[b2a37b0] | 1 | #include "relaxed_list.hpp" |
---|
| 2 | |
---|
| 3 | #include <array> |
---|
[c921712] | 4 | #include <iomanip> |
---|
[b2a37b0] | 5 | #include <iostream> |
---|
| 6 | #include <locale> |
---|
| 7 | #include <string> |
---|
| 8 | #include <thread> |
---|
| 9 | #include <vector> |
---|
| 10 | |
---|
[9421f3d8] | 11 | #include <getopt.h> |
---|
[b2a37b0] | 12 | #include <unistd.h> |
---|
| 13 | #include <sys/sysinfo.h> |
---|
| 14 | |
---|
| 15 | #include "utils.hpp" |
---|
| 16 | |
---|
[50aeb6f] | 17 | struct __attribute__((aligned(64))) Node { |
---|
[b2a37b0] | 18 | static std::atomic_size_t creates; |
---|
| 19 | static std::atomic_size_t destroys; |
---|
| 20 | |
---|
| 21 | _LinksFields_t<Node> _links; |
---|
| 22 | |
---|
| 23 | int value; |
---|
[807a632] | 24 | int id; |
---|
[b2a37b0] | 25 | |
---|
[9421f3d8] | 26 | Node() { creates++; } |
---|
| 27 | Node(int value): value(value) { creates++; } |
---|
| 28 | ~Node() { destroys++; } |
---|
[b2a37b0] | 29 | }; |
---|
| 30 | |
---|
| 31 | std::atomic_size_t Node::creates = { 0 }; |
---|
| 32 | std::atomic_size_t Node::destroys = { 0 }; |
---|
| 33 | |
---|
| 34 | bool enable_stats = false; |
---|
| 35 | |
---|
[9421f3d8] | 36 | template<> |
---|
| 37 | thread_local relaxed_list<Node>::TLS relaxed_list<Node>::tls = {}; |
---|
| 38 | |
---|
| 39 | template<> |
---|
[807a632] | 40 | relaxed_list<Node> * relaxed_list<Node>::head = nullptr; |
---|
| 41 | |
---|
| 42 | #ifndef NO_STATS |
---|
| 43 | template<> |
---|
| 44 | relaxed_list<Node>::GlobalStats relaxed_list<Node>::global_stats = {}; |
---|
| 45 | #endif |
---|
[9421f3d8] | 46 | |
---|
| 47 | // ================================================================================================ |
---|
| 48 | // UTILS |
---|
| 49 | // ================================================================================================ |
---|
| 50 | |
---|
[50aeb6f] | 51 | struct local_stat_t { |
---|
| 52 | size_t in = 0; |
---|
| 53 | size_t out = 0; |
---|
| 54 | size_t empty = 0; |
---|
| 55 | size_t crc_in = 0; |
---|
| 56 | size_t crc_out = 0; |
---|
[807a632] | 57 | size_t valmax = 0; |
---|
| 58 | size_t valmin = 100000000ul; |
---|
[50aeb6f] | 59 | }; |
---|
| 60 | |
---|
[9421f3d8] | 61 | struct global_stat_t { |
---|
| 62 | std::atomic_size_t in = { 0 }; |
---|
| 63 | std::atomic_size_t out = { 0 }; |
---|
| 64 | std::atomic_size_t empty = { 0 }; |
---|
| 65 | std::atomic_size_t crc_in = { 0 }; |
---|
| 66 | std::atomic_size_t crc_out = { 0 }; |
---|
[807a632] | 67 | std::atomic_size_t valmax = { 0 }; |
---|
| 68 | std::atomic_size_t valmin = { 100000000ul }; |
---|
[9421f3d8] | 69 | }; |
---|
| 70 | |
---|
[807a632] | 71 | void atomic_max(std::atomic_size_t & target, size_t value) { |
---|
| 72 | for(;;) { |
---|
| 73 | size_t expect = target.load(std::memory_order_relaxed); |
---|
| 74 | if(value <= expect) return; |
---|
| 75 | bool success = target.compare_exchange_strong(expect, value); |
---|
| 76 | if(success) return; |
---|
| 77 | } |
---|
| 78 | } |
---|
| 79 | |
---|
| 80 | void atomic_min(std::atomic_size_t & target, size_t value) { |
---|
| 81 | for(;;) { |
---|
| 82 | size_t expect = target.load(std::memory_order_relaxed); |
---|
| 83 | if(value >= expect) return; |
---|
| 84 | bool success = target.compare_exchange_strong(expect, value); |
---|
| 85 | if(success) return; |
---|
| 86 | } |
---|
| 87 | } |
---|
| 88 | |
---|
[9421f3d8] | 89 | void tally_stats(global_stat_t & global, local_stat_t & local) { |
---|
[807a632] | 90 | |
---|
[9421f3d8] | 91 | global.in += local.in; |
---|
| 92 | global.out += local.out; |
---|
| 93 | global.empty += local.empty; |
---|
| 94 | |
---|
| 95 | global.crc_in += local.crc_in; |
---|
| 96 | global.crc_out += local.crc_out; |
---|
| 97 | |
---|
[807a632] | 98 | atomic_max(global.valmax, local.valmax); |
---|
| 99 | atomic_min(global.valmin, local.valmin); |
---|
[9421f3d8] | 100 | |
---|
[807a632] | 101 | relaxed_list<Node>::stats_tls_tally(); |
---|
[9421f3d8] | 102 | } |
---|
| 103 | |
---|
| 104 | void waitfor(double & duration, barrier_t & barrier, std::atomic_bool & done) { |
---|
| 105 | std::cout << "Starting" << std::endl; |
---|
| 106 | auto before = Clock::now(); |
---|
| 107 | barrier.wait(0); |
---|
| 108 | |
---|
| 109 | while(true) { |
---|
| 110 | usleep(100000); |
---|
| 111 | auto now = Clock::now(); |
---|
