fxt_tool.c 23 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941
  1. /*
  2. * StarPU
  3. * Copyright (C) INRIA 2008-2010 (see AUTHORS file)
  4. *
  5. * This program is free software; you can redistribute it and/or modify
  6. * it under the terms of the GNU Lesser General Public License as published by
  7. * the Free Software Foundation; either version 2.1 of the License, or (at
  8. * your option) any later version.
  9. *
  10. * This program is distributed in the hope that it will be useful, but
  11. * WITHOUT ANY WARRANTY; without even the implied warranty of
  12. * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
  13. *
  14. * See the GNU Lesser General Public License in COPYING.LGPL for more details.
  15. */
  16. #include "fxt_tool.h"
  17. #include <inttypes.h>
  18. /*
  19. * Default user options
  20. */
  21. static unsigned per_task_colour = 0;
  22. static unsigned generate_distrib = 0;
  23. static unsigned no_counter = 0;
  24. static unsigned no_bus = 0;
  25. /* TODO don't make that global ? */
  26. struct fxt_ev_64 ev;
  27. /* In case we are going to gather multiple traces (eg in the case of MPI
  28. * processes), we may need to prefix the name of the containers. */
  29. char *prefix = "";
  30. uint64_t offset = 0;
  31. int rank = -1;
  32. static uint64_t start_time = 0;
  33. static uint64_t end_time = 0;
  34. static int nworkers = 0;
  35. //static char *filename = NULL;
  36. /* XXX remove the 64 ... */
  37. unsigned ninputfiles = 0;
  38. static char *filenames[64];
  39. LIST_TYPE(symbol_name,
  40. char *name;
  41. );
  42. static symbol_name_list_t symbol_list;
  43. LIST_TYPE(communication,
  44. unsigned comid;
  45. float comm_start;
  46. float bandwidth;
  47. unsigned node;
  48. );
  49. static communication_list_t communication_list;
  50. /*
  51. * Paje trace file tools
  52. */
  53. static char *out_paje_path = "paje.trace";
  54. static FILE *out_paje_file;
  55. static char *distrib_time_path = "distrib.data";
  56. static FILE *distrib_time;
  57. static void paje_output_file_init(void)
  58. {
  59. /* create a new file */
  60. out_paje_file = fopen(out_paje_path, "w+");
  61. write_paje_header(out_paje_file);
  62. fprintf(out_paje_file, " \n \
  63. 1 MPIP 0 \"MPI Program\" \n \
  64. 1 P MPIP \"Program\" \n \
  65. 1 Mn P \"Memory Node\" \n \
  66. 1 T Mn \"Worker\" \n \
  67. 1 Sc P \"Scheduler State\" \n \
  68. 2 event T \"event type\" \n \
  69. 3 S T \"Thread State\" \n \
  70. 3 MS Mn \"Memory Node State\" \n \
  71. 4 ntask Sc \"Number of tasks\" \n \
  72. 4 bw Mn \"Bandwidth\" \n \
  73. 6 I S Initializing \"0.0 .7 1.0\" \n \
  74. 6 D S Deinitializing \"0.0 .1 .7\" \n \
  75. 6 Fi S FetchingInput \"1.0 .1 1.0\" \n \
  76. 6 Po S PushingOutput \"0.1 1.0 1.0\" \n \
  77. 6 E S Executing \".0 .6 .4\" \n \
  78. 6 C S Callback \".0 .3 .8\" \n \
  79. 6 B S Blocked \".9 .1 .0\" \n \
  80. 6 P S Progressing \".4 .1 .6\" \n \
  81. 6 A MS Allocating \".4 .1 .0\" \n \
  82. 6 Ar MS AllocatingReuse \".1 .1 .8\" \n \
  83. 6 R MS Reclaiming \".0 .1 .4\" \n \
  84. 6 Co MS DriverCopy \".3 .5 .1\" \n \
  85. 6 No MS Nothing \".0 .0 .0\" \n \
  86. 5 MPIL MPIP P P MPIL\n \
  87. 5 L P Mn Mn L\n");
  88. fprintf(out_paje_file, "7 0.0 MPIroot MPIP 0 root\n");
  89. }
  90. /*
  91. * Generic tools
  92. */
  93. static float get_event_time_stamp(void)
  94. {
  95. return (float)((ev.time-offset)/1000000.0);
  96. }
  97. static int register_worker_id(unsigned long tid)
  98. {
  99. int workerid = nworkers++;
  100. /* create a new key in the htable */
  101. char *tidstr = malloc(16*sizeof(char));
  102. sprintf(tidstr, "%ld", tid);
  103. ENTRY item;
  104. item.key = tidstr;
  105. item.data = (void *)(uintptr_t)workerid;
