starpu_mpi_insert_task.c 18 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403
  1. /* StarPU --- Runtime system for heterogeneous multicore architectures.
  2. *
  3. * Copyright (C) 2011, 2012 Centre National de la Recherche Scientifique
  4. * Copyright (C) 2011 Université de Bordeaux 1
  5. *
  6. * StarPU is free software; you can redistribute it and/or modify
  7. * it under the terms of the GNU Lesser General Public License as published by
  8. * the Free Software Foundation; either version 2.1 of the License, or (at
  9. * your option) any later version.
  10. *
  11. * StarPU is distributed in the hope that it will be useful, but
  12. * WITHOUT ANY WARRANTY; without even the implied warranty of
  13. * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
  14. *
  15. * See the GNU Lesser General Public License in COPYING.LGPL for more details.
  16. */
  17. #include <stdarg.h>
  18. #include <mpi.h>
  19. #include <starpu.h>
  20. #include <starpu_data.h>
  21. #include <common/utils.h>
  22. #include <starpu_hash.h>
  23. #include <common/htable32.h>
  24. #include <util/starpu_insert_task_utils.h>
  25. //#define STARPU_MPI_VERBOSE 1
  26. #include <starpu_mpi_private.h>
  27. /* Whether we are allowed to keep copies of remote data. Does not work
  28. * yet: the sender has to know whether the receiver has it, keeping it
  29. * in an array indexed by node numbers. */
  30. //#define MPI_CACHE
  31. #include <starpu_mpi_insert_task_cache.h>
  32. static void _starpu_mpi_tables_init()
  33. {
  34. if (sent_data == NULL) {
  35. int nb_nodes;
  36. int i;
  37. MPI_Comm_size(MPI_COMM_WORLD, &nb_nodes);
  38. _STARPU_MPI_DEBUG("Initialising hash table for cache\n");
  39. sent_data = malloc(nb_nodes * sizeof(struct starpu_htbl32_node *));
  40. for(i=0 ; i<nb_nodes ; i++) sent_data[i] = NULL;
  41. received_data = malloc(nb_nodes * sizeof(struct starpu_htbl32_node *));
  42. for(i=0 ; i<nb_nodes ; i++) received_data[i] = NULL;
  43. }
  44. }
  45. void _starpu_data_deallocate(starpu_data_handle_t data_handle)
  46. {
  47. #ifdef STARPU_DEVEL
  48. #warning _starpu_data_deallocate not implemented yet
  49. #endif
  50. }
  51. int starpu_mpi_insert_task(MPI_Comm comm, struct starpu_codelet *codelet, ...)
  52. {
  53. int arg_type;
  54. va_list varg_list;
  55. int me, do_execute, xrank, nb_nodes;
  56. size_t arg_buffer_size = 0;
  57. char *arg_buffer;
  58. int dest=0, inconsistent_execute;
  59. _STARPU_MPI_LOG_IN();
  60. MPI_Comm_rank(comm, &me);
  61. MPI_Comm_size(comm, &nb_nodes);
  62. _starpu_mpi_tables_init();
  63. /* Get the number of buffers and the size of the arguments */
  64. va_start(varg_list, codelet);
  65. arg_buffer_size = _starpu_insert_task_get_arg_size(varg_list);
  66. va_start(varg_list, codelet);
  67. _starpu_pack_cl_args(arg_buffer_size, &arg_buffer, varg_list);
  68. /* Find out whether we are to execute the data because we own the data to be written to. */
  69. inconsistent_execute = 0;
  70. do_execute = -1;
  71. xrank = -1;
  72. va_start(varg_list, codelet);
  73. while ((arg_type = va_arg(varg_list, int)) != 0) {
  74. if (arg_type==STARPU_EXECUTE_ON_NODE) {
  75. xrank = va_arg(varg_list, int);
  76. _STARPU_MPI_DEBUG("Executing on node %d\n", xrank);
  77. do_execute = 1;
  78. }
  79. else if (arg_type==STARPU_EXECUTE_ON_DATA) {
  80. starpu_data_handle_t data = va_arg(varg_list, starpu_data_handle_t);
  81. xrank = starpu_data_get_rank(data);
  82. _STARPU_MPI_DEBUG("Executing on data node %d\n", xrank);
  83. STARPU_ASSERT(xrank <= nb_nodes);
  84. do_execute = 1;
  85. }
  86. if (arg_type==STARPU_R || arg_type==STARPU_W || arg_type==STARPU_REDUX || arg_type==STARPU_RW || arg_type == STARPU_SCRATCH) {
  87. starpu_data_handle_t data = va_arg(varg_list, starpu_data_handle_t);
  88. if (arg_type & STARPU_W || arg_type & STARPU_REDUX) {
  89. if (!data) {
  90. /* We don't have anything allocated for this.