| 112 | duration_t durr = now - before; |
---|
| 113 | if( durr.count() > duration ) { |
---|
| 114 | done = true; |
---|
| 115 | break; |
---|
| 116 | } |
---|
| 117 | std::cout << "\r" << std::setprecision(4) << durr.count(); |
---|
| 118 | std::cout.flush(); |
---|
| 119 | } |
---|
| 120 | |
---|
| 121 | barrier.wait(0); |
---|
| 122 | auto after = Clock::now(); |
---|
| 123 | duration_t durr = after - before; |
---|
| 124 | duration = durr.count(); |
---|
| 125 | std::cout << "\rClosing down" << std::endl; |
---|
| 126 | } |
---|
| 127 | |
---|
[807a632] | 128 | void waitfor(double & duration, barrier_t & barrier, const std::atomic_size_t & count) { |
---|
| 129 | std::cout << "Starting" << std::endl; |
---|
| 130 | auto before = Clock::now(); |
---|
| 131 | barrier.wait(0); |
---|
| 132 | |
---|
| 133 | while(true) { |
---|
| 134 | usleep(100000); |
---|
| 135 | size_t c = count.load(); |
---|
| 136 | if( c == 0 ) { |
---|
| 137 | break; |
---|
| 138 | } |
---|
| 139 | std::cout << "\r" << c; |
---|
| 140 | std::cout.flush(); |
---|
| 141 | } |
---|
| 142 | |
---|
| 143 | barrier.wait(0); |
---|
| 144 | auto after = Clock::now(); |
---|
| 145 | duration_t durr = after - before; |
---|
| 146 | duration = durr.count(); |
---|
| 147 | std::cout << "\rClosing down" << std::endl; |
---|
| 148 | } |
---|
| 149 | |
---|
[9421f3d8] | 150 | void print_stats(double duration, unsigned nthread, global_stat_t & global) { |
---|
| 151 | assert(Node::creates == Node::destroys); |
---|
| 152 | assert(global.crc_in == global.crc_out); |
---|
| 153 | |
---|
| 154 | std::cout << "Done" << std::endl; |
---|
| 155 | |
---|
| 156 | size_t ops = global.in + global.out; |
---|
| 157 | size_t ops_sec = size_t(double(ops) / duration); |
---|
| 158 | size_t ops_thread = ops_sec / nthread; |
---|
| 159 | auto dur_nano = duration_cast<std::nano>(1.0); |
---|
| 160 | |
---|
| 161 | std::cout << "Duration : " << duration << "s\n"; |
---|
| 162 | std::cout << "ns/Op : " << ( dur_nano / ops_thread )<< "\n"; |
---|
| 163 | std::cout << "Ops/sec/thread: " << ops_thread << "\n"; |
---|
| 164 | std::cout << "Ops/sec : " << ops_sec << "\n"; |
---|
| 165 | std::cout << "Total ops : " << ops << "(" << global.in << "i, " << global.out << "o, " << global.empty << "e)\n"; |
---|
[807a632] | 166 | if(global.valmax != 0) { |
---|
| 167 | std::cout << "Max runs : " << global.valmax << "\n"; |
---|
| 168 | std::cout << "Min runs : " << global.valmin << "\n"; |
---|
| 169 | } |
---|
[9421f3d8] | 170 | #ifndef NO_STATS |
---|
[807a632] | 171 | relaxed_list<Node>::stats_print(std::cout); |
---|
[9421f3d8] | 172 | #endif |
---|
| 173 | } |
---|
| 174 | |
---|
[807a632] | 175 | void save_fairness(const int data[], int factor, unsigned nthreads, size_t columns, size_t rows, const std::string & output); |
---|
| 176 | |
---|
[9421f3d8] | 177 | // ================================================================================================ |
---|
| 178 | // EXPERIMENTS |
---|
| 179 | // ================================================================================================ |
---|
| 180 | |
---|
| 181 | // ================================================================================================ |
---|
| 182 | __attribute__((noinline)) void runChurn_body( |
---|
[50aeb6f] | 183 | std::atomic<bool>& done, |
---|
| 184 | Random & rand, |
---|
[9421f3d8] | 185 | Node * my_nodes[], |
---|
| 186 | unsigned nslots, |
---|
[50aeb6f] | 187 | local_stat_t & local, |
---|
| 188 | relaxed_list<Node> & list |
---|
| 189 | ) { |
---|
| 190 | while(__builtin_expect(!done.load(std::memory_order_relaxed), true)) { |
---|
[9421f3d8] | 191 | int idx = rand.next() % nslots; |
---|
[50aeb6f] | 192 | if (auto node = my_nodes[idx]) { |
---|
| 193 | local.crc_in += node->value; |
---|
| 194 | list.push(node); |
---|
| 195 | my_nodes[idx] = nullptr; |
---|
| 196 | local.in++; |
---|
| 197 | } |
---|
| 198 | else if(auto node = list.pop()) { |
---|
| 199 | local.crc_out += node->value; |
---|
| 200 | my_nodes[idx] = node; |
---|
| 201 | local.out++; |
---|
| 202 | } |
---|
| 203 | else { |
---|
| 204 | local.empty++; |
---|
| 205 | } |
---|
| 206 | } |
---|
| 207 | } |
---|
[b2a37b0] | 208 | |
---|
[9421f3d8] | 209 | void runChurn(unsigned nthread, unsigned nqueues, double duration, unsigned nnodes, const unsigned nslots) { |
---|
| 210 | std::cout << "Churn Benchmark" << std::endl; |