  106. ENTRY *res;
  107. res = hsearch(item, FIND);
  108. /* only register a thread once */
  109. STARPU_ASSERT(res == NULL);
  110. res = hsearch(item, ENTER);
  111. STARPU_ASSERT(res);
  112. return workerid;
  113. }
  114. static int find_worker_id(unsigned long tid)
  115. {
  116. char tidstr[16];
  117. sprintf(tidstr, "%ld", tid);
  118. ENTRY item;
  119. item.key = tidstr;
  120. item.data = NULL;
  121. ENTRY *res;
  122. res = hsearch(item, FIND);
  123. if (!res)
  124. return -1;
  125. int id = (uintptr_t)(res->data);
  126. return id;
  127. }
  128. /*
  129. * Initialization
  130. */
  131. static void handle_new_mem_node(void)
  132. {
  133. fprintf(out_paje_file, "7 %f %"PRIu64" Mn %sp %sMEMNODE%"PRIu64"\n", get_event_time_stamp(), ev.param[0], prefix, prefix, ev.param[0]);
  134. if (!no_bus)
  135. fprintf(out_paje_file, "13 %f bw %sMEMNODE%"PRIu64" 0.0\n", 0.0f, prefix, ev.param[0]);
  136. }
  137. static void handle_worker_init_start(void)
  138. {
  139. /*
  140. arg0 : type of worker (cuda, cpu ..)
  141. arg1 : memory node
  142. arg2 : thread id
  143. */
  144. fprintf(out_paje_file, "7 %f %s%"PRIu64" T %sMEMNODE%"PRIu64" %s%"PRIu64"\n",
  145. get_event_time_stamp(), prefix, ev.param[2], prefix, ev.param[1], prefix, ev.param[2]);
  146. int workerid = register_worker_id(ev.param[2]);
  147. switch (ev.param[0]) {
  148. case STARPU_FUT_APPS_KEY:
  149. set_next_other_worker_color(workerid);
  150. break;
  151. case STARPU_FUT_CPU_KEY:
  152. set_next_cpu_worker_color(workerid);
  153. break;
  154. case STARPU_FUT_CUDA_KEY:
  155. set_next_cuda_worker_color(workerid);
  156. break;
  157. default:
  158. STARPU_ABORT();
  159. }
  160. /* start initialization */
  161. fprintf(out_paje_file, "10 %f S %s%"PRIu64" I\n",
  162. get_event_time_stamp(), prefix, ev.param[2]);
  163. }
  164. static void handle_worker_init_end(void)
  165. {
  166. fprintf(out_paje_file, "10 %f S %s%"PRIu64" B\n",
  167. get_event_time_stamp(), prefix, ev.param[0]);
  168. }
  169. static void handle_worker_deinit_start(void)
  170. {
  171. fprintf(out_paje_file, "10 %f S %s%"PRIu64" D\n",
  172. get_event_time_stamp(), prefix, ev.param[0]);
  173. }
  174. static void handle_worker_deinit_end(void)
  175. {
  176. fprintf(out_paje_file, "8 %f %s%"PRIu64" T\n",
  177. get_event_time_stamp(), prefix, ev.param[1]);
  178. }
  179. static void create_paje_state_if_not_found(char *name)
  180. {
  181. symbol_name_itor_t itor;
  182. for (itor = symbol_name_list_begin(symbol_list);
  183. itor != symbol_name_list_end(symbol_list);
  184. itor = symbol_name_list_next(itor))
  185. {
  186. if (!strcmp(name, itor->name))
  187. {
  188. /* we found an entry */
  189. return;
  190. }
  191. }
  192. /* it's the first time ... */
  193. symbol_name_t entry = symbol_name_new();
  194. entry->name = malloc(strlen(name));
  195. strcpy(entry->name, name);
  196. symbol_name_list_push_front(symbol_list, entry);
  197. /* choose some colour ... that's disguting yes */
  198. unsigned hash_symbol_red = get_colour_symbol_red(name);
  199. unsigned hash_symbol_green = get_colour_symbol_green(name);
  200. unsigned hash_symbol_blue = get_colour_symbol_blue(name);
  201. fprintf(stderr, "name %s hash red %d green %d blue %d \n", name, hash_symbol_red, hash_symbol_green, hash_symbol_blue);
  202. uint32_t hash_sum = hash_symbol_red + hash_symbol_green + hash_symbol_blue;
  203. float red = (1.0f * hash_symbol_red) / hash_sum;
  204. float green = (1.0f * hash_symbol_green) / hash_sum;
  205. float blue = (1.0f * hash_symbol_blue) / hash_sum;
  206. /* create the Paje state */
  207. fprintf(out_paje_file, "6 %s S %s \"%f %f %f\" \n", name, name, red, green, blue);
  208. }