  91. * The application knows we won't do anything
  92. * about this task */
  93. /* Yes, the app could actually not call
  94. * insert_task at all itself, this is just a
  95. * safeguard. */
  96. _STARPU_MPI_DEBUG("oh oh\n");
  97. _STARPU_MPI_LOG_OUT();
  98. return -EINVAL;
  99. }
  100. int mpi_rank = starpu_data_get_rank(data);
  101. if (mpi_rank == me) {
  102. if (do_execute == 0) {
  103. inconsistent_execute = 1;
  104. }
  105. else {
  106. do_execute = 1;
  107. }
  108. }
  109. else if (mpi_rank != -1) {
  110. if (do_execute == 1) {
  111. inconsistent_execute = 1;
  112. }
  113. else {
  114. do_execute = 0;
  115. dest = mpi_rank;
  116. /* That's the rank which needs the data to be sent to */
  117. }
  118. }
  119. else {
  120. _STARPU_ERROR("rank invalid\n");
  121. }
  122. }
  123. }
  124. else if (arg_type==STARPU_VALUE) {
  125. va_arg(varg_list, void *);
  126. va_arg(varg_list, size_t);
  127. }
  128. else if (arg_type==STARPU_CALLBACK) {
  129. va_arg(varg_list, void (*)(void *));
  130. }
  131. else if (arg_type==STARPU_CALLBACK_WITH_ARG) {
  132. va_arg(varg_list, void (*)(void *));
  133. va_arg(varg_list, void *);
  134. }
  135. else if (arg_type==STARPU_CALLBACK_ARG) {
  136. va_arg(varg_list, void *);
  137. }
  138. else if (arg_type==STARPU_PRIORITY) {
  139. va_arg(varg_list, int);
  140. }
  141. else if (arg_type==STARPU_EXECUTE_ON_NODE) {
  142. va_arg(varg_list, int);
  143. }
  144. else if (arg_type==STARPU_EXECUTE_ON_DATA) {
  145. va_arg(varg_list, int);
  146. }
  147. }
  148. va_end(varg_list);
  149. STARPU_ASSERT(do_execute != -1 && "StarPU needs to see a W or a REDUX data which will tell it where to execute the task");
  150. if (inconsistent_execute == 1) {
  151. if (xrank == -1) {
  152. _STARPU_MPI_DEBUG("Different tasks are owning W data. Needs to specify which one is to execute the codelet, using STARPU_EXECUTE_ON_NODE or STARPU_EXECUTE_ON_DATA\n");
  153. return -EINVAL;
  154. }
  155. else {
  156. do_execute = (me == xrank);
  157. dest = xrank;
  158. }
  159. }
  160. else if (xrank != -1) {
  161. _STARPU_MPI_DEBUG("Property STARPU_EXECUTE_ON_NODE or STARPU_EXECUTE_ON_DATA overwriting node defined by data model\n");
  162. do_execute = (me == xrank);
  163. dest = xrank;
  164. }
  165. _STARPU_MPI_DEBUG("Executing %d - Sending to node %d\n", do_execute, dest);
  166. /* Send and receive data as requested */
  167. va_start(varg_list, codelet);
  168. while ((arg_type = va_arg(varg_list, int)) != 0) {
  169. if (arg_type==STARPU_R || arg_type==STARPU_W || arg_type==STARPU_RW || arg_type == STARPU_SCRATCH) {
  170. starpu_data_handle_t data = va_arg(varg_list, starpu_data_handle_t);
  171. if (data && arg_type & STARPU_R) {
  172. int mpi_rank = starpu_data_get_rank(data);
  173. int mpi_tag = starpu_data_get_tag(data);
  174. STARPU_ASSERT(mpi_tag >= 0 && "StarPU needs to be told the MPI rank of this data, using starpu_data_set_rank");
  175. /* The task needs to read this data */
  176. if (do_execute && mpi_rank != me && mpi_rank != -1) {
  177. /* I will have to execute but I don't have the data, receive */
  178. #ifdef MPI_CACHE
  179. uint32_t key = starpu_crc32_be((uintptr_t)data, 0);
  180. void *already_received = _starpu_htbl_search_32(received_data[mpi_rank], key);
  181. if (!already_received) {
  182. _starpu_htbl_insert_32(&received_data[mpi_rank], key, data);
  183. }
  184. else {
  185. _STARPU_MPI_DEBUG("Do not receive data %p from node %d as it is already available\n", data, mpi_rank);