---|
| 211 | assert(nnodes <= nslots); |
---|
[b2a37b0] | 212 | // List being tested |
---|
| 213 | |
---|
| 214 | // Barrier for synchronization |
---|
| 215 | barrier_t barrier(nthread + 1); |
---|
| 216 | |
---|
| 217 | // Data to check everything is OK |
---|
[9421f3d8] | 218 | global_stat_t global; |
---|
[b2a37b0] | 219 | |
---|
| 220 | // Flag to signal termination |
---|
| 221 | std::atomic_bool done = { false }; |
---|
| 222 | |
---|
| 223 | // Prep nodes |
---|
[c921712] | 224 | std::cout << "Initializing "; |
---|
| 225 | size_t npushed = 0; |
---|
[807a632] | 226 | relaxed_list<Node> list = { nthread * nqueues }; |
---|
| 227 | { |
---|
| 228 | Node** all_nodes[nthread]; |
---|
| 229 | for(auto & nodes : all_nodes) { |
---|
| 230 | nodes = new __attribute__((aligned(64))) Node*[nslots + 8]; |
---|
| 231 | Random rand(rdtscl()); |
---|
| 232 | for(unsigned i = 0; i < nnodes; i++) { |
---|
| 233 | nodes[i] = new Node(rand.next() % 100); |
---|
| 234 | } |
---|
[9421f3d8] | 235 | |
---|
[807a632] | 236 | for(unsigned i = nnodes; i < nslots; i++) { |
---|
| 237 | nodes[i] = nullptr; |
---|
| 238 | } |
---|
[b2a37b0] | 239 | |
---|
[807a632] | 240 | for(int i = 0; i < 10 && i < (int)nslots; i++) { |
---|
| 241 | int idx = rand.next() % nslots; |
---|
| 242 | if (auto node = nodes[idx]) { |
---|
| 243 | global.crc_in += node->value; |
---|
| 244 | list.push(node); |
---|
| 245 | npushed++; |
---|
| 246 | nodes[idx] = nullptr; |
---|
| 247 | } |
---|
[b2a37b0] | 248 | } |
---|
| 249 | } |
---|
| 250 | |
---|
[807a632] | 251 | std::cout << nnodes << " nodes (" << nslots << " slots)" << std::endl; |
---|
[c921712] | 252 | |
---|
[807a632] | 253 | enable_stats = true; |
---|
[b2a37b0] | 254 | |
---|
[807a632] | 255 | std::thread * threads[nthread]; |
---|
| 256 | unsigned i = 1; |
---|
| 257 | for(auto & t : threads) { |
---|
| 258 | auto & my_nodes = all_nodes[i - 1]; |
---|
| 259 | t = new std::thread([&done, &list, &barrier, &global, &my_nodes, nslots](unsigned tid) { |
---|
| 260 | Random rand(tid + rdtscl()); |
---|
[b2a37b0] | 261 | |
---|
[807a632] | 262 | local_stat_t local; |
---|
[b2a37b0] | 263 | |
---|
[807a632] | 264 | // affinity(tid); |
---|
[b2a37b0] | 265 | |
---|
[807a632] | 266 | barrier.wait(tid); |
---|
[b2a37b0] | 267 | |
---|
[807a632] | 268 | // EXPERIMENT START |
---|
[b2a37b0] | 269 | |
---|
[807a632] | 270 | runChurn_body(done, rand, my_nodes, nslots, local, list); |
---|
[b2a37b0] | 271 | |
---|
[807a632] | 272 | // EXPERIMENT END |
---|
[b2a37b0] | 273 | |
---|
[807a632] | 274 | barrier.wait(tid); |
---|
[b2a37b0] | 275 | |
---|
[807a632] | 276 | tally_stats(global, local); |
---|
[b2a37b0] | 277 | |
---|
[807a632] | 278 | for(unsigned i = 0; i < nslots; i++) { |
---|
| 279 | delete my_nodes[i]; |
---|
| 280 | } |
---|
| 281 | }, i++); |
---|
| 282 | } |
---|
[b2a37b0] | 283 | |
---|
[807a632] | 284 | waitfor(duration, barrier, done); |
---|
[b2a37b0] | 285 | |
---|
[807a632] | 286 | for(auto t : threads) { |
---|
| 287 | t->join(); |
---|
| 288 | delete t; |
---|
| 289 | } |
---|
[b2a37b0] | 290 | |
---|
[807a632] | 291 | enable_stats = false; |
---|
[b2a37b0] | 292 | |
---|
[807a632] | 293 | while(auto node = list.pop()) { |
---|
| 294 | global.crc_out += node->value; |
---|
| 295 | delete node; |
---|
| 296 | } |
---|
[b2a37b0] | 297 | |
---|
[807a632] | 298 | for(auto nodes : all_nodes) { |
---|
| 299 | delete[] nodes; |
---|
| 300 | } |
---|
[9421f3d8] | 301 | } |
---|
[b2a37b0] | 302 | |
---|
[9421f3d8] | 303 | print_stats(duration, nthread, global); |
---|
| 304 | } |
---|
[b2a37b0] | 305 | |
---|
[9421f3d8] | 306 | // ================================================================================================ |
---|
| 307 | __attribute__((noinline)) void runPingPong_body( |
---|
| 308 | std::atomic<bool>& done, |
---|
| 309 | Node initial_nodes[], |
---|
| 310 | unsigned nnodes, |
---|
| 311 | local_stat_t & local, |
---|
| 312 | relaxed_list<Node> & list |
---|
| 313 | ) { |
---|
| 314 | Node * nodes[nnodes]; |
---|
| 315 | { |
---|
| 316 | unsigned i = 0; |
---|
| 317 | for(auto & n : nodes) { |
---|
| 318 | n = &initial_nodes[i++]; |
---|
| 319 | } |
---|
| 320 | } |
---|
[b2a37b0] | 321 | |
---|
[9421f3d8] | 322 | while(__builtin_expect(!done.load(std::memory_order_relaxed), true)) { |
---|
| 323 | |
---|
| 324 | for(Node * & node : nodes) { |
---|
