) //-V128 | |
| 289 | << "thread_pool::run: " << pool_name_ | |
| 290 | << " number of processing units available: " //-V128 | |
| 291 | << threads::hardware_concurrency(); | |
| 292 | LTM_(info) //-V128 | |
| 293 | << "thread_pool::run: " << pool_name_ | |
| 294 | << " creating " << num_threads << " OS thread(s)"; //-V128 | |
| 295 | ||
| 296 | if (0 == num_threads) { | |
| 297 | HPX_THROW_EXCEPTION(bad_parameter, | |
| 298 | "thread_pool::run", "number of threads is zero"); | |
| 299 | } | |
| 300 | ||
| 301 | if (!threads_.empty() || sched_.has_reached_state(state_running)) | |
| 302 | return true; // do nothing if already running | |
| 303 | ||
| 304 | executed_threads_.resize(num_threads); | |
| 305 | executed_thread_phases_.resize(num_threads); | |
| 306 | ||
| 307 | tfunc_times_.resize(num_threads); | |
| 308 | exec_times_.resize(num_threads); | |
| 309 | ||
| 310 | reset_tfunc_times_.resize(num_threads); | |
| 311 | ||
| 312 | // scale timestamps to nanoseconds | |
| 313 | std::uint64_t base_timestamp = util::hardware::timestamp(); | |
| 314 | std::uint64_t base_time = util::high_resolution_clock::now(); | |
| 315 | std::uint64_t curr_timestamp = util::hardware::timestamp(); | |
| 316 | std::uint64_t curr_time = util::high_resolution_clock::now(); | |
| 317 | ||
| 318 | while ((curr_time - base_time) <= 100000) | |
| 319 | { | |
| 320 | curr_timestamp = util::hardware::timestamp(); | |
| 321 | curr_time = util::high_resolution_clock::now(); | |
| 322 | } | |
| 323 | ||
| 324 | if (curr_timestamp - base_timestamp != 0) | |
| 325 | { | |
| 326 | timestamp_scale_ = double(curr_time - base_time) / | |
| 327 | double(curr_timestamp - base_timestamp); | |
| 328 | } | |
| 329 | ||
| 330 | #if defined(HPX_HAVE_THREAD_CUMULATIVE_COUNTS) | |
| 331 | // timestamps/values of last reset operation for various performance | |
| 332 | // counters | |
| 333 | reset_executed_threads_.resize(num_threads); | |
| 334 | reset_executed_thread_phases_.resize(num_threads); | |
| 335 | ||
| 336 | #if defined(HPX_HAVE_THREAD_IDLE_RATES) | |
| 337 | // timestamps/values of last reset operation for various performance | |
| 338 | // counters | |
| 339 | reset_thread_duration_.resize(num_threads); | |
| 340 | reset_thread_duration_times_.resize(num_threads); | |
| 341 | ||
| 342 | reset_thread_overhead_.resize(num_threads); | |
| 343 | reset_thread_overhead_times_.resize(num_threads); | |
| 344 | reset_thread_overhead_times_total_.resize(num_threads); | |
| 345 | ||
| 346 | reset_thread_phase_duration_.resize(num_threads); | |
| 347 | reset_thread_phase_duration_times_.resize(num_threads); | |
| 348 | ||
| 349 | reset_thread_phase_overhead_.resize(num_threads); | |
| 350 | reset_thread_phase_overhead_times_.resize(num_threads); | |
| 351 | reset_thread_phase_overhead_times_total_.resize(num_threads); | |
| 352 | ||
| 353 | reset_cumulative_thread_duration_.resize(num_threads); | |
| 354 | ||
| 355 | reset_cumulative_thread_overhead_.resize(num_threads); | |
| 356 | reset_cumulative_thread_overhead_total_.resize(num_threads); | |
| 357 | #endif | |
| 358 | #endif | |
| 359 | ||
| 360 | #if defined(HPX_HAVE_THREAD_IDLE_RATES) | |
| 361 | reset_idle_rate_time_.resize(num_threads); | |
| 362 | reset_idle_rate_time_total_.resize(num_threads); | |
| 363 | ||
| 364 | #if defined(HPX_HAVE_THREAD_CREATION_AND_CLEANUP_RATES) | |
| 365 | reset_creation_idle_rate_time_.resize(num_threads); | |
| 366 | reset_creation_idle_rate_time_total_.resize(num_threads); | |
| 367 | ||
| 368 | reset_cleanup_idle_rate_time_.resize(num_threads); | |
| 369 | reset_cleanup_idle_rate_time_total_.resize(num_threads); | |
| 370 | #endif | |
| 371 | #endif | |
| 372 | ||
| 373 | LTM_(info) | |
| 374 | << "thread_pool::run: " << pool_name_ | |
| 375 | << " timestamp_scale: " << timestamp_scale_; //-V128 | |
| 376 | ||
| 377 | try { | |
| 378 | HPX_ASSERT(startup_.get() == nullptr); | |
| 379 | startup_.reset( | |
| 380 | new boost::barrier(static_cast<unsigned>(num_threads+1)) | |
| 381 | ); | |
| 382 | ||
| 383 | // run threads and wait for initialization to complete | |
| 384 | ||
| 385 | topology const& topology_ = get_topology(); | |
| 386 | ||
| 387 | std::size_t thread_num = num_threads; | |
| 388 | while (thread_num-- != 0) { | |
| 389 | threads::mask_cref_type mask = | |
| 390 | sched_.Scheduler::get_pu_mask(topology_, thread_num); | |
| 391 | ||
| 392 | LTM_(info) //-V128 | |
| 393 | << "thread_pool::run: " << pool_name_ | |
| 394 | << " create OS thread " << thread_num //-V128 | |
| 395 | << ": will run on processing units within this mask: " | |
| 396 | #if !defined(HPX_HAVE_MORE_THAN_64_THREADS) || \ | |
| 397 | (defined(HPX_HAVE_MAX_CPU_COUNT) && HPX_HAVE_MAX_CPU_COUNT <= 64) | |
| 398 | << std::hex << "0x" << mask; | |
| 399 | #else | |
| 400 | << "0b" << mask; | |
| 401 | #endif | |
| 402 | ||
| 403 | // create a new thread | |
| 404 | threads_.push_back(boost::thread( | |
| 405 | &thread_pool::thread_func, this, thread_num, | |
| 406 | std::ref(topology_), std::ref(*startup_) | |
| 407 | )); | |
| 408 | ||
| 409 | // set the new threads affinity (on Windows systems) | |
| 410 | if (any(mask)) | |
| 411 | { | |
| 412 | error_code ec(lightweight); | |
| 413 | topology_.set_thread_affinity_mask(threads_.back(), mask, ec); | |
| 414 | if (ec) | |
| 415 | { | |
| 416 | LTM_(warning) //-V128 | |
| 417 | << "thread_pool::run: " << pool_name_ | |
| 418 | << " setting thread affinity on OS thread " //-V128 | |
| 419 | << thread_num << " failed with: " | |
| 420 | << ec.get_message(); | |
| 421 | } | |
| 422 | } | |
| 423 | else | |
| 424 | { | |
| 425 | LTM_(debug) //-V128 | |
| 426 | << "thread_pool::run: " << pool_name_ | |
| 427 | << " setting thread affinity on OS thread " //-V128 | |
