| 1 | // -------------------------------------------------------------------------- |
| 2 | // |
| 3 | // Copyright |
| 4 | // Markus Wittmann, 2016-2017 |
| 5 | // RRZE, University of Erlangen-Nuremberg, Germany |
| 6 | // markus.wittmann -at- fau.de or hpc -at- rrze.fau.de |
| 7 | // |
| 8 | // Viktor Haag, 2016 |
| 9 | // LSS, University of Erlangen-Nuremberg, Germany |
| 10 | // |
| 11 | // This file is part of the Lattice Boltzmann Benchmark Kernels (LbmBenchKernels). |
| 12 | // |
| 13 | // LbmBenchKernels is free software: you can redistribute it and/or modify |
| 14 | // it under the terms of the GNU General Public License as published by |
| 15 | // the Free Software Foundation, either version 3 of the License, or |
| 16 | // (at your option) any later version. |
| 17 | // |
| 18 | // LbmBenchKernels is distributed in the hope that it will be useful, |
| 19 | // but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 20 | // MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 21 | // GNU General Public License for more details. |
| 22 | // |
| 23 | // You should have received a copy of the GNU General Public License |
| 24 | // along with LbmBenchKernels. If not, see <http://www.gnu.org/licenses/>. |
| 25 | // |
| 26 | // -------------------------------------------------------------------------- |
| 27 | #ifndef _GNU_SOURCE |
| 28 | #define _GNU_SOURCE |
| 29 | #endif |
| 30 | #include <sched.h> |
| 31 | #include <errno.h> |
| 32 | |
| 33 | |
| 34 | #include "Base.h" |
| 35 | #include "Pinning.h" |
| 36 | |
| 37 | |
| 38 | |
| 39 | |
| 40 | // ----------------------------------------------------------------------- |
| 41 | // |
| 42 | // Binds the calling thread to specified core. |
| 43 | // |
| 44 | // Return value: 0 = success, else error. |
| 45 | // |
| 46 | // ----------------------------------------------------------------------- |
| 47 | |
| 48 | int PinCurrentThreadToCore(int coreNumber) |
| 49 | { |
| 50 | int error = 0; |
| 51 | |
| 52 | cpu_set_t cpu_set; |
| 53 | CPU_ZERO(&cpu_set); |
| 54 | CPU_SET(coreNumber, &cpu_set); |
| 55 | |
| 56 | error = sched_setaffinity((pid_t)0, sizeof(cpu_set_t), &cpu_set); |
| 57 | |
| 58 | if (error != 0) { |
| 59 | Error("pinning thread to core %d failed (%d): %s\n", |
| 60 | coreNumber, error, strerror(error)); |
| 61 | } |
| 62 | |
| 63 | return error; |
| 64 | } |
| 65 | |
| 66 | |
| 67 | // ----------------------------------------------------------------------- |
| 68 | // |
| 69 | // Binds the calling thread to specified core by a cpu list specified |
| 70 | // in the given environment variable. |
| 71 | // |
| 72 | // Return value: 0 = success, else error. |
| 73 | // |
| 74 | // ----------------------------------------------------------------------- |
| 75 | |
| 76 | int PinCurrentThreadByEnvVar(const char * envVarName, |
| 77 | int mpiRank, int nodeRank, int threadNumber) |
| 78 | { |
| 79 | const char * envVarValue; |
| 80 | int core; |
| 81 | |
| 82 | envVarValue = getenv(envVarName); |
| 83 | |
| 84 | if (envVarValue == NULL) { |
| 85 | if (mpiRank == 0) { |
| 86 | Print("skip pinning: env var %s not set\n", envVarName); |
| 87 | } |
| 88 | |
| 89 | return 0; |
| 90 | } |
| 91 | |
| 92 | core = PinParseCpuList(envVarValue, mpiRank, nodeRank, threadNumber); |
| 93 | |
| 94 | if (core < 0) { |
| 95 | return core; |
| 96 | } |
| 97 | |
| 98 | return PinCurrentThreadToCore(core); |
| 99 | } |
| 100 | |
| 101 | |
| 102 | // ----------------------------------------------------------------------- |
| 103 | // |
| 104 | // Binds the calling thread to a core specified in the CPU list. |
| 105 | // |
| 106 | // Return value: 0 = success, else error. |
| 107 | // |
| 108 | // ----------------------------------------------------------------------- |
| 109 | |
| 110 | int PinCurrentThreadByCpuList(const char * cpuList, |
| 111 | int mpiRank, int nodeRank, int threadNumber) |
| 112 | { |
| 113 | int core; |
| 114 | |
| 115 | if (cpuList == NULL) { |