| 325 | local.crc_in += node->value; |
---|
| 326 | list.push(node); |
---|
| 327 | local.in++; |
---|
| 328 | } |
---|
| 329 | |
---|
| 330 | // ----- |
---|
| 331 | |
---|
| 332 | for(Node * & node : nodes) { |
---|
| 333 | node = list.pop(); |
---|
| 334 | assert(node); |
---|
| 335 | local.crc_out += node->value; |
---|
| 336 | local.out++; |
---|
| 337 | } |
---|
| 338 | } |
---|
| 339 | } |
---|
| 340 | |
---|
| 341 | void runPingPong(unsigned nthread, unsigned nqueues, double duration, unsigned nnodes) { |
---|
| 342 | std::cout << "PingPong Benchmark" << std::endl; |
---|
| 343 | |
---|
| 344 | |
---|
| 345 | // Barrier for synchronization |
---|
| 346 | barrier_t barrier(nthread + 1); |
---|
| 347 | |
---|
| 348 | // Data to check everything is OK |
---|
| 349 | global_stat_t global; |
---|
| 350 | |
---|
| 351 | // Flag to signal termination |
---|
| 352 | std::atomic_bool done = { false }; |
---|
| 353 | |
---|
| 354 | std::cout << "Initializing "; |
---|
[807a632] | 355 | // List being tested |
---|
| 356 | relaxed_list<Node> list = { nthread * nqueues }; |
---|
| 357 | { |
---|
| 358 | enable_stats = true; |
---|
[9421f3d8] | 359 | |
---|
[807a632] | 360 | std::thread * threads[nthread]; |
---|
| 361 | unsigned i = 1; |
---|
| 362 | for(auto & t : threads) { |
---|
| 363 | t = new std::thread([&done, &list, &barrier, &global, nnodes](unsigned tid) { |
---|
| 364 | Random rand(tid + rdtscl()); |
---|
[9421f3d8] | 365 | |
---|
[807a632] | 366 | Node nodes[nnodes]; |
---|
| 367 | for(auto & n : nodes) { |
---|
| 368 | n.value = (int)rand.next() % 100; |
---|
| 369 | } |
---|
[9421f3d8] | 370 | |
---|
[807a632] | 371 | local_stat_t local; |
---|
[9421f3d8] | 372 | |
---|
[807a632] | 373 | // affinity(tid); |
---|
[9421f3d8] | 374 | |
---|
[807a632] | 375 | barrier.wait(tid); |
---|
[9421f3d8] | 376 | |
---|
[807a632] | 377 | // EXPERIMENT START |
---|
[9421f3d8] | 378 | |
---|
[807a632] | 379 | runPingPong_body(done, nodes, nnodes, local, list); |
---|
[9421f3d8] | 380 | |
---|
[807a632] | 381 | // EXPERIMENT END |
---|
[9421f3d8] | 382 | |
---|
[807a632] | 383 | barrier.wait(tid); |
---|
[9421f3d8] | 384 | |
---|
[807a632] | 385 | tally_stats(global, local); |
---|
| 386 | }, i++); |
---|
| 387 | } |
---|
| 388 | |
---|
| 389 | waitfor(duration, barrier, done); |
---|
| 390 | |
---|
| 391 | for(auto t : threads) { |
---|
| 392 | t->join(); |
---|
| 393 | delete t; |
---|
| 394 | } |
---|
| 395 | |
---|
| 396 | enable_stats = false; |
---|
[9421f3d8] | 397 | } |
---|
| 398 | |
---|
[807a632] | 399 | print_stats(duration, nthread, global); |
---|
| 400 | } |
---|
[9421f3d8] | 401 | |
---|
[807a632] | 402 | // ================================================================================================ |
---|
| 403 | __attribute__((noinline)) void runFairness_body( |
---|
| 404 | unsigned tid, |
---|
| 405 | size_t width, |
---|
| 406 | size_t length, |
---|
| 407 | int output[], |
---|
| 408 | std::atomic_size_t & count, |
---|
| 409 | Node initial_nodes[], |
---|
| 410 | unsigned nnodes, |
---|
| 411 | local_stat_t & local, |
---|
| 412 | relaxed_list<Node> & list |
---|
| 413 | ) { |
---|
| 414 | Node * nodes[nnodes]; |
---|
| 415 | { |
---|
| 416 | unsigned i = 0; |
---|
| 417 | for(auto & n : nodes) { |
---|
| 418 | n = &initial_nodes[i++]; |
---|
| 419 | } |
---|
[9421f3d8] | 420 | } |
---|
| 421 | |
---|
[807a632] | 422 | while(__builtin_expect(0 != count.load(std::memory_order_relaxed), true)) { |
---|
| 423 | |
---|
| 424 | for(Node * & node : nodes) { |
---|
| 425 | local.crc_in += node->id; |
---|
| 426 | list.push(node); |
---|
| 427 | local.in++; |
---|
| 428 | } |
---|
| 429 | |
---|
| 430 | // ----- |
---|
| 431 | |
---|
| 432 | for(Node * & node : nodes) { |
---|
| 433 | node = list.pop(); |
---|
| 434 | assert(node); |
---|
| 435 | |
---|
| 436 | if (unsigned(node->value) < length) { |
---|
| 437 | size_t idx = (node->value * width) + node->id; |
---|
| 438 | assert(idx < (width * length)); |
---|
| 439 | output[idx] = tid; |
---|
| 440 | } |
---|
| 441 | |
---|
| 442 | node->value++; |
---|
| 443 | if(unsigned(node->value) == length) count--; |
---|
| 444 | |
---|
| 445 | local.crc_out += node->id; |
---|
| 446 | local.out++; |
---|
| 447 | } |
---|
| 448 | } |
---|
| 449 | } |
---|
| 450 | |
---|
| 451 | void runFairness(unsigned nthread, unsigned nqueues, double duration, unsigned nnodes, const std::string & output) { |