| 428 | << thread_num << " was explicitly disabled."; | |
| 429 | } | |
| 430 | } | |
| 431 | ||
| 432 | // the main thread needs to have a unique thread_num | |
| 433 | init_tss(num_threads); | |
| 434 | startup_->wait(); | |
| 435 | ||
| 436 | // The scheduler is now running. | |
| 437 | sched_.set_all_states(state_running); | |
| 438 | } | |
| 439 | catch (std::exception const& e) { | |
| 440 | LTM_(always) | |
| 441 | << "thread_pool::run: " << pool_name_ | |
| 442 | << " failed with: " << e.what(); | |
| 443 | ||
| 444 | // trigger the barrier | |
| 445 | if (startup_.get() != nullptr) | |
| 446 | { | |
| 447 | while (num_threads-- != 0 && !startup_->wait()) | |
| 448 | ; | |
| 449 | } | |
| 450 | ||
| 451 | stop(l); | |
| 452 | threads_.clear(); | |
| 453 | ||
| 454 | return false; | |
| 455 | } | |
| 456 | ||
| 457 | LTM_(info) << "thread_pool::run: " << pool_name_ << " running"; | |
| 458 | return true; | |
| 459 | } | |
| 460 | ||
| 461 | /////////////////////////////////////////////////////////////////////////// | |
| 462 | template <typename Scheduler> | |
| 463 | void thread_pool<Scheduler>::stop ( | |
| 464 | std::unique_lock<boost::mutex>& l, bool blocking) | |
| 465 | { | |
| 466 | HPX_ASSERT(l.owns_lock()); | |
| 467 | ||
| 468 | return stop_locked(l, blocking); | |
| 469 | } | |
| 470 | ||
| 471 | template <typename Scheduler> | |
| 472 | template <typename Lock> | |
| 473 | void thread_pool<Scheduler>::stop_locked(Lock& l, bool blocking) | |
| 474 | { | |
| 475 | LTM_(info) | |
| 476 | << "thread_pool::stop: " << pool_name_ | |
| 477 | << " blocking(" << std::boolalpha << blocking << ")"; | |
| 478 | ||
| 479 | deinit_tss(); | |
| 480 | ||
| 481 | if (!threads_.empty()) { | |
| 482 | // set state to stopping | |
| 483 | sched_.set_all_states(state_stopping); | |
| 484 | ||
| 485 | // make sure we're not waiting | |
| 486 | sched_.Scheduler::do_some_work(std::size_t(-1)); | |
| 487 | ||
| 488 | if (blocking) { | |
| 489 | for (std::size_t i = 0; i != threads_.size(); ++i) | |
| 490 | { | |
| 491 | // make sure no OS thread is waiting | |
| 492 | LTM_(info) | |
| 493 | << "thread_pool::stop: " << pool_name_ | |
| 494 | << " notify_all"; | |
| 495 | ||
| 496 | sched_.Scheduler::do_some_work(std::size_t(-1)); | |
| 497 | ||
| 498 | LTM_(info) //-V128 | |
| 499 | << "thread_pool::stop: " << pool_name_ | |
| 500 | << " join:" << i; //-V128 | |
| 501 | ||
| 502 | // unlock the lock while joining | |
| 503 | util::unlock_guard<Lock> ul(l); | |
| 504 | threads_[i].join(); | |
| 505 | } | |
| 506 | threads_.clear(); | |
| 507 | } | |
| 508 | } | |
| 509 | } | |
| 510 | ||
| 511 | /////////////////////////////////////////////////////////////////////////// | |
| 512 | struct manage_active_thread_count | |
| 513 | { | |
| 514 | manage_active_thread_count(boost::atomic<long>& counter) | |
| 515 | : counter_(counter) | |
| 516 | { | |
| 517 | ++counter_; | |
| 518 | } | |
| 519 | ~manage_active_thread_count() | |
| 520 | { | |
| 521 | --counter_; | |
| 522 | } | |
| 523 | ||
| 524 | boost::atomic<long>& counter_; | |
| 525 | }; | |
| 526 | ||
| 527 | template <typename Scheduler> | |
| 528 | struct init_tss_helper | |
| 529 | { | |
| 530 | init_tss_helper(thread_pool<Scheduler>& pool, std::size_t thread_num) | |
| 531 | : pool_(pool), thread_num_(thread_num) | |
| 532 | { | |
| 533 | pool.notifier_.on_start_thread(thread_num); | |
| 534 | pool.init_tss(thread_num); | |
| 535 | pool.sched_.Scheduler::on_start_thread(thread_num); | |
| 536 | } | |
| 537 | ~init_tss_helper() | |
| 538 | { | |
| 539 | pool_.sched_.Scheduler::on_stop_thread(thread_num_); | |
| 540 | pool_.deinit_tss(); | |
| 541 | pool_.notifier_.on_stop_thread(thread_num_); | |
| 542 | } | |
| 543 | ||
| 544 | thread_pool<Scheduler>& pool_; | |
| 545 | std::size_t thread_num_; | |
| 546 | }; | |
| 547 | ||
| 548 | template <typename Scheduler> | |
| 549 | void thread_pool<Scheduler>::thread_func(std::size_t num_thread, | |
| 550 | topology const& topology, boost::barrier& startup) | |
| 551 | { | |
| 552 | // Set the affinity for the current thread. | |
| 553 | threads::mask_cref_type mask = | |
| 554 | sched_.Scheduler::get_pu_mask(topology, num_thread); | |
| 555 | ||
| 556 | if (LHPX_ENABLED(debug)) | |
| 557 | topology.write_to_log(); | |
| 558 | ||
| 559 | error_code ec(lightweight); | |
| 560 | if (any(mask)) | |
| 561 | { | |
| 562 | topology.set_thread_affinity_mask(mask, ec); | |
| 563 | if (ec) | |
| 564 | { | |
| 565 | LTM_(warning) //-V128 | |
| 566 | << "thread_pool::thread_func: " << pool_name_ | |
| 567 | << " setting thread affinity on OS thread " //-V128 | |
| 568 | << num_thread << " failed with: " << ec.get_message(); | |
| 569 | } | |
| 570 | } | |
| 571 | else | |
| 572 | { | |
| 573 | LTM_(debug) //-V128 | |
| 574 | << "thread_pool::thread_func: " << pool_name_ | |
| 575 | << " setting thread affinity on OS thread " //-V128 | |
| 576 | << num_thread << " was explicitly disabled."; | |
| 577 | } | |
| 578 | ||
| 579 | // Setting priority of worker threads to a lower priority, this needs to | |
| 580 | // be done in order to give the parcel pool threads higher priority | |
| 581 | if ((mode_ & policies::reduce_thread_priority) && | |
| 582 | any(mask & used_processing_units_)) | |
| 583 | { | |
| 584 | topology.reduce_thread_priority(ec); | |
| 585 | if (ec) | |
| 586 | { | |
| 587 | LTM_(warning) //-V128 | |
| 588 | << "thread_pool::thread_func: " << pool_name_ | |
| 589 | << " reducing thread priority on OS thread " //-V128 | |
| 590 | << num_thread << " failed with: " << ec.get_message(); | |
| 591 | } | |
| 592 | } | |
| 593 | ||
| 594 | // manage the number of this thread in its TSS | |