  209. static double last_codelet_start[MAXWORKERS];
  210. static uint64_t last_codelet_hash[MAXWORKERS];
  211. static char last_codelet_symbol[128][MAXWORKERS];
  212. static void handle_start_codelet_body(void)
  213. {
  214. int worker;
  215. worker = find_worker_id(ev.param[1]);
  216. if (worker < 0) return;
  217. unsigned long has_name = ev.param[2];
  218. char *name = has_name?(char *)&ev.param[3]:"unknown";
  219. snprintf(last_codelet_symbol[worker], 128, "%s", name);
  220. /* TODO */
  221. last_codelet_hash[worker] = 0;
  222. float start_codelet_time = get_event_time_stamp();
  223. last_codelet_start[worker] = start_codelet_time;
  224. if (per_task_colour)
  225. {
  226. create_paje_state_if_not_found(name);
  227. fprintf(out_paje_file, "101 %f S %s%"PRIu64" E %s\n", start_codelet_time, prefix, ev.param[1], name);
  228. }
  229. else {
  230. fprintf(out_paje_file, "10 %f S %s%"PRIu64" E\n", start_codelet_time, prefix, ev.param[1]);
  231. }
  232. end_time = STARPU_MAX(end_time, ev.time);
  233. }
  234. static void handle_end_codelet_body(void)
  235. {
  236. int worker;
  237. worker = find_worker_id(ev.param[1]);
  238. if (worker < 0) return;
  239. float end_codelet_time = get_event_time_stamp();
  240. fprintf(out_paje_file, "10 %f S %s%"PRIu64" B\n", end_codelet_time, prefix, ev.param[1]);
  241. float codelet_length = (end_codelet_time - last_codelet_start[worker]);
  242. if (generate_distrib)
  243. fprintf(distrib_time, "%s\t%s%d\t%"PRIx64"\t%f\n", last_codelet_symbol[worker],
  244. prefix, worker, last_codelet_hash[worker], codelet_length);
  245. end_time = STARPU_MAX(end_time, ev.time);
  246. }
  247. static void handle_user_event(void)
  248. {
  249. int worker;
  250. unsigned code;
  251. code = ev.param[2];
  252. worker = find_worker_id(ev.param[1]);
  253. if (worker < 0)
  254. {
  255. fprintf(out_paje_file, "9 %f event %sp %d\n", get_event_time_stamp(), prefix, rank);
  256. }
  257. else {
  258. fprintf(out_paje_file, "9 %f event %s%"PRIu64" %d\n", get_event_time_stamp(), prefix, ev.param[1], code);
  259. }
  260. }
  261. static void handle_start_callback(void)
  262. {
  263. int worker;
  264. worker = find_worker_id(ev.param[1]);
  265. if (worker < 0) return;
  266. fprintf(out_paje_file, "10 %f S %s%"PRIu64" C\n", get_event_time_stamp(), prefix, ev.param[1] );
  267. }
  268. static void handle_end_callback(void)
  269. {
  270. int worker;
  271. worker = find_worker_id(ev.param[1]);
  272. if (worker < 0) return;
  273. fprintf(out_paje_file, "10 %f S %s%"PRIu64" B\n", get_event_time_stamp(), prefix, ev.param[1] );
  274. }
  275. static void handle_worker_status(const char *newstatus)
  276. {
  277. int worker;
  278. worker = find_worker_id(ev.param[1]);
  279. if (worker < 0) return;
  280. fprintf(out_paje_file, "10 %f S %s%"PRIu64" %s\n",
  281. get_event_time_stamp(), prefix, ev.param[1], newstatus);
  282. end_time = STARPU_MAX(end_time, ev.time);
  283. }
  284. static void handle_data_copy(void)
  285. {
  286. }
  287. static void handle_start_driver_copy(void)
  288. {
  289. unsigned src = ev.param[0];
  290. unsigned dst = ev.param[1];
  291. unsigned size = ev.param[2];
  292. unsigned comid = ev.param[3];
  293. if (!no_bus)
  294. {
  295. fprintf(out_paje_file, "10 %f MS %sMEMNODE%d Co\n", get_event_time_stamp(), prefix, dst);
  296. fprintf(out_paje_file, "18 %f L %sp %d %sMEMNODE%d com_%d\n", get_event_time_stamp(), prefix, size, prefix, src, comid);
  297. /* create a structure to store the start of the communication, this will be matched later */
  298. communication_t com = communication_new();
  299. com->comid = comid;
  300. com->comm_start = get_event_time_stamp();
  301. /* that's a hack: either src or dst is non null */