---|
| 452 | std::cout << "Fairness Benchmark, outputing to : " << output << std::endl; |
---|
| 453 | |
---|
| 454 | // Barrier for synchronization |
---|
| 455 | barrier_t barrier(nthread + 1); |
---|
| 456 | |
---|
| 457 | // Data to check everything is OK |
---|
| 458 | global_stat_t global; |
---|
| 459 | |
---|
| 460 | std::cout << "Initializing "; |
---|
| 461 | |
---|
| 462 | // Check fairness by creating a png of where the threads ran |
---|
| 463 | size_t width = nthread * nnodes; |
---|
| 464 | size_t length = 100000; |
---|
| 465 | |
---|
| 466 | std::unique_ptr<int[]> data_out { new int[width * length] }; |
---|
| 467 | |
---|
| 468 | // Flag to signal termination |
---|
| 469 | std::atomic_size_t count = width; |
---|
| 470 | |
---|
| 471 | // List being tested |
---|
| 472 | relaxed_list<Node> list = { nthread * nqueues }; |
---|
| 473 | { |
---|
| 474 | enable_stats = true; |
---|
| 475 | |
---|
| 476 | std::thread * threads[nthread]; |
---|
| 477 | unsigned i = 1; |
---|
| 478 | for(auto & t : threads) { |
---|
| 479 | t = new std::thread([&count, &list, &barrier, &global, nnodes, width, length, data_out = data_out.get()](unsigned tid) { |
---|
| 480 | unsigned int start = (tid - 1) * nnodes; |
---|
| 481 | Node nodes[nnodes]; |
---|
| 482 | for(auto & n : nodes) { |
---|
| 483 | n.id = start; |
---|
| 484 | n.value = 0; |
---|
| 485 | start++; |
---|
| 486 | } |
---|
| 487 | |
---|
| 488 | local_stat_t local; |
---|
| 489 | |
---|
| 490 | // affinity(tid); |
---|
| 491 | |
---|
| 492 | barrier.wait(tid); |
---|
| 493 | |
---|
| 494 | // EXPERIMENT START |
---|
| 495 | |
---|
| 496 | runFairness_body(tid, width, length, data_out, count, nodes, nnodes, local, list); |
---|
| 497 | |
---|
| 498 | // EXPERIMENT END |
---|
| 499 | |
---|
| 500 | barrier.wait(tid); |
---|
| 501 | |
---|
| 502 | for(const auto & n : nodes) { |
---|
| 503 | local.valmax = max(local.valmax, size_t(n.value)); |
---|
| 504 | local.valmin = min(local.valmin, size_t(n.value)); |
---|
| 505 | } |
---|
| 506 | |
---|
| 507 | tally_stats(global, local); |
---|
| 508 | }, i++); |
---|
| 509 | } |
---|
| 510 | |
---|
| 511 | waitfor(duration, barrier, count); |
---|
| 512 | |
---|
| 513 | for(auto t : threads) { |
---|
| 514 | t->join(); |
---|
| 515 | delete t; |
---|
| 516 | } |
---|
| 517 | |
---|
| 518 | enable_stats = false; |
---|
| 519 | } |
---|
[9421f3d8] | 520 | |
---|
| 521 | print_stats(duration, nthread, global); |
---|
[807a632] | 522 | |
---|
| 523 | save_fairness(data_out.get(), 100, nthread, width, length, output); |
---|
[b2a37b0] | 524 | } |
---|
| 525 | |
---|
[807a632] | 526 | // ================================================================================================ |
---|
| 527 | |
---|
[9421f3d8] | 528 | bool iequals(const std::string& a, const std::string& b) |
---|
| 529 | { |
---|
| 530 | return std::equal(a.begin(), a.end(), |
---|
| 531 | b.begin(), b.end(), |
---|
| 532 | [](char a, char b) { |
---|
| 533 | return std::tolower(a) == std::tolower(b); |
---|
| 534 | }); |
---|
[b2a37b0] | 535 | } |
---|
| 536 | |
---|
| 537 | int main(int argc, char * argv[]) { |
---|
| 538 | |
---|
| 539 | double duration = 5.0; |
---|
| 540 | unsigned nthreads = 2; |
---|
[9421f3d8] | 541 | unsigned nqueues = 4; |
---|
| 542 | unsigned nnodes = 100; |
---|
| 543 | unsigned nslots = 100; |
---|
[807a632] | 544 | std::string out = "fairness.png"; |
---|
[9421f3d8] | 545 | |
---|
| 546 | enum { |
---|
| 547 | Churn, |
---|
| 548 | PingPong, |
---|
[807a632] | 549 | Fairness, |
---|
[9421f3d8] | 550 | NONE |
---|
| 551 | } benchmark = NONE; |
---|
[b2a37b0] | 552 | |
---|
| 553 | std::cout.imbue(std::locale("")); |
---|
| 554 | |
---|
[9421f3d8] | 555 | for(;;) { |
---|
| 556 | static struct option options[] = { |
---|
| 557 | {"duration", required_argument, 0, 'd'}, |
---|
| 558 | {"nthreads", required_argument, 0, 't'}, |
---|
| 559 | {"nqueues", required_argument, 0, 'q'}, |
---|
| 560 | {"benchmark", required_argument, 0, 'b'}, |
---|
| 561 | {0, 0, 0, 0} |
---|
| 562 | }; |
---|
| 563 | |
---|
| 564 | int idx = 0; |
---|
| 565 | int opt = getopt_long(argc, argv, "d:t:q:b:", options, &idx); |
---|
| 566 | |
---|
| 567 | std::string arg = optarg ? optarg : ""; |
---|
| 568 | size_t len = 0; |
---|
| 569 | switch(opt) { |
---|
| 570 | // Exit Case |
---|
| 571 | case -1: |
---|