| 595 | init_tss_helper<Scheduler> tss_helper(*this, num_thread); | |
| 596 | ||
| 597 | // wait for all threads to start up before before starting HPX work | |
| 598 | startup.wait(); | |
| 599 | ||
| 600 | { | |
| 601 | LTM_(info) //-V128 | |
| 602 | << "thread_pool::thread_func: " << pool_name_ | |
| 603 | << " starting OS thread: " << num_thread; //-V128 | |
| 604 | ||
| 605 | try { | |
| 606 | try { | |
| 607 | manage_active_thread_count count(thread_count_); | |
| 608 | ||
| 609 | // run the work queue | |
| 610 | hpx::threads::coroutines::prepare_main_thread main_thread; | |
| 611 | ||
| 612 | // run main Scheduler loop until terminated | |
| 613 | detail::scheduling_counters counters( | |
| 614 | executed_threads_[num_thread], | |
| 615 | executed_thread_phases_[num_thread], | |
| 616 | tfunc_times_[num_thread], exec_times_[num_thread]); | |
| 617 | ||
| 618 | detail::scheduling_callbacks callbacks( | |
| 619 | util::bind( //-V107 | |
| 620 | &policies::scheduler_base::idle_callback, | |
| 621 | &sched_, num_thread | |
| 622 | ), | |
| 623 | detail::scheduling_callbacks::callback_type()); | |
| 624 | ||
| 625 | if (mode_ & policies::do_background_work) | |
| 626 | { | |
| 627 | callbacks.background_ = util::bind( //-V107 | |
| 628 | &policies::scheduler_base::background_callback, | |
| 629 | &sched_, num_thread); | |
| 630 | } | |
| 631 | ||
| 632 | sched_.set_scheduler_mode(mode_); | |
| 633 | detail::scheduling_loop(num_thread, sched_, counters, | |
| 634 | callbacks); | |
| 635 | ||
| 636 | // the OS thread is allowed to exit only if no more HPX | |
| 637 | // threads exist or if some other thread has terminated | |
| 638 | HPX_ASSERT(!sched_.Scheduler::get_thread_count( | |
| 639 | unknown, thread_priority_default, num_thread) || | |
| 640 | sched_.get_state(num_thread) == state_terminating); | |
| 641 | } | |
| 642 | catch (hpx::exception const& e) { | |
| 643 | LFATAL_ //-V128 | |
| 644 | << "thread_pool::thread_func: " << pool_name_ | |
| 645 | << " thread_num:" << num_thread //-V128 | |
| 646 | << " : caught hpx::exception: " | |
| 647 | << e.what() << ", aborted thread execution"; | |
| 648 | ||
| 649 | report_error(num_thread, boost::current_exception()); | |
| 650 | return; | |
| 651 | } | |
| 652 | catch (boost::system::system_error const& e) { | |
| 653 | LFATAL_ //-V128 | |
| 654 | << "thread_pool::thread_func: " << pool_name_ | |
| 655 | << " thread_num:" << num_thread //-V128 | |
| 656 | << " : caught boost::system::system_error: " | |
| 657 | << e.what() << ", aborted thread execution"; | |
| 658 | ||
| 659 | report_error(num_thread, boost::current_exception()); | |
| 660 | return; | |
| 661 | } | |
| 662 | catch (std::exception const& e) { | |
| 663 | // Repackage exceptions to avoid slicing. | |
| 664 | boost::throw_exception(boost::enable_error_info( | |
| 665 | hpx::exception(unhandled_exception, e.what()))); | |
| 666 | } | |
| 667 | } | |
| 668 | catch (...) { | |
| 669 | LFATAL_ //-V128 | |
| 670 | << "thread_pool::thread_func: " << pool_name_ | |
| 671 | << " thread_num:" << num_thread //-V128 | |
| 672 | << " : caught unexpected " //-V128 | |
| 673 | "exception, aborted thread execution"; | |
| 674 | ||
| 675 | report_error(num_thread, boost::current_exception()); | |
| 676 | return; | |
| 677 | } | |
| 678 | ||
| 679 | LTM_(info) //-V128 | |
| 680 | << "thread_pool::thread_func: " << pool_name_ | |
| 681 | << " thread_num: " << num_thread | |
| 682 | << " : ending OS thread, " //-V128 | |
| 683 | "executed " << executed_threads_[num_thread] | |
| 684 | << " HPX threads"; | |
| 685 | } | |
| 686 | } | |
| 687 | ||
| 688 | /////////////////////////////////////////////////////////////////////////// | |
| 689 | // performance counters | |
| 690 | #if defined(HPX_HAVE_THREAD_CUMULATIVE_COUNTS) | |
| 691 | template <typename Scheduler> | |
| 692 | std::int64_t thread_pool<Scheduler>:: | |
| 693 | get_executed_threads(std::size_t num, bool reset) | |
| 694 | { | |
| 695 | std::int64_t executed_threads = 0; | |
| 696 | std::int64_t reset_executed_threads = 0; | |
| 697 | ||
| 698 | if (num != std::size_t(-1)) | |
| 699 | { | |
| 700 | executed_threads = executed_threads_[num]; | |
| 701 | reset_executed_threads = reset_executed_threads_[num]; | |
| 702 | ||
| 703 | if (reset) | |
| 704 | reset_executed_threads_[num] = executed_threads; | |
| 705 | } | |
| 706 | else | |
| 707 | { | |
| 708 | executed_threads = std::accumulate(executed_threads_.begin(), | |
| 709 | executed_threads_.end(), std::int64_t(0)); | |
| 710 | reset_executed_threads = std::accumulate( | |
| 711 | reset_executed_threads_.begin(), | |
| 712 | reset_executed_threads_.end(), std::int64_t(0)); | |
| 713 | ||
| 714 | if (reset) | |
| 715 | { | |
| 716 | std::copy(executed_threads_.begin(), executed_threads_.end(), | |
| 717 | reset_executed_threads_.begin()); | |
| 718 | } | |
| 719 | } | |
| 720 | ||
| 721 | HPX_ASSERT(executed_threads >= reset_executed_threads); | |
| 722 | ||
| 723 | return executed_threads - reset_executed_threads; | |
| 724 | } | |
| 725 | ||
| 726 | template <typename Scheduler> | |
| 727 | std::int64_t thread_pool<Scheduler>:: | |
| 728 | get_executed_thread_phases(std::size_t num, bool reset) | |
| 729 | { | |
| 730 | std::int64_t executed_phases = 0; | |
| 731 | std::int64_t reset_executed_phases = 0; | |
| 732 | ||
| 733 | if (num != std::size_t(-1)) | |
| 734 | { | |
| 735 | executed_phases = executed_thread_phases_[num]; | |
| 736 | reset_executed_phases = reset_executed_thread_phases_[num]; | |
| 737 | ||
| 738 | if (reset) | |
| 739 | reset_executed_thread_phases_[num] = executed_phases; | |
| 740 | } | |
| 741 | else | |
| 742 | { | |
| 743 | executed_phases = std::accumulate(executed_thread_phases_.begin(), | |