  302. com->node = (src + dst);
  303. communication_list_push_back(communication_list, com);
  304. }
  305. }
  306. static void handle_end_driver_copy(void)
  307. {
  308. unsigned dst = ev.param[1];
  309. unsigned size = ev.param[2];
  310. unsigned comid = ev.param[3];
  311. if (!no_bus)
  312. {
  313. fprintf(out_paje_file, "10 %f MS %sMEMNODE%d No\n", get_event_time_stamp(), prefix, dst);
  314. fprintf(out_paje_file, "19 %f L %sp %d %sMEMNODE%d com_%d\n", get_event_time_stamp(), prefix, size, prefix, dst, comid);
  315. /* look for a data transfer to match */
  316. communication_itor_t itor;
  317. for (itor = communication_list_begin(communication_list);
  318. itor != communication_list_end(communication_list);
  319. itor = communication_list_next(itor))
  320. {
  321. if (itor->comid == comid)
  322. {
  323. float comm_end = get_event_time_stamp();
  324. float bandwidth = (float)((0.001*size)/(comm_end - itor->comm_start));
  325. itor->bandwidth = bandwidth;
  326. communication_t com = communication_new();
  327. com->comid = comid;
  328. com->comm_start = get_event_time_stamp();
  329. com->bandwidth = -bandwidth;
  330. com->node = itor->node;
  331. communication_list_push_back(communication_list, com);
  332. break;
  333. }
  334. }
  335. }
  336. }
  337. static void display_bandwidth_evolution(void)
  338. {
  339. float current_bandwidth = 0.0;
  340. float current_bandwidth_per_node[32] = {0.0};
  341. communication_itor_t itor;
  342. for (itor = communication_list_begin(communication_list);
  343. itor != communication_list_end(communication_list);
  344. itor = communication_list_next(itor))
  345. {
  346. current_bandwidth += itor->bandwidth;
  347. fprintf(out_paje_file, "13 %f bw %sMEMNODE0 %f\n",
  348. itor->comm_start, prefix, current_bandwidth);
  349. current_bandwidth_per_node[itor->node] += itor->bandwidth;
  350. fprintf(out_paje_file, "13 %f bw %sMEMNODE%d %f\n",
  351. itor->comm_start, prefix, itor->node, current_bandwidth_per_node[itor->node]);
  352. }
  353. }
  354. static void handle_memnode_event(const char *eventstr)
  355. {
  356. unsigned memnode = ev.param[0];
  357. fprintf(out_paje_file, "10 %f MS %sMEMNODE%d %s\n",
  358. get_event_time_stamp(), prefix, memnode, eventstr);
  359. }
  360. /*
  361. * Number of task submitted to the scheduler
  362. */
  363. static int curq_size = 0;
  364. static void handle_job_push(void)
  365. {
  366. curq_size++;
  367. fprintf(out_paje_file, "13 %f ntask %ssched %f\n", get_event_time_stamp(), prefix, (float)curq_size);
  368. }
  369. static void handle_job_pop(void)
  370. {
  371. curq_size--;
  372. fprintf(out_paje_file, "13 %f ntask %ssched %f\n", get_event_time_stamp(), prefix, (float)curq_size);
  373. }
  374. static void handle_codelet_tag_deps(void)
  375. {
  376. uint64_t child;
  377. uint64_t father;
  378. child = ev.param[0];
  379. father = ev.param[1];
  380. add_deps(child, father);
  381. }
  382. static void handle_task_done(void)
  383. {
  384. uint64_t tag_id;
  385. tag_id = ev.param[0];
  386. unsigned long has_name = ev.param[2];
  387. char *name = has_name?(char *)&ev.param[3]:"unknown";
  388. int worker;
  389. worker = find_worker_id(ev.param[1]);
  390. char *colour;
  391. char buffer[32];
  392. if (per_task_colour) {
  393. snprintf(buffer, 32, "%.4f,%.4f,%.4f",
  394. get_colour_symbol_red(name)/1024.0,
  395. get_colour_symbol_green(name)/1024.0,
  396. get_colour_symbol_blue(name)/1024.0);
  397. colour = &buffer[0];
  398. }
  399. else {
  400. colour= (worker < 0)?"0.0,0.0,0.0":get_worker_color(worker);
  401. }
  402. dot_set_tag_done(tag_id, colour);
  403. }
  404. static void handle_mpi_barrier(void)
  405. {
  406. rank = ev.param[0];