  186. }
  187. if (!already_received)
  188. #endif
  189. {
  190. _STARPU_MPI_DEBUG("Receive data %p from %d\n", data, mpi_rank);
  191. starpu_mpi_irecv_detached(data, mpi_rank, mpi_tag, comm, NULL, NULL);
  192. }
  193. }
  194. if (!do_execute && mpi_rank == me) {
  195. /* Somebody else will execute it, and I have the data, send it. */
  196. #ifdef MPI_CACHE
  197. uint32_t key = starpu_crc32_be((uintptr_t)data, 0);
  198. void *already_sent = _starpu_htbl_search_32(sent_data[dest], key);
  199. if (!already_sent) {
  200. _starpu_htbl_insert_32(&sent_data[dest], key, data);
  201. }
  202. else {
  203. _STARPU_MPI_DEBUG("Do not sent data %p to node %d as it has already been sent\n", data, dest);
  204. }
  205. if (!already_sent)
  206. #endif
  207. {
  208. _STARPU_MPI_DEBUG("Send data %p to %d\n", data, dest);
  209. starpu_mpi_isend_detached(data, dest, mpi_tag, comm, NULL, NULL);
  210. }
  211. }
  212. }
  213. }
  214. else if (arg_type==STARPU_VALUE) {
  215. va_arg(varg_list, void *);
  216. va_arg(varg_list, size_t);
  217. }
  218. else if (arg_type==STARPU_CALLBACK) {
  219. va_arg(varg_list, void (*)(void *));
  220. }
  221. else if (arg_type==STARPU_CALLBACK_WITH_ARG) {
  222. va_arg(varg_list, void (*)(void *));
  223. va_arg(varg_list, void *);
  224. }
  225. else if (arg_type==STARPU_CALLBACK_ARG) {
  226. va_arg(varg_list, void *);
  227. }
  228. else if (arg_type==STARPU_PRIORITY) {
  229. va_arg(varg_list, int);
  230. }
  231. else if (arg_type==STARPU_EXECUTE_ON_NODE) {
  232. va_arg(varg_list, int);
  233. }
  234. else if (arg_type==STARPU_EXECUTE_ON_DATA) {
  235. va_arg(varg_list, starpu_data_handle_t);
  236. }
  237. }
  238. va_end(varg_list);
  239. if (do_execute) {
  240. _STARPU_MPI_DEBUG("Execution of the codelet %p (%s)\n", codelet, codelet->name);
  241. va_start(varg_list, codelet);
  242. struct starpu_task *task = starpu_task_create();
  243. int ret = _starpu_insert_task_create_and_submit(arg_buffer, codelet, &task, varg_list);
  244. _STARPU_MPI_DEBUG("ret: %d\n", ret);
  245. STARPU_ASSERT(ret==0);
  246. }
  247. if (inconsistent_execute) {
  248. va_start(varg_list, codelet);
  249. while ((arg_type = va_arg(varg_list, int)) != 0) {
  250. if (arg_type==STARPU_R || arg_type==STARPU_W || arg_type==STARPU_RW || arg_type == STARPU_SCRATCH) {
  251. starpu_data_handle_t data = va_arg(varg_list, starpu_data_handle_t);
  252. if (arg_type & STARPU_W) {
  253. int mpi_rank = starpu_data_get_rank(data);
  254. int mpi_tag = starpu_data_get_tag(data);
  255. STARPU_ASSERT(mpi_tag >= 0 && "StarPU needs to be told the MPI rank of this data, using starpu_data_set_rank");
  256. if (mpi_rank == me) {
  257. if (xrank != -1 && me != xrank) {
  258. _STARPU_MPI_DEBUG("Receive data %p back from the task %d which executed the codelet ...\n", data, dest);
  259. starpu_mpi_irecv_detached(data, dest, mpi_tag, comm, NULL, NULL);
  260. }
  261. }
  262. else if (do_execute) {
  263. _STARPU_MPI_DEBUG("Send data %p back to its owner %d...\n", data, mpi_rank);
  264. starpu_mpi_isend_detached(data, mpi_rank, mpi_tag, comm, NULL, NULL);
  265. }
  266. }
  267. }
  268. else if (arg_type==STARPU_VALUE) {
  269. va_arg(varg_list, void *);
  270. va_arg(varg_list, size_t);
  271. }
  272. else if (arg_type==STARPU_CALLBACK) {
  273. va_arg(varg_list, void (*)(void *));
  274. }
  275. else if (arg_type==STARPU_CALLBACK_WITH_ARG) {
  276. va_arg(varg_list, void (*)(void *));
  277. va_arg(varg_list, void *);
  278. }