| 744 | executed_thread_phases_.end(), std::int64_t(0)); | |
| 745 | reset_executed_phases = std::accumulate( | |
| 746 | reset_executed_thread_phases_.begin(), | |
| 747 | reset_executed_thread_phases_.end(), std::int64_t(0)); | |
| 748 | ||
| 749 | if (reset) | |
| 750 | { | |
| 751 | std::copy(executed_thread_phases_.begin(), | |
| 752 | executed_thread_phases_.end(), | |
| 753 | reset_executed_thread_phases_.begin()); | |
| 754 | } | |
| 755 | } | |
| 756 | ||
| 757 | HPX_ASSERT(executed_phases >= reset_executed_phases); | |
| 758 | ||
| 759 | return executed_phases - reset_executed_phases; | |
| 760 | } | |
| 761 | ||
| 762 | #if defined(HPX_HAVE_THREAD_IDLE_RATES) | |
| 763 | template <typename Scheduler> | |
| 764 | std::int64_t thread_pool<Scheduler>:: | |
| 765 | get_thread_phase_duration(std::size_t num, bool reset) | |
| 766 | { | |
| 767 | std::uint64_t exec_total = 0ul; | |
| 768 | std::int64_t num_phases = 0l; | |
| 769 | std::uint64_t reset_exec_total = 0ul; | |
| 770 | std::int64_t reset_num_phases = 0l; | |
| 771 | ||
| 772 | if (num != std::size_t(-1)) | |
| 773 | { | |
| 774 | exec_total = exec_times_[num]; | |
| 775 | num_phases = executed_thread_phases_[num]; | |
| 776 | ||
| 777 | reset_exec_total = reset_thread_phase_duration_times_[num]; | |
| 778 | reset_num_phases = reset_thread_phase_duration_[num]; | |
| 779 | ||
| 780 | if (reset) | |
| 781 | { | |
| 782 | reset_thread_phase_duration_[num] = num_phases; | |
| 783 | reset_thread_phase_duration_times_[num] = exec_total; | |
| 784 | } | |
| 785 | } | |
| 786 | else | |
| 787 | { | |
| 788 | exec_total = std::accumulate(exec_times_.begin(), | |
| 789 | exec_times_.end(), std::uint64_t(0)); | |
| 790 | num_phases = std::accumulate(executed_thread_phases_.begin(), | |
| 791 | executed_thread_phases_.end(), std::int64_t(0)); | |
| 792 | ||
| 793 | reset_exec_total = std::accumulate( | |
| 794 | reset_thread_phase_duration_times_.begin(), | |
| 795 | reset_thread_phase_duration_times_.end(), std::uint64_t(0)); | |
| 796 | reset_num_phases = std::accumulate( | |
| 797 | reset_thread_phase_duration_.begin(), | |
| 798 | reset_thread_phase_duration_.end(), std::int64_t(0)); | |
| 799 | ||
| 800 | if (reset) | |
| 801 | { | |
| 802 | std::copy(exec_times_.begin(), exec_times_.end(), | |
| 803 | reset_thread_phase_duration_times_.begin()); | |
| 804 | std::copy(executed_thread_phases_.begin(), | |
| 805 | executed_thread_phases_.end(), | |
| 806 | reset_thread_phase_duration_.begin()); | |
| 807 | } | |
| 808 | } | |
| 809 | ||
| 810 | HPX_ASSERT(exec_total >= reset_exec_total); | |
| 811 | HPX_ASSERT(num_phases >= reset_num_phases); | |
| 812 | ||
| 813 | exec_total -= reset_exec_total; | |
| 814 | num_phases -= reset_num_phases; | |
| 815 | ||
| 816 | return std::uint64_t( | |
| 817 | (double(exec_total) * timestamp_scale_) / double(num_phases) | |
| 818 | ); | |
| 819 | } | |
| 820 | ||
| 821 | template <typename Scheduler> | |
| 822 | std::int64_t thread_pool<Scheduler>:: | |
| 823 | get_thread_duration(std::size_t num, bool reset) | |
| 824 | { | |
| 825 | std::uint64_t exec_total = 0ul; | |
| 826 | std::int64_t num_threads = 0l; | |
| 827 | std::uint64_t reset_exec_total = 0ul; | |
| 828 | std::int64_t reset_num_threads = 0l; | |
| 829 | ||
| 830 | if (num != std::size_t(-1)) | |
| 831 | { | |
| 832 | exec_total = exec_times_[num]; | |
| 833 | num_threads = executed_threads_[num]; | |
| 834 | ||
| 835 | reset_exec_total = reset_thread_duration_times_[num]; | |
| 836 | reset_num_threads = reset_thread_duration_[num]; | |
| 837 | ||
| 838 | if (reset) | |
| 839 | { | |
| 840 | reset_thread_duration_[num] = num_threads; | |
| 841 | reset_thread_duration_times_[num] = exec_total; | |
| 842 | } | |
| 843 | } | |
| 844 | else | |
| 845 | { | |
| 846 | exec_total = std::accumulate(exec_times_.begin(), | |
| 847 | exec_times_.end(), std::uint64_t(0)); | |
| 848 | num_threads = std::accumulate(executed_threads_.begin(), | |
| 849 | executed_threads_.end(), std::int64_t(0)); | |
| 850 | ||
| 851 | reset_exec_total = std::accumulate( | |
| 852 | reset_thread_duration_times_.begin(), | |
| 853 | reset_thread_duration_times_.end(), | |
| 854 | std::uint64_t(0)); | |
| 855 | reset_num_threads = std::accumulate( | |
| 856 | reset_thread_duration_.begin(), | |
| 857 | reset_thread_duration_.end(), | |
| 858 | std::int64_t(0)); | |
| 859 | ||
| 860 | if (reset) | |
| 861 | { | |
| 862 | std::copy(exec_times_.begin(), exec_times_.end(), | |
| 863 | reset_thread_duration_times_.begin()); | |
| 864 | std::copy(executed_threads_.begin(), | |
| 865 | executed_threads_.end(), | |
| 866 | reset_thread_duration_.begin()); | |
| 867 | } | |
| 868 | } | |
| 869 | ||
| 870 | HPX_ASSERT(exec_total >= reset_exec_total); | |
| 871 | HPX_ASSERT(num_threads >= reset_num_threads); | |
| 872 | ||
| 873 | exec_total -= reset_exec_total; | |
| 874 | num_threads -= reset_num_threads; | |
| 875 | ||
| 876 | return std::uint64_t( | |
| 877 | (double(exec_total) * timestamp_scale_) / double(num_threads) | |
| 878 | ); | |
| 879 | } | |
| 880 | ||
| 881 | template <typename Scheduler> | |
| 882 | std::int64_t thread_pool<Scheduler>:: | |
| 883 | get_thread_phase_overhead(std::size_t num, bool reset) | |
| 884 | { | |
| 885 | std::uint64_t exec_total = 0; | |
| 886 | std::uint64_t tfunc_total = 0; | |
| 887 | std::int64_t num_phases = 0; | |
| 888 | ||
| 889 | std::uint64_t reset_exec_total = 0; | |
| 890 | std::uint64_t reset_tfunc_total = 0; | |
| 891 | std::int64_t reset_num_phases = 0; | |
| 892 | ||
| 893 | if (num != std::size_t(-1)) | |
| 894 | { | |
| 895 | exec_total = exec_times_[num]; | |
| 896 | tfunc_total = tfunc_times_[num]; | |