| 572 | /* paranoid */ assert(optind <= argc); |
---|
| 573 | switch(benchmark) { |
---|
| 574 | case NONE: |
---|
| 575 | std::cerr << "Must specify a benchmark" << std::endl; |
---|
| 576 | goto usage; |
---|
| 577 | case PingPong: |
---|
| 578 | nnodes = 1; |
---|
| 579 | nslots = 1; |
---|
| 580 | switch(argc - optind) { |
---|
| 581 | case 0: break; |
---|
| 582 | case 1: |
---|
| 583 | try { |
---|
| 584 | arg = optarg = argv[optind]; |
---|
| 585 | nnodes = stoul(optarg, &len); |
---|
| 586 | if(len != arg.size()) { throw std::invalid_argument(""); } |
---|
| 587 | } catch(std::invalid_argument &) { |
---|
| 588 | std::cerr << "Number of nodes must be a positive integer, was " << arg << std::endl; |
---|
| 589 | goto usage; |
---|
| 590 | } |
---|
| 591 | break; |
---|
| 592 | default: |
---|
| 593 | std::cerr << "'PingPong' benchmark doesn't accept more than 2 extra arguments" << std::endl; |
---|
| 594 | goto usage; |
---|
| 595 | } |
---|
| 596 | break; |
---|
| 597 | case Churn: |
---|
| 598 | nnodes = 100; |
---|
| 599 | nslots = 100; |
---|
| 600 | switch(argc - optind) { |
---|
| 601 | case 0: break; |
---|
| 602 | case 1: |
---|
| 603 | try { |
---|
| 604 | arg = optarg = argv[optind]; |
---|
| 605 | nnodes = stoul(optarg, &len); |
---|
| 606 | if(len != arg.size()) { throw std::invalid_argument(""); } |
---|
| 607 | nslots = nnodes; |
---|
| 608 | } catch(std::invalid_argument &) { |
---|
| 609 | std::cerr << "Number of nodes must be a positive integer, was " << arg << std::endl; |
---|
| 610 | goto usage; |
---|
| 611 | } |
---|
| 612 | break; |
---|
| 613 | case 2: |
---|
| 614 | try { |
---|
| 615 | arg = optarg = argv[optind]; |
---|
| 616 | nnodes = stoul(optarg, &len); |
---|
| 617 | if(len != arg.size()) { throw std::invalid_argument(""); } |
---|
| 618 | } catch(std::invalid_argument &) { |
---|
| 619 | std::cerr << "Number of nodes must be a positive integer, was " << arg << std::endl; |
---|
| 620 | goto usage; |
---|
| 621 | } |
---|
| 622 | try { |
---|
| 623 | arg = optarg = argv[optind + 1]; |
---|
| 624 | nslots = stoul(optarg, &len); |
---|
| 625 | if(len != arg.size()) { throw std::invalid_argument(""); } |
---|
| 626 | } catch(std::invalid_argument &) { |
---|
| 627 | std::cerr << "Number of slots must be a positive integer, was " << arg << std::endl; |
---|
| 628 | goto usage; |
---|
| 629 | } |
---|
| 630 | break; |
---|
| 631 | default: |
---|
| 632 | std::cerr << "'Churn' benchmark doesn't accept more than 2 extra arguments" << std::endl; |
---|
| 633 | goto usage; |
---|
| 634 | } |
---|
| 635 | break; |
---|
[807a632] | 636 | case Fairness: |
---|
| 637 | nnodes = 1; |
---|
| 638 | switch(argc - optind) { |
---|
| 639 | case 0: break; |
---|
| 640 | case 1: |
---|
| 641 | arg = optarg = argv[optind]; |
---|
| 642 | out = arg; |
---|
| 643 | break; |
---|
| 644 | default: |
---|
| 645 | std::cerr << "'Churn' benchmark doesn't accept more than 2 extra arguments" << std::endl; |
---|
| 646 | goto usage; |
---|
| 647 | } |
---|
[9421f3d8] | 648 | } |
---|
| 649 | goto run; |
---|
| 650 | // Benchmarks |
---|
| 651 | case 'b': |
---|
| 652 | if(benchmark != NONE) { |
---|
| 653 | std::cerr << "Only when benchmark can be run" << std::endl; |
---|
| 654 | goto usage; |
---|
| 655 | } |
---|
| 656 | if(iequals(arg, "churn")) { |
---|
| 657 | benchmark = Churn; |
---|
| 658 | break; |
---|
| 659 | } |
---|
| 660 | if(iequals(arg, "pingpong")) { |
---|
| 661 | benchmark = PingPong; |
---|
| 662 | break; |
---|
| 663 | } |
---|
[807a632] | 664 | if(iequals(arg, "fairness")) { |
---|
| 665 | benchmark = Fairness; |
---|
| 666 | break; |
---|
| 667 | } |
---|
[9421f3d8] | 668 | std::cerr << "Unkown benchmark " << arg << std::endl; |
---|
| 669 | goto usage; |
---|
| 670 | // Numeric Arguments |
---|
| 671 | case 'd': |
---|
| 672 | try { |
---|
| 673 | duration = stod(optarg, &len); |
---|
| 674 | if(len != arg.size()) { throw std::invalid_argument(""); } |
---|
| 675 | } catch(std::invalid_argument &) { |
---|
| 676 | std::cerr << "Duration must be a valid double, was " << arg << std::endl; |
---|
| 677 | goto usage; |
---|
| 678 | } |
---|
| 679 | break; |
---|
| 680 | case 't': |
---|
| 681 | try { |
---|
| 682 | nthreads = stoul(optarg, &len); |
---|
| 683 | if(len != arg.size()) { throw std::invalid_argument(""); } |
---|
| 684 | } catch(std::invalid_argument &) { |
---|