  279. else if (arg_type==STARPU_CALLBACK_ARG) {
  280. va_arg(varg_list, void *);
  281. }
  282. else if (arg_type==STARPU_PRIORITY) {
  283. va_arg(varg_list, int);
  284. }
  285. else if (arg_type==STARPU_EXECUTE_ON_NODE) {
  286. va_arg(varg_list, int);
  287. }
  288. else if (arg_type==STARPU_EXECUTE_ON_DATA) {
  289. va_arg(varg_list, starpu_data_handle_t);
  290. }
  291. }
  292. va_end(varg_list);
  293. }
  294. va_start(varg_list, codelet);
  295. while ((arg_type = va_arg(varg_list, int)) != 0) {
  296. if (arg_type==STARPU_R || arg_type==STARPU_W || arg_type==STARPU_RW || arg_type == STARPU_SCRATCH) {
  297. starpu_data_handle_t data = va_arg(varg_list, starpu_data_handle_t);
  298. #ifdef MPI_CACHE
  299. if (arg_type & STARPU_W) {
  300. uint32_t key = starpu_crc32_be((uintptr_t)data, 0);
  301. if (do_execute) {
  302. /* Note that all copies I've sent to neighbours are now invalid */
  303. int n, size;
  304. MPI_Comm_size(comm, &size);
  305. for(n=0 ; n<size ; n++) {
  306. void *already_sent = _starpu_htbl_search_32(sent_data[n], key);
  307. if (already_sent) {
  308. _STARPU_MPI_DEBUG("Posting request to clear send cache for data %p\n", data);
  309. _starpu_mpi_clear_cache_request(data, n, _STARPU_MPI_CLEAR_SENT_DATA);
  310. }
  311. }
  312. }
  313. else {
  314. int mpi_rank = starpu_data_get_rank(data);
  315. void *already_received = _starpu_htbl_search_32(received_data[mpi_rank], key);
  316. if (already_received) {
  317. /* Somebody else will write to the data, so discard our cached copy if any */
  318. /* TODO: starpu_mpi could just remember itself. */
  319. _STARPU_MPI_DEBUG("Posting request to clear receive cache for data %p\n", data);
  320. _starpu_mpi_clear_cache_request(data, mpi_rank, _STARPU_MPI_CLEAR_RECEIVED_DATA);
  321. _starpu_data_deallocate(data);
  322. }
  323. }
  324. }
  325. #else
  326. /* We allocated a temporary buffer for the received data, now drop it */
  327. if ((arg_type & STARPU_R) && do_execute) {
  328. int mpi_rank = starpu_data_get_rank(data);
  329. if (mpi_rank != me && mpi_rank != -1) {
  330. _starpu_data_deallocate(data);
  331. }
  332. }
  333. #endif
  334. }
  335. else if (arg_type==STARPU_VALUE) {
  336. va_arg(varg_list, void *);
  337. va_arg(varg_list, size_t);
  338. }
  339. else if (arg_type==STARPU_CALLBACK) {
  340. va_arg(varg_list, void (*)(void *));
  341. }
  342. else if (arg_type==STARPU_CALLBACK_WITH_ARG) {
  343. va_arg(varg_list, void (*)(void *));
  344. va_arg(varg_list, void *);
  345. }
  346. else if (arg_type==STARPU_CALLBACK_ARG) {
  347. va_arg(varg_list, void *);
  348. }
  349. else if (arg_type==STARPU_PRIORITY) {
  350. va_arg(varg_list, int);
  351. }
  352. else if (arg_type==STARPU_EXECUTE_ON_NODE) {
  353. va_arg(varg_list, int);
  354. }
  355. else if (arg_type==STARPU_EXECUTE_ON_DATA) {
  356. va_arg(varg_list, starpu_data_handle_t);
  357. }
  358. }
  359. va_end(varg_list);
  360. _STARPU_MPI_LOG_OUT();
  361. return 0;
  362. }
  363. void starpu_mpi_get_data_on_node(MPI_Comm comm, starpu_data_handle_t data_handle, int node)
  364. {
  365. int me, rank;
  366. rank = starpu_data_get_rank(data_handle);
  367. MPI_Comm_rank(comm, &me);
  368. if (node == rank) return;
  369. if (me == node)
  370. {
  371. starpu_mpi_irecv_detached(data_handle, rank, 42, comm, NULL, NULL);
  372. }
  373. else if (me == rank)
  374. {
  375. starpu_mpi_isend_detached(data_handle, node, 42, comm, NULL, NULL);
  376. }
  377. starpu_task_wait_for_all();
  378. }