| 897 | num_phases = executed_thread_phases_[num]; | |
| 898 | ||
| 899 | reset_exec_total = reset_thread_phase_overhead_times_[num]; | |
| 900 | reset_tfunc_total = reset_thread_phase_overhead_times_total_[num]; | |
| 901 | reset_num_phases = reset_thread_phase_overhead_[num]; | |
| 902 | ||
| 903 | if (reset) | |
| 904 | { | |
| 905 | reset_thread_phase_overhead_times_[num] = exec_total; | |
| 906 | reset_thread_phase_overhead_times_total_[num] = tfunc_total; | |
| 907 | reset_thread_phase_overhead_[num] = num_phases; | |
| 908 | } | |
| 909 | } | |
| 910 | else | |
| 911 | { | |
| 912 | exec_total = std::accumulate(exec_times_.begin(), | |
| 913 | exec_times_.end(), std::uint64_t(0)); | |
| 914 | tfunc_total = std::accumulate(tfunc_times_.begin(), | |
| 915 | tfunc_times_.end(), std::uint64_t(0)); | |
| 916 | num_phases = std::accumulate( | |
| 917 | executed_thread_phases_.begin(), | |
| 918 | executed_thread_phases_.end(), std::int64_t(0)); | |
| 919 | ||
| 920 | reset_exec_total = std::accumulate( | |
| 921 | reset_thread_phase_overhead_times_.begin(), | |
| 922 | reset_thread_phase_overhead_times_.end(), std::uint64_t(0)); | |
| 923 | reset_tfunc_total = std::accumulate( | |
| 924 | reset_thread_phase_overhead_times_total_.begin(), | |
| 925 | reset_thread_phase_overhead_times_total_.end(), | |
| 926 | std::uint64_t(0)); | |
| 927 | reset_num_phases = std::accumulate( | |
| 928 | reset_thread_phase_overhead_.begin(), | |
| 929 | reset_thread_phase_overhead_.end(), std::int64_t(0)); | |
| 930 | ||
| 931 | if (reset) | |
| 932 | { | |
| 933 | std::copy(exec_times_.begin(), exec_times_.end(), | |
| 934 | reset_thread_phase_overhead_times_.begin()); | |
| 935 | std::copy(tfunc_times_.begin(), tfunc_times_.end(), | |
| 936 | reset_thread_phase_overhead_times_total_.begin()); | |
| 937 | std::copy(executed_thread_phases_.begin(), | |
| 938 | executed_thread_phases_.end(), | |
| 939 | reset_thread_phase_overhead_.begin()); | |
| 940 | } | |
| 941 | } | |
| 942 | ||
| 943 | HPX_ASSERT(exec_total >= reset_exec_total); | |
| 944 | HPX_ASSERT(tfunc_total >= reset_tfunc_total); | |
| 945 | HPX_ASSERT(num_phases >= reset_num_phases); | |
| 946 | ||
| 947 | exec_total -= reset_exec_total; | |
| 948 | tfunc_total -= reset_tfunc_total; | |
| 949 | num_phases -= reset_num_phases; | |
| 950 | ||
| 951 | if (num_phases == 0) // avoid division by zero | |
| 952 | return 0; | |
| 953 | ||
| 954 | HPX_ASSERT(tfunc_total >= exec_total); | |
| 955 | ||
| 956 | return std::uint64_t( | |
| 957 | double((tfunc_total - exec_total) * timestamp_scale_) / | |
| 958 | double(num_phases) | |
| 959 | ); | |
| 960 | } | |
| 961 | ||
| 962 | template <typename Scheduler> | |
| 963 | std::int64_t thread_pool<Scheduler>:: | |
| 964 | get_thread_overhead(std::size_t num, bool reset) | |
| 965 | { | |
| 966 | std::uint64_t exec_total = 0; | |
| 967 | std::uint64_t tfunc_total = 0; | |
| 968 | std::int64_t num_threads = 0; | |
| 969 | ||
| 970 | std::uint64_t reset_exec_total = 0; | |
| 971 | std::uint64_t reset_tfunc_total = 0; | |
| 972 | std::int64_t reset_num_threads = 0; | |
| 973 | ||
| 974 | if (num != std::size_t(-1)) | |
| 975 | { | |
| 976 | exec_total = exec_times_[num]; | |
| 977 | tfunc_total = tfunc_times_[num]; | |
| 978 | num_threads = executed_threads_[num]; | |
| 979 | ||
| 980 | reset_exec_total = reset_thread_overhead_times_[num]; | |
| 981 | reset_tfunc_total = reset_thread_overhead_times_total_[num]; | |
| 982 | reset_num_threads = reset_thread_overhead_[num]; | |
| 983 | ||
| 984 | if (reset) | |
| 985 | { | |
| 986 | reset_thread_overhead_times_[num] = exec_total; | |
| 987 | reset_thread_overhead_times_total_[num] = tfunc_total; | |
| 988 | reset_thread_overhead_[num] = num_threads; | |
| 989 | } | |
| 990 | } | |
| 991 | else | |
| 992 | { | |
| 993 | exec_total = std::accumulate(exec_times_.begin(), | |
| 994 | exec_times_.end(), std::uint64_t(0)); | |
| 995 | tfunc_total = std::accumulate(tfunc_times_.begin(), | |
| 996 | tfunc_times_.end(), std::uint64_t(0)); | |
| 997 | num_threads = std::accumulate(executed_threads_.begin(), | |
| 998 | executed_threads_.end(), std::int64_t(0)); | |
| 999 | ||
| 1000 | reset_exec_total = std::accumulate( | |
| 1001 | reset_thread_overhead_times_.begin(), | |
| 1002 | reset_thread_overhead_times_.end(), std::uint64_t(0)); | |
| 1003 | reset_tfunc_total = std::accumulate( | |
| 1004 | reset_thread_overhead_times_total_.begin(), | |
| 1005 | reset_thread_overhead_times_total_.end(), | |
| 1006 | std::uint64_t(0)); | |
| 1007 | reset_num_threads = std::accumulate( | |
| 1008 | reset_thread_overhead_.begin(), | |
| 1009 | reset_thread_overhead_.end(), std::int64_t(0)); | |
| 1010 | ||
| 1011 | if (reset) | |
| 1012 | { | |
| 1013 | std::copy(exec_times_.begin(), exec_times_.end(), | |
| 1014 | reset_thread_overhead_times_.begin()); | |
| 1015 | std::copy(tfunc_times_.begin(), tfunc_times_.end(), | |
| 1016 | reset_thread_overhead_times_total_.begin()); | |
| 1017 | std::copy(executed_threads_.begin(), | |
| 1018 | executed_threads_.end(), | |
| 1019 | reset_thread_overhead_.begin()); | |
| 1020 | } | |
| 1021 | } | |
| 1022 | ||
| 1023 | HPX_ASSERT(exec_total >= reset_exec_total); | |
| 1024 | HPX_ASSERT(tfunc_total >= reset_tfunc_total); | |
| 1025 | HPX_ASSERT(num_threads >= reset_num_threads); | |
| 1026 | ||
| 1027 | exec_total -= reset_exec_total; | |
| 1028 | tfunc_total -= reset_tfunc_total; | |
| 1029 | num_threads -= reset_num_threads; | |
| 1030 | ||
| 1031 | if (num_threads == 0) // avoid division by zero | |
| 1032 | return 0; | |
| 1033 | ||
| 1034 | HPX_ASSERT(tfunc_total >= exec_total); | |
| 1035 | ||
| 1036 | return std::uint64_t( | |
| 1037 | double((tfunc_total - exec_total) * timestamp_scale_) / | |