| 685 | std::cerr << "Number of threads must be a positive integer, was " << arg << std::endl; |
---|
| 686 | goto usage; |
---|
| 687 | } |
---|
| 688 | break; |
---|
| 689 | case 'q': |
---|
| 690 | try { |
---|
| 691 | nqueues = stoul(optarg, &len); |
---|
| 692 | if(len != arg.size()) { throw std::invalid_argument(""); } |
---|
| 693 | } catch(std::invalid_argument &) { |
---|
| 694 | std::cerr << "Number of queues must be a positive integer, was " << arg << std::endl; |
---|
| 695 | goto usage; |
---|
| 696 | } |
---|
| 697 | break; |
---|
| 698 | // Other cases |
---|
| 699 | default: /* ? */ |
---|
| 700 | std::cerr << opt << std::endl; |
---|
| 701 | usage: |
---|
| 702 | std::cerr << "Usage: " << argv[0] << ": [options] -b churn [NNODES] [NSLOTS = NNODES]" << std::endl; |
---|
| 703 | std::cerr << " or: " << argv[0] << ": [options] -b pingpong [NNODES]" << std::endl; |
---|
| 704 | std::cerr << std::endl; |
---|
| 705 | std::cerr << " -d, --duration=DURATION Duration of the experiment, in seconds" << std::endl; |
---|
| 706 | std::cerr << " -t, --nthreads=NTHREADS Number of kernel threads" << std::endl; |
---|
| 707 | std::cerr << " -q, --nqueues=NQUEUES Number of queues per threads" << std::endl; |
---|
| 708 | std::exit(1); |
---|
[b2a37b0] | 709 | } |
---|
| 710 | } |
---|
[9421f3d8] | 711 | run: |
---|
[b2a37b0] | 712 | |
---|
| 713 | check_cache_line_size(); |
---|
| 714 | |
---|
[50aeb6f] | 715 | std::cout << "Running " << nthreads << " threads (" << (nthreads * nqueues) << " queues) for " << duration << " seconds" << std::endl; |
---|
[9421f3d8] | 716 | switch(benchmark) { |
---|
| 717 | case Churn: |
---|
| 718 | runChurn(nthreads, nqueues, duration, nnodes, nslots); |
---|
| 719 | break; |
---|
| 720 | case PingPong: |
---|
| 721 | runPingPong(nthreads, nqueues, duration, nnodes); |
---|
| 722 | break; |
---|
[807a632] | 723 | case Fairness: |
---|
| 724 | runFairness(nthreads, nqueues, duration, nnodes, out); |
---|
| 725 | break; |
---|
[9421f3d8] | 726 | default: |
---|
| 727 | abort(); |
---|
| 728 | } |
---|
[b2a37b0] | 729 | return 0; |
---|
| 730 | } |
---|
| 731 | |
---|
[807a632] | 732 | const char * __my_progname = "Relaxed List"; |
---|
| 733 | |
---|
| 734 | struct rgb_t { |
---|
| 735 | double r; // a fraction between 0 and 1 |
---|
| 736 | double g; // a fraction between 0 and 1 |
---|
| 737 | double b; // a fraction between 0 and 1 |
---|
| 738 | }; |
---|
| 739 | |
---|
| 740 | struct hsv_t { |
---|
| 741 | double h; // angle in degrees |
---|
| 742 | double s; // a fraction between 0 and 1 |
---|
| 743 | double v; // a fraction between 0 and 1 |
---|
| 744 | }; |
---|
| 745 | |
---|
| 746 | rgb_t hsv2rgb(hsv_t in) { |
---|
| 747 | double hh, p, q, t, ff; |
---|
| 748 | long i; |
---|
| 749 | rgb_t out; |
---|
| 750 | |
---|
| 751 | if(in.s <= 0.0) { // < is bogus, just shuts up warnings |
---|
| 752 | out.r = in.v; |
---|
| 753 | out.g = in.v; |
---|
| 754 | out.b = in.v; |
---|
| 755 | return out; |
---|
| 756 | } |
---|
| 757 | hh = in.h; |
---|
| 758 | if(hh >= 360.0) hh = 0.0; |
---|
| 759 | hh /= 60.0; |
---|
| 760 | i = (long)hh; |
---|
| 761 | ff = hh - i; |
---|
| 762 | p = in.v * (1.0 - in.s); |
---|
| 763 | q = in.v * (1.0 - (in.s * ff)); |
---|
| 764 | t = in.v * (1.0 - (in.s * (1.0 - ff))); |
---|
| 765 | |
---|
| 766 | switch(i) { |
---|
| 767 | case 0: |
---|
| 768 | out.r = in.v; |
---|
| 769 | out.g = t; |
---|
| 770 | out.b = p; |
---|
| 771 | break; |
---|
| 772 | case 1: |
---|
| 773 | out.r = q; |
---|
| 774 | out.g = in.v; |
---|
| 775 | out.b = p; |
---|
| 776 | break; |
---|
| 777 | case 2: |
---|
| 778 | out.r = p; |
---|
| 779 | out.g = in.v; |
---|
| 780 | out.b = t; |
---|
| 781 | break; |
---|
| 782 | |
---|
| 783 | case 3: |
---|
| 784 | out.r = p; |
---|
| 785 | out.g = q; |
---|
| 786 | out.b = in.v; |
---|
| 787 | break; |
---|
| 788 | case 4: |
---|
| 789 | out.r = t; |
---|
| 790 | out.g = p; |
---|
| 791 | out.b = in.v; |
---|
| 792 | break; |
---|
| 793 | case 5: |
---|
| 794 | default: |
---|
| 795 | out.r = in.v; |
---|
| 796 | out.g = p; |
---|
| 797 | out.b = q; |
---|
| 798 | break; |
---|
| 799 | } |
---|
| 800 | return out; |
---|
| 801 | } |
---|
| 802 | |
---|
| 803 | void save_fairness(const int data[], int factor, unsigned nthreads, size_t columns, size_t rows, const std::string & output) { |
---|
| 804 | std::ofstream os(output); |
---|