| 116 | if (mpiRank == 0) { |
| 117 | printf("ERROR: cpu list is NULL.\n"); |
| 118 | } |
| 119 | |
| 120 | exit(1); |
| 121 | } |
| 122 | |
| 123 | core = PinParseCpuList(cpuList, mpiRank, nodeRank, threadNumber); |
| 124 | |
| 125 | if (core < 0) { |
| 126 | return core; |
| 127 | } |
| 128 | |
| 129 | return PinCurrentThreadToCore(core); |
| 130 | } |
| 131 | |
| 132 | |
| 133 | // ----------------------------------------------------------------------- |
| 134 | // |
| 135 | // Parses the provided cpu list and returns the core number for the |
| 136 | // specified MPI rank, local rank, and thread. |
| 137 | // |
| 138 | // The cpu list has for example a format of: 0,1,2 or 0,1,2_3,4,5 |
| 139 | // |
| 140 | // Blocks (0,1,2 or 3,4,5) separated by "_" specify pinning inside a |
| 141 | // node rank. The first block maps to node rank 1, the second to node |
| 142 | // rank 2, etc. |
| 143 | // |
| 144 | // Inside a block the core numbers specify where the threads should |
| 145 | // be pinned to. They are separated by "," and the first number maps |
| 146 | // to the first core, the second number to the second core, etc. |
| 147 | // |
| 148 | // For example: 0,2,4_6,8,10 |
| 149 | // |
| 150 | // Node rank 0 thread 0 pinned to core 0 |
| 151 | // 0 1 2 |
| 152 | // 0 2 4 |
| 153 | // 1 0 6 |
| 154 | // 1 1 8 |
| 155 | // 1 2 10 |
| 156 | // |
| 157 | // ----------------------------------------------------------------------- |
| 158 | |
| 159 | int PinParseCpuList(const char * cpuList, |
| 160 | int mpiRank, int nodeRank, int threadNumber) |
| 161 | { |
| 162 | int cpu = -1; |
| 163 | |
| 164 | if (cpuList == NULL) { |
| 165 | return -1; |
| 166 | } |
| 167 | |
| 168 | const char * c = cpuList; |
| 169 | |
| 170 | // Ensure only valid characters are in the cpu list. |
| 171 | // Cpu list is in the format of "0,1,2_3,4,5". |
| 172 | while (((*c >= '0' && *c <= '9') || *c == ',' || *c == '_')) { |
| 173 | ++c; |
| 174 | } |
| 175 | |
| 176 | if (*c != 0x00) { |
| 177 | // Invalid character detected. |
| 178 | return -2; |
| 179 | } |
| 180 | |
| 181 | c = cpuList; |
| 182 | |
| 183 | int i = 0; |
| 184 | |
| 185 | // Move variable c after the "nodeRank"th "_" in the cpu list. |
| 186 | while (i < nodeRank && *c != 0x00) { |
| 187 | if (*c == '_') ++i; |
| 188 | ++c; |
| 189 | } |
| 190 | |
| 191 | if (i != nodeRank || *c < '0' || *c > '9') { |
| 192 | // Cpu list for this node rank not found. |
| 193 | return -3; |
| 194 | } |
| 195 | |
| 196 | // Now find the core for the specified thread. |
| 197 | |
| 198 | int t = 0; |
| 199 | |
| 200 | while (t < threadNumber && *c != 0x00) { |
| 201 | if (*c == ',') { |
| 202 | ++t; |
| 203 | } |
| 204 | else if (*c == '_') { |
| 205 | // Unexpected character at this position. |
| 206 | break; |
| 207 | } |
| 208 | |
| 209 | ++c; |
| 210 | } |
| 211 | |
| 212 | if (t != threadNumber || *c < '0' || *c > '9') { |
| 213 | // Cpu for this threadNumber not found. |
| 214 | return -4; |
| 215 | } |
| 216 | |
| 217 | cpu = atoi(c); |
| 218 | |
| 219 | return cpu; |
| 220 | } |
| 221 | |
| 222 | |
| 223 | |
| 224 | // ----------------------------------------------------------------------- |
| 225 | // |
| 226 | // Returns the first core from the calling thread's affinity set. |
| 227 | // |
| 228 | // On error a value < 0 is returned. |
| 229 | // |
| 230 | // ----------------------------------------------------------------------- |
| 231 | |
| 232 | int PinCurrentCore() |
| 233 | { |
| 234 | int core = -1; |
| 235 | int err; |
| 236 | |
| 237 | cpu_set_t cpu_set; |
| 238 | CPU_ZERO(&cpu_set); |
| 239 | |
| 240 | err = sched_getaffinity((pid_t)0, sizeof(cpu_set_t), &cpu_set); |
| 241 | |
| 242 | // constant CPU_SETSIZE is one larger than the maximum CPU |
| 243 | // number that can be stored in a CPU set |
| 244 | for (int i = 0; i < CPU_SETSIZE; ++i) { |
| 245 | if (CPU_ISSET(i, &cpu_set)) { |
| 246 | core = i; |
| 247 | break; |
| 248 | } |
| 249 | } |
| 250 | |
| 251 | if (err != 0) { |