| 1038 | double(num_threads) | |
| 1039 | ); | |
| 1040 | } | |
| 1041 | ||
| 1042 | template <typename Scheduler> | |
| 1043 | std::int64_t thread_pool<Scheduler>:: | |
| 1044 | get_cumulative_thread_duration(std::size_t num, bool reset) | |
| 1045 | { | |
| 1046 | std::uint64_t exec_total = 0ul; | |
| 1047 | std::uint64_t reset_exec_total = 0ul; | |
| 1048 | ||
| 1049 | if (num != std::size_t(-1)) | |
| 1050 | { | |
| 1051 | exec_total = exec_times_[num]; | |
| 1052 | reset_exec_total = reset_cumulative_thread_duration_[num]; | |
| 1053 | ||
| 1054 | if (reset) | |
| 1055 | reset_cumulative_thread_duration_[num] = exec_total; | |
| 1056 | } | |
| 1057 | else | |
| 1058 | { | |
| 1059 | exec_total = std::accumulate(exec_times_.begin(), | |
| 1060 | exec_times_.end(), std::uint64_t(0)); | |
| 1061 | reset_exec_total = std::accumulate( | |
| 1062 | reset_cumulative_thread_duration_.begin(), | |
| 1063 | reset_cumulative_thread_duration_.end(), | |
| 1064 | std::uint64_t(0)); | |
| 1065 | ||
| 1066 | if (reset) | |
| 1067 | { | |
| 1068 | std::copy(exec_times_.begin(), exec_times_.end(), | |
| 1069 | reset_cumulative_thread_duration_.begin()); | |
| 1070 | } | |
| 1071 | } | |
| 1072 | ||
| 1073 | HPX_ASSERT(exec_total >= reset_exec_total); | |
| 1074 | ||
| 1075 | exec_total -= reset_exec_total; | |
| 1076 | ||
| 1077 | return std::uint64_t(double(exec_total) * timestamp_scale_); | |
| 1078 | } | |
| 1079 | ||
| 1080 | template <typename Scheduler> | |
| 1081 | std::int64_t thread_pool<Scheduler>:: | |
| 1082 | get_cumulative_thread_overhead(std::size_t num, bool reset) | |
| 1083 | { | |
| 1084 | std::uint64_t exec_total = 0ul; | |
| 1085 | std::uint64_t reset_exec_total = 0ul; | |
| 1086 | std::uint64_t tfunc_total = 0ul; | |
| 1087 | std::uint64_t reset_tfunc_total = 0ul; | |
| 1088 | ||
| 1089 | if (num != std::size_t(-1)) | |
| 1090 | { | |
| 1091 | exec_total = exec_times_[num]; | |
| 1092 | tfunc_total = tfunc_times_[num]; | |
| 1093 | ||
| 1094 | reset_exec_total = reset_cumulative_thread_overhead_[num]; | |
| 1095 | reset_tfunc_total = reset_cumulative_thread_overhead_total_[num]; | |
| 1096 | ||
| 1097 | if (reset) | |
| 1098 | { | |
| 1099 | reset_cumulative_thread_overhead_[num] = exec_total; | |
| 1100 | reset_cumulative_thread_overhead_total_[num] = tfunc_total; | |
| 1101 | } | |
| 1102 | } | |
| 1103 | else | |
| 1104 | { | |
| 1105 | exec_total = std::accumulate(exec_times_.begin(), | |
| 1106 | exec_times_.end(), std::uint64_t(0)); | |
| 1107 | reset_exec_total = std::accumulate( | |
| 1108 | reset_cumulative_thread_overhead_.begin(), | |
| 1109 | reset_cumulative_thread_overhead_.end(), | |
| 1110 | std::uint64_t(0)); | |
| 1111 | ||
| 1112 | tfunc_total = std::accumulate(tfunc_times_.begin(), | |
| 1113 | tfunc_times_.end(), std::uint64_t(0)); | |
| 1114 | reset_tfunc_total = std::accumulate( | |
| 1115 | reset_cumulative_thread_overhead_total_.begin(), | |
| 1116 | reset_cumulative_thread_overhead_total_.end(), | |
| 1117 | std::uint64_t(0)); | |
| 1118 | ||
| 1119 | if (reset) | |
| 1120 | { | |
| 1121 | std::copy(exec_times_.begin(), exec_times_.end(), | |
| 1122 | reset_cumulative_thread_overhead_.begin()); | |
| 1123 | std::copy(tfunc_times_.begin(), tfunc_times_.end(), | |
| 1124 | reset_cumulative_thread_overhead_total_.begin()); | |
| 1125 | } | |
| 1126 | } | |
| 1127 | ||
| 1128 | HPX_ASSERT(exec_total >= reset_exec_total); | |
| 1129 | HPX_ASSERT(tfunc_total >= reset_tfunc_total); | |
| 1130 | ||
| 1131 | exec_total -= reset_exec_total; | |
| 1132 | tfunc_total -= reset_tfunc_total; | |
| 1133 | ||
| 1134 | return std::uint64_t( | |
| 1135 | (double(tfunc_total) - double(exec_total)) * timestamp_scale_ | |
| 1136 | ); | |
| 1137 | } | |
| 1138 | #endif | |
| 1139 | #endif | |
| 1140 | ||
| 1141 | template <typename Scheduler> | |
| 1142 | std::int64_t thread_pool<Scheduler>:: | |
| 1143 | get_cumulative_duration(std::size_t num, bool reset) | |
| 1144 | { | |
| 1145 | std::uint64_t tfunc_total = 0ul; | |
| 1146 | std::uint64_t reset_tfunc_total = 0ul; | |
| 1147 | ||
| 1148 | if (num != std::size_t(-1)) | |
| 1149 | { | |
| 1150 | tfunc_total = tfunc_times_[num]; | |
| 1151 | reset_tfunc_total = reset_tfunc_times_[num]; | |
| 1152 | ||
| 1153 | if (reset) | |
| 1154 | reset_tfunc_times_[num] = tfunc_total; | |
| 1155 | } | |
| 1156 | else | |
| 1157 | { | |
| 1158 | tfunc_total = std::accumulate(tfunc_times_.begin(), | |
| 1159 | tfunc_times_.end(), std::uint64_t(0)); | |
| 1160 | reset_tfunc_total = std::accumulate( | |
| 1161 | reset_tfunc_times_.begin(), reset_tfunc_times_.end(), | |
| 1162 | std::uint64_t(0)); | |
| 1163 | ||
| 1164 | if (reset) | |
| 1165 | { | |
| 1166 | std::copy(tfunc_times_.begin(), tfunc_times_.end(), | |
| 1167 | reset_tfunc_times_.begin()); | |
| 1168 | } | |
| 1169 | } | |
| 1170 | ||
| 1171 | HPX_ASSERT(tfunc_total >= reset_tfunc_total); | |
| 1172 | ||
| 1173 | tfunc_total -= reset_tfunc_total; | |
| 1174 | ||
| 1175 | return std::uint64_t(double(tfunc_total) * timestamp_scale_); | |
| 1176 | } | |
| 1177 | ||
| 1178 | #if defined(HPX_HAVE_THREAD_IDLE_RATES) | |
| 1179 | /////////////////////////////////////////////////////////////////////////// | |
| 1180 | template <typename Scheduler> | |
| 1181 | std::int64_t thread_pool<Scheduler>::avg_idle_rate(bool reset) | |
| 1182 | { | |
| 1183 | std::uint64_t exec_total = std::accumulate(exec_times_.begin(), | |
| 1184 | exec_times_.end(), std::uint64_t(0)); | |
| 1185 | std::uint64_t tfunc_total = std::accumulate(tfunc_times_.begin(), | |
| 1186 | tfunc_times_.end(), std::uint64_t(0)); | |
| 1187 | std::uint64_t reset_exec_total = std::accumulate( | |
| 1188 | reset_idle_rate_time_.begin(), | |