| 805 | os << "<html>\n"; |
---|
| 806 | os << "<head>\n"; |
---|
| 807 | os << "<style>\n"; |
---|
| 808 | os << "</style>\n"; |
---|
| 809 | os << "</head>\n"; |
---|
| 810 | os << "<body>\n"; |
---|
| 811 | os << "<table style=\"width=100%\">\n"; |
---|
| 812 | |
---|
| 813 | size_t idx = 0; |
---|
| 814 | for(size_t r = 0ul; r < rows; r++) { |
---|
| 815 | os << "<tr>\n"; |
---|
| 816 | for(size_t c = 0ul; c < columns; c++) { |
---|
| 817 | os << "<td class=\"custom custom" << data[idx] << "\"></td>\n"; |
---|
| 818 | idx++; |
---|
| 819 | } |
---|
| 820 | os << "</tr>\n"; |
---|
| 821 | } |
---|
| 822 | |
---|
| 823 | os << "</table>\n"; |
---|
| 824 | os << "</body>\n"; |
---|
| 825 | os << "</html>\n"; |
---|
| 826 | os << std::endl; |
---|
| 827 | } |
---|
| 828 | |
---|
| 829 | #include <png.h> |
---|
| 830 | #include <setjmp.h> |
---|
| 831 | |
---|
| 832 | /* |
---|
| 833 | void save_fairness(const int data[], int factor, unsigned nthreads, size_t columns, size_t rows, const std::string & output) { |
---|
| 834 | int width = columns * factor; |
---|
| 835 | int height = rows / factor; |
---|
| 836 | |
---|
| 837 | int code = 0; |
---|
| 838 | int idx = 0; |
---|
| 839 | FILE *fp = NULL; |
---|
| 840 | png_structp png_ptr = NULL; |
---|
| 841 | png_infop info_ptr = NULL; |
---|
| 842 | png_bytep row = NULL; |
---|
| 843 | |
---|
| 844 | // Open file for writing (binary mode) |
---|
| 845 | fp = fopen(output.c_str(), "wb"); |
---|
| 846 | if (fp == NULL) { |
---|
| 847 | fprintf(stderr, "Could not open file %s for writing\n", output.c_str()); |
---|
| 848 | code = 1; |
---|
| 849 | goto finalise; |
---|
| 850 | } |
---|
| 851 | |
---|
| 852 | // Initialize write structure |
---|
| 853 | png_ptr = png_create_write_struct(PNG_LIBPNG_VER_STRING, NULL, NULL, NULL); |
---|
| 854 | if (png_ptr == NULL) { |
---|
| 855 | fprintf(stderr, "Could not allocate write struct\n"); |
---|
| 856 | code = 1; |
---|
| 857 | goto finalise; |
---|
| 858 | } |
---|
| 859 | |
---|
| 860 | // Initialize info structure |
---|
| 861 | info_ptr = png_create_info_struct(png_ptr); |
---|
| 862 | if (info_ptr == NULL) { |
---|
| 863 | fprintf(stderr, "Could not allocate info struct\n"); |
---|
| 864 | code = 1; |
---|
| 865 | goto finalise; |
---|
| 866 | } |
---|
| 867 | |
---|
| 868 | // Setup Exception handling |
---|
| 869 | if (setjmp(png_jmpbuf(png_ptr))) { |
---|
| 870 | fprintf(stderr, "Error during png creation\n"); |
---|
| 871 | code = 1; |
---|
| 872 | goto finalise; |
---|
| 873 | } |
---|
| 874 | |
---|
| 875 | png_init_io(png_ptr, fp); |
---|
| 876 | |
---|
| 877 | // Write header (8 bit colour depth) |
---|
| 878 | png_set_IHDR(png_ptr, info_ptr, width, height, |
---|
| 879 | 8, PNG_COLOR_TYPE_RGB, PNG_INTERLACE_NONE, |
---|
| 880 | PNG_COMPRESSION_TYPE_BASE, PNG_FILTER_TYPE_BASE); |
---|
| 881 | |
---|
| 882 | png_write_info(png_ptr, info_ptr); |
---|
| 883 | |
---|
| 884 | // Allocate memory for one row (3 bytes per pixel - RGB) |
---|
| 885 | row = (png_bytep) malloc(3 * width * sizeof(png_byte)); |
---|
| 886 | |
---|
| 887 | // Write image data |
---|
| 888 | int x, y; |
---|
| 889 | for (y=0 ; y<height ; y++) { |
---|
| 890 | for (x=0 ; x<width ; x++) { |
---|
| 891 | auto & r = row[(x * 3) + 0]; |
---|
| 892 | auto & g = row[(x * 3) + 1]; |
---|
| 893 | auto & b = row[(x * 3) + 2]; |
---|
| 894 | assert(idx < (rows * columns)); |
---|
| 895 | int color = data[idx] - 1; |
---|
| 896 | assert(color < nthreads); |
---|
| 897 | assert(color >= 0); |
---|
| 898 | idx++; |
---|
| 899 | |
---|
| 900 | double angle = double(color) / double(nthreads); |
---|
| 901 | |
---|
| 902 | auto c = hsv2rgb({ 360.0 * angle, 0.8, 0.8 }); |
---|
| 903 | |
---|
| 904 | r = char(c.r * 255.0); |
---|
| 905 | g = char(c.g * 255.0); |
---|
| 906 | b = char(c.b * 255.0); |
---|
| 907 | |
---|
| 908 | } |
---|
| 909 | png_write_row(png_ptr, row); |
---|
| 910 | } |
---|
| 911 | |
---|
| 912 | assert(idx == (rows * columns)); |
---|
| 913 | |
---|
| 914 | // End write |
---|
| 915 | png_write_end(png_ptr, NULL); |
---|
| 916 | |
---|
| 917 | finalise: |
---|
| 918 | if (fp != NULL) fclose(fp); |
---|
| 919 | if (info_ptr != NULL) png_free_data(png_ptr, info_ptr, PNG_FREE_ALL, -1); |
---|
| 920 | if (png_ptr != NULL) png_destroy_write_struct(&png_ptr, (png_infopp)NULL); |
---|
| 921 | if (row != NULL) free(row); |
---|
| 922 | } |
---|
| 923 | */ |
---|