| 252 | Error("getting thread affinty failed (%d): %s\n", err, strerror(err)); |
| 253 | return -1; |
| 254 | } |
| 255 | |
| 256 | return core; |
| 257 | } |
| 258 | |
| 259 | |
| 260 | |
| 261 | // ----------------------------------------------------------------------- |
| 262 | // |
| 263 | // Returns the all cores from the calling thread's affinity set. |
| 264 | // |
| 265 | // On error a value < 0 is returned. |
| 266 | // |
| 267 | // ----------------------------------------------------------------------- |
| 268 | |
| 269 | typedef cpu_set_t CpuSet; |
| 270 | |
| 271 | |
| 272 | static CpuSet PinCurrentCores() |
| 273 | { |
| 274 | CpuSet cpuSet; |
| 275 | int err; |
| 276 | |
| 277 | cpu_set_t cpu_set; |
| 278 | CPU_ZERO(&cpu_set); |
| 279 | |
| 280 | err = sched_getaffinity((pid_t)0, sizeof(cpu_set_t), &cpu_set); |
| 281 | |
| 282 | cpuSet = cpu_set; |
| 283 | |
| 284 | if (err != 0) { |
| 285 | Error("getting thread affinty failed (%d): %s\n", err, strerror(err)); |
| 286 | return cpuSet; |
| 287 | } |
| 288 | |
| 289 | return cpuSet; |
| 290 | } |
| 291 | |
| 292 | static char * CpuSetToString(cpu_set_t * cpu_set) |
| 293 | { |
| 294 | int previousSetCore = -2; |
| 295 | int rangeBeginCore = -2; |
| 296 | |
| 297 | char * buffer1 = (char *)malloc(1024); |
| 298 | Assert(buffer1 != NULL); |
| 299 | char * buffer2 = (char *)malloc(1024); |
| 300 | Assert(buffer2 != NULL); |
| 301 | |
| 302 | buffer1[0] = 0x00; |
| 303 | buffer2[0] = 0x00; |
| 304 | |
| 305 | char * buffer = buffer1; |
| 306 | char * bufferOld = buffer2; |
| 307 | |
| 308 | const char * empty = ""; |
| 309 | const char * realComma = ","; |
| 310 | const char * comma = empty; |
| 311 | |
| 312 | // TODO: use snprintf |
| 313 | // TODO: increase allocated buffer if necessary |
| 314 | |
| 315 | for (int i = 0; i < CPU_SETSIZE; ++i) { |
| 316 | if (!CPU_ISSET(i, cpu_set)) { |
| 317 | continue; |
| 318 | } |
| 319 | |
| 320 | if (i == previousSetCore + 1) { |
| 321 | previousSetCore = i; |
| 322 | continue; |
| 323 | } |
| 324 | |
| 325 | // Now we reached the end of a range. |
| 326 | // The range can also consist of only one core. |
| 327 | // Be aware, that this core is not part of the range. |
| 328 | |
| 329 | // TODO: this code is repeated below -> use it only once |
| 330 | if (rangeBeginCore >= 0 && previousSetCore >= 0) { |
| 331 | char * tmp; |
| 332 | |
| 333 | tmp = buffer; |
| 334 | buffer = bufferOld; |
| 335 | bufferOld = tmp; |
| 336 | |
| 337 | if (rangeBeginCore < previousSetCore) { |
| 338 | sprintf(buffer, "%s%s%d-%d", bufferOld, comma, rangeBeginCore, previousSetCore); |
| 339 | } |
| 340 | else { |
| 341 | sprintf(buffer, "%s%s%d", bufferOld, comma, previousSetCore); |
| 342 | } |
| 343 | |
| 344 | comma = realComma; |
| 345 | } |
| 346 | |
| 347 | // With this core a new range begins. |
| 348 | rangeBeginCore = i; |
| 349 | previousSetCore = i; |
| 350 | } |
| 351 | |
| 352 | if (rangeBeginCore >= 0 && previousSetCore >= 0) { |
| 353 | char * tmp; |
| 354 | |
| 355 | tmp = buffer; |
| 356 | buffer = bufferOld; |
| 357 | bufferOld = tmp; |
| 358 | |
| 359 | if (rangeBeginCore < previousSetCore) { |
| 360 | sprintf(buffer, "%s%s%d-%d", bufferOld, comma, rangeBeginCore, previousSetCore); |
| 361 | } |
| 362 | else { |
| 363 | sprintf(buffer, "%s%s%d", bufferOld, comma, previousSetCore); |
| 364 | } |
| 365 | } |
| 366 | |
| 367 | free(bufferOld); bufferOld = NULL; |
| 368 | |
| 369 | return buffer; |
| 370 | } |
| 371 | |
| 372 | char * PinCpuListAsString() |
| 373 | { |
| 374 | CpuSet cpuSet = PinCurrentCores(); |
| 375 | |
| 376 | return CpuSetToString(&cpuSet); |
| 377 | } |
| 378 | |
| 379 | #ifdef TEST |
| 380 | |
| 381 | int main(int argc, char * argv[]) |
| 382 | { |
| 383 | char * cpuList = PinCpuListAsString(); |
| 384 | |
| 385 | printf("pinned to cores: %s\n", cpuList); |
| 386 | |
| 387 | free(cpuList); cpuList = NULL; |
| 388 | |
| 389 | return 0; |
| 390 | } |
| 391 | |
| 392 | #endif // TEST |
| 393 | |