| 1189 | reset_idle_rate_time_.end(), std::uint64_t(0)); | |
| 1190 | std::uint64_t reset_tfunc_total = std::accumulate( | |
| 1191 | reset_idle_rate_time_total_.begin(), | |
| 1192 | reset_idle_rate_time_total_.end(), std::uint64_t(0)); | |
| 1193 | ||
| 1194 | if (reset) | |
| 1195 | { | |
| 1196 | std::copy(exec_times_.begin(), exec_times_.end(), | |
| 1197 | reset_idle_rate_time_.begin()); | |
| 1198 | std::copy(tfunc_times_.begin(), tfunc_times_.end(), | |
| 1199 | reset_idle_rate_time_total_.begin()); | |
| 1200 | } | |
| 1201 | ||
| 1202 | HPX_ASSERT(exec_total >= reset_exec_total); | |
| 1203 | HPX_ASSERT(tfunc_total >= reset_tfunc_total); | |
| 1204 | ||
| 1205 | exec_total -= reset_exec_total; | |
| 1206 | tfunc_total -= reset_tfunc_total; | |
| 1207 | ||
| 1208 | if (tfunc_total == 0) // avoid division by zero | |
| 1209 | return 10000LL; | |
| 1210 | ||
| 1211 | HPX_ASSERT(tfunc_total >= exec_total); | |
| 1212 | ||
| 1213 | double const percent = 1. - (double(exec_total) / double(tfunc_total)); | |
| 1214 | return std::int64_t(10000. * percent); // 0.01 percent | |
| 1215 | } | |
| 1216 | ||
| 1217 | template <typename Scheduler> | |
| 1218 | std::int64_t thread_pool<Scheduler>::avg_idle_rate( | |
| 1219 | std::size_t num_thread, bool reset) | |
| 1220 | { | |
| 1221 | std::uint64_t exec_time = exec_times_[num_thread]; | |
| 1222 | std::uint64_t tfunc_time = tfunc_times_[num_thread]; | |
| 1223 | std::uint64_t reset_exec_time = reset_idle_rate_time_[num_thread]; | |
| 1224 | std::uint64_t reset_tfunc_time = reset_idle_rate_time_total_[num_thread]; | |
| 1225 | ||
| 1226 | if (reset) | |
| 1227 | { | |
| 1228 | reset_idle_rate_time_[num_thread] = exec_time; | |
| 1229 | reset_idle_rate_time_total_[num_thread] = tfunc_time; | |
| 1230 | } | |
| 1231 | ||
| 1232 | HPX_ASSERT(exec_time >= reset_exec_time); | |
| 1233 | HPX_ASSERT(tfunc_time >= reset_tfunc_time); | |
| 1234 | ||
| 1235 | exec_time -= reset_exec_time; | |
| 1236 | tfunc_time -= reset_tfunc_time; | |
| 1237 | ||
| 1238 | if (tfunc_time == 0) // avoid division by zero | |
| 1239 | return 10000LL; | |
| 1240 | ||
| 1241 | HPX_ASSERT(tfunc_time > exec_time); | |
| 1242 | ||
| 1243 | double const percent = 1. - (double(exec_time) / double(tfunc_time)); | |
| 1244 | return std::int64_t(10000. * percent); // 0.01 percent | |
| 1245 | } | |
| 1246 | ||
| 1247 | #if defined(HPX_HAVE_THREAD_CREATION_AND_CLEANUP_RATES) | |
| 1248 | template <typename Scheduler> | |
| 1249 | std::int64_t thread_pool<Scheduler>::avg_creation_idle_rate(bool reset) | |
| 1250 | { | |
| 1251 | double const creation_total = | |
| 1252 | static_cast<double>(sched_.get_creation_time(reset)); | |
| 1253 | ||
| 1254 | std::uint64_t exec_total = std::accumulate(exec_times_.begin(), | |
| 1255 | exec_times_.end(), std::uint64_t(0)); | |
| 1256 | std::uint64_t tfunc_total = std::accumulate(tfunc_times_.begin(), | |
| 1257 | tfunc_times_.end(), std::uint64_t(0)); | |
| 1258 | std::uint64_t reset_exec_total = std::accumulate( | |
| 1259 | reset_creation_idle_rate_time_.begin(), | |
| 1260 | reset_creation_idle_rate_time_.end(), std::uint64_t(0)); | |
| 1261 | std::uint64_t reset_tfunc_total = std::accumulate( | |
| 1262 | reset_creation_idle_rate_time_total_.begin(), | |
| 1263 | reset_creation_idle_rate_time_total_.end(), std::uint64_t(0)); | |
| 1264 | ||
| 1265 | if (reset) | |
| 1266 | { | |
| 1267 | std::copy(exec_times_.begin(), exec_times_.end(), | |
| 1268 | reset_creation_idle_rate_time_.begin()); | |
| 1269 | std::copy(tfunc_times_.begin(), tfunc_times_.end(), | |
| 1270 | reset_creation_idle_rate_time_.begin()); | |
| 1271 | } | |
| 1272 | ||
| 1273 | HPX_ASSERT(exec_total >= reset_exec_total); | |
| 1274 | HPX_ASSERT(tfunc_total >= reset_tfunc_total); | |
| 1275 | ||
| 1276 | exec_total -= reset_exec_total; | |
| 1277 | tfunc_total -= reset_tfunc_total; | |
| 1278 | ||
| 1279 | if (tfunc_total == exec_total) // avoid division by zero | |
| 1280 | return 10000LL; | |
| 1281 | ||
| 1282 | HPX_ASSERT(tfunc_total > exec_total); | |
| 1283 | ||
| 1284 | double const percent = (creation_total / double(tfunc_total - exec_total)); | |
| 1285 | return std::int64_t(10000. * percent); // 0.01 percent | |
| 1286 | } | |
| 1287 | ||
| 1288 | template <typename Scheduler> | |
| 1289 | std::int64_t thread_pool<Scheduler>::avg_cleanup_idle_rate(bool reset) | |
| 1290 | { | |
| 1291 | double const cleanup_total = | |
| 1292 | static_cast<double>(sched_.get_cleanup_time(reset)); | |
| 1293 | ||
| 1294 | std::uint64_t exec_total = std::accumulate(exec_times_.begin(), | |
| 1295 | exec_times_.end(), std::uint64_t(0)); | |
| 1296 | std::uint64_t tfunc_total = std::accumulate(tfunc_times_.begin(), | |
| 1297 | tfunc_times_.end(), std::uint64_t(0)); | |
| 1298 | std::uint64_t reset_exec_total = std::accumulate( | |
| 1299 | reset_cleanup_idle_rate_time_.begin(), | |
| 1300 | reset_cleanup_idle_rate_time_.end(), std::uint64_t(0)); | |
| 1301 | std::uint64_t reset_tfunc_total = std::accumulate( | |
| 1302 | reset_cleanup_idle_rate_time_total_.begin(), | |
| 1303 | reset_cleanup_idle_rate_time_total_.end(), std::uint64_t(0)); | |
| 1304 | ||
| 1305 | if (reset) | |
| 1306 | { | |
| 1307 | std::copy(exec_times_.begin(), exec_times_.end(), | |
| 1308 | reset_cleanup_idle_rate_time_.begin()); | |
| 1309 | std::copy(tfunc_times_.begin(), tfunc_times_.end(), | |
| 1310 | reset_cleanup_idle_rate_time_.begin()); | |
| 1311 | } | |
| 1312 | ||
| 1313 | HPX_ASSERT(exec_total >= reset_exec_total); | |
| 1314 | HPX_ASSERT(tfunc_total >= reset_tfunc_total); | |
| 1315 | ||
| 1316 | exec_total -= reset_exec_total; | |
| 1317 | tfunc_total -= reset_tfunc_total; | |
| 1318 | ||
| 1319 | if (tfunc_total == exec_total) // avoid division by zero | |
| 1320 | return 10000LL; | |
| 1321 | ||
| 1322 | HPX_ASSERT(tfunc_total > exec_total); | |
| 1323 | ||
| 1324 | double const percent = (cleanup_total / double(tfunc_total - exec_total)); | |
| 1325 | return std::int64_t(10000. * percent); // 0.01 percent | |
| 1326 | } | |
| 1327 | #endif | |
| 1328 | #endif | |
| 1329 | ||
| 1330 | template <typename Scheduler> | |
| 1331 | std::int64_t thread_pool<Scheduler>:: | |
| 1332 | get_queue_length(std::size_t num_thread) const | |
| 1333 | { | |
| 1334 | return sched_.Scheduler::get_queue_length(num_thread); | |
| 1335 | } | |
| 1336 | ||
| 1337 | #ifdef HPX_HAVE_THREAD_QUEUE_WAITTIME | |
| 1338 | template <typename Scheduler> | |
| 1339 | std::int64_t thread_pool<Scheduler>:: | |
| 1340 | get_average_thread_wait_time(std::size_t num_thread) const | |
| 1341 | { | |
| 1342 | return sched_.Scheduler::get_average_thread_wait_time(num_thread); | |
| 1343 | } | |
| 1344 | ||
| 1345 | template <typename Scheduler> | |
| 1346 | std::int64_t thread_pool<Scheduler>:: | |
| 1347 | get_average_task_wait_time(std::size_t num_thread) const | |
| 1348 | { | |
| 1349 | return sched_.Scheduler::get_average_task_wait_time(num_thread); | |
| 1350 | } | |
| 1351 | #endif | |
| 1352 | ||
| 1353 | #ifdef HPX_HAVE_THREAD_STEALING_COUNTS | |
| 1354 | template <typename Scheduler> | |
| 1355 | std::int64_t thread_pool<Scheduler>:: | |
| 1356 | get_num_pending_misses(std::size_t num, bool reset) | |
| 1357 | { | |
| 1358 | return sched_.Scheduler::get_num_pending_misses(num, reset); | |
| 1359 | } | |
| 1360 | ||
| 1361 | template <typename Scheduler> | |
| 1362 | std::int64_t thread_pool<Scheduler>:: | |
| 1363 | get_num_pending_accesses(std::size_t num, bool reset) | |
| 1364 | { | |
| 1365 | return sched_.Scheduler::get_num_pending_accesses(num, reset); | |
| 1366 | } | |
| 1367 | ||
| 1368 | template <typename Scheduler> | |
| 1369 | std::int64_t thread_pool<Scheduler>:: | |
| 1370 | get_num_stolen_from_pending(std::size_t num, bool reset) | |
| 1371 | { | |
| 1372 | return sched_.Scheduler::get_num_stolen_from_pending(num, reset); | |
| 1373 | } | |
| 1374 | ||
| 1375 | template <typename Scheduler> | |
| 1376 | std::int64_t thread_pool<Scheduler>:: | |
| 1377 | get_num_stolen_to_pending(std::size_t num, bool reset) | |
| 1378 | { | |
| 1379 | return sched_.Scheduler::get_num_stolen_to_pending(num, reset); | |
| 1380 | } | |
| 1381 | ||
| 1382 | template <typename Scheduler> | |
| 1383 | std::int64_t thread_pool<Scheduler>:: | |
| 1384 | get_num_stolen_from_staged(std::size_t num, bool reset) | |
| 1385 | { | |
| 1386 | return sched_.Scheduler::get_num_stolen_from_staged(num, reset); | |
| 1387 | } | |
| 1388 | ||
| 1389 | template <typename Scheduler> | |
| 1390 | std::int64_t thread_pool<Scheduler>:: | |
| 1391 | get_num_stolen_to_staged(std::size_t num, bool reset) | |
| 1392 | { | |
| 1393 | return sched_.Scheduler::get_num_stolen_to_staged(num, reset); | |
| 1394 | } | |
| 1395 | #endif | |
| 1396 | ||
| 1397 | }}} | |
| 1398 | ||
| 1399 | /////////////////////////////////////////////////////////////////////////////// | |
| 1400 | /// explicit template instantiation for the thread manager of our choice | |
| 1401 | #if defined(HPX_HAVE_THROTTLE_SCHEDULER) && defined(HPX_HAVE_APEX) | |
| 1402 | #include <hpx/runtime/threads/policies/throttle_queue_scheduler.hpp> | |
| 1403 | template class HPX_EXPORT hpx::threads::detail::thread_pool< | |
| 1404 | hpx::threads::policies::throttle_queue_scheduler<> >; | |
| 1405 | #endif | |
| 1406 | ||
| 1407 | #if defined(HPX_HAVE_LOCAL_SCHEDULER) | |
| 1408 | #include <hpx/runtime/threads/policies/local_queue_scheduler.hpp> | |
| 1409 | template class HPX_EXPORT hpx::threads::detail::thread_pool< | |
| 1410 | hpx::threads::policies::local_queue_scheduler<> >; | |
| 1411 | #endif | |
| 1412 | ||
| 1413 | #if defined(HPX_HAVE_STATIC_SCHEDULER) | |
| 1414 | #include <hpx/runtime/threads/policies/static_queue_scheduler.hpp> | |
| 1415 | template class HPX_EXPORT hpx::threads::detail::thread_pool< | |
| 1416 | hpx::threads::policies::static_queue_scheduler<> >; | |
| 1417 | #endif | |
| 1418 | ||
| 1419 | #if defined(HPX_HAVE_STATIC_PRIORITY_SCHEDULER) | |
| 1420 | #include <hpx/runtime/threads/policies/static_priority_queue_scheduler.hpp> | |
| 1421 | template class HPX_EXPORT hpx::threads::detail::thread_pool< | |
| 1422 | hpx::threads::policies::static_priority_queue_scheduler<> >; | |
| 1423 | #endif | |
| 1424 | ||
| 1425 | #include <hpx/runtime/threads/policies/local_priority_queue_scheduler.hpp> | |
| 1426 | template class HPX_EXPORT hpx::threads::detail::thread_pool< | |
| 1427 | hpx::threads::policies::local_priority_queue_scheduler<> >; | |
| 1428 | ||
| 1429 | #if defined(HPX_HAVE_ABP_SCHEDULER) | |
| 1430 | template class HPX_EXPORT hpx::threads::detail::thread_pool< | |
| 1431 | hpx::threads::policies::abp_fifo_priority_queue_scheduler>; | |
| 1432 | #endif | |
| 1433 | ||
| 1434 | #if defined(HPX_HAVE_HIERARCHY_SCHEDULER) | |
| 1435 | #include <hpx/runtime/threads/policies/hierarchy_scheduler.hpp> | |
| 1436 | template class HPX_EXPORT hpx::threads::detail::thread_pool< | |
| 1437 | hpx::threads::policies::hierarchy_scheduler<> >; | |
| 1438 | #endif | |
| 1439 | ||
| 1440 | #if defined(HPX_HAVE_PERIODIC_PRIORITY_SCHEDULER) | |
| 1441 | #include <hpx/runtime/threads/policies/periodic_priority_queue_scheduler.hpp> | |
| 1442 | template class HPX_EXPORT hpx::threads::detail::thread_pool< | |
| 1443 | hpx::threads::policies::periodic_priority_queue_scheduler<> >; | |
| 1444 | #endif | |
| 1445 | ||
| 1446 |
Copyright (c) 2006-2012 Rogue Wave Software, Inc. All Rights Reserved.
Patents pending.