  407. /* Add an event in the trace */
  408. fprintf(out_paje_file, "9 %f event %sp %d\n", get_event_time_stamp(), prefix, rank);
  409. }
  410. static void handle_mpi_isend(void)
  411. {
  412. int dest = ev.param[0];
  413. int mpi_tag = ev.param[1];
  414. size_t size = ev.param[2];
  415. float date = get_event_time_stamp();
  416. add_mpi_send_transfer(rank, dest, mpi_tag, size, date);
  417. }
  418. static void handle_mpi_irecv_end(void)
  419. {
  420. int src = ev.param[0];
  421. int mpi_tag = ev.param[1];
  422. float date = get_event_time_stamp();
  423. add_mpi_recv_transfer(src, rank, mpi_tag, date);
  424. }
  425. static void parse_args(int argc, char **argv)
  426. {
  427. /* We want to support arguments such as "fxt_tool -i trace_*" */
  428. unsigned reading_input_filenames = 0;
  429. int i;
  430. for (i = 1; i < argc; i++) {
  431. if (strcmp(argv[i], "-c") == 0) {
  432. per_task_colour = 1;
  433. reading_input_filenames = 0;
  434. continue;
  435. }
  436. if (strcmp(argv[i], "-o") == 0) {
  437. out_paje_path = argv[++i];
  438. reading_input_filenames = 0;
  439. continue;
  440. }
  441. if (strcmp(argv[i], "-i") == 0) {
  442. filenames[ninputfiles++] = argv[++i];
  443. reading_input_filenames = 1;
  444. continue;
  445. }
  446. if (strcmp(argv[i], "-no-counter") == 0) {
  447. no_counter = 1;
  448. reading_input_filenames = 0;
  449. continue;
  450. }
  451. if (strcmp(argv[i], "-no-bus") == 0) {
  452. no_bus = 1;
  453. reading_input_filenames = 0;
  454. continue;
  455. }
  456. if (strcmp(argv[i], "-d") == 0) {
  457. generate_distrib = 1;
  458. reading_input_filenames = 0;
  459. continue;
  460. }
  461. if (strcmp(argv[i], "-h") == 0) {
  462. fprintf(stderr, "Usage : %s [-c] [-no-counter] [-no-bus] [-i input_filename] [-o output_filename]\n", argv[0]);
  463. fprintf(stderr, "\t-c: use a different colour for every type of task.\n");
  464. exit(-1);
  465. }
  466. /* That's pretty dirty: if the reading_input_filenames flag is
  467. * set, and that the argument does not match an option, we
  468. * assume this may be another filename */
  469. if (reading_input_filenames)
  470. {
  471. filenames[ninputfiles++] = argv[i];
  472. continue;
  473. }
  474. }
  475. }
  476. void parse_new_file(char *filename_in, char *file_prefix, uint64_t file_offset)
  477. {
  478. prefix = file_prefix;
  479. offset = file_offset;
  480. /* Open the trace file */
  481. int fd_in;
  482. fd_in = open(filename_in, O_RDONLY);
  483. if (fd_in < 0) {
  484. perror("open failed :");
  485. exit(-1);
  486. }
  487. static fxt_t fut;
  488. fut = fxt_fdopen(fd_in);
  489. if (!fut) {
  490. perror("fxt_fdopen :");
  491. exit(-1);
  492. }
  493. fxt_blockev_t block;
  494. block = fxt_blockev_enter(fut);
  495. /* create a htable to identify each worker(tid) */
  496. hcreate(MAXWORKERS);
  497. symbol_list = symbol_name_list_new();
  498. communication_list = communication_list_new();
  499. /* TODO starttime ...*/
  500. /* create the "program" container */
  501. fprintf(out_paje_file, "7 0.0 %sp P MPIroot program%s \n", prefix, prefix);
  502. /* create a variable with the number of tasks */
  503. if (!no_counter)
  504. {
  505. fprintf(out_paje_file, "7 %f %ssched Sc %sp scheduler \n", 0.0, prefix, prefix);
  506. fprintf(out_paje_file, "13 0.0 ntask %ssched 0.0\n", prefix);
  507. }
  508. unsigned first_event = 1;
  509. while(1) {
  510. int ret = fxt_next_ev(block, FXT_EV_TYPE_64, (struct fxt_ev *)&ev);
  511. if (ret != FXT_EV_OK) {
  512. fprintf(stderr, "no more block ...\n");
  513. break;
  514. }
  515. __attribute__ ((unused)) int nbparam = ev.nb_params;
  516. if (first_event)
  517. {
  518. first_event = 0;
  519. start_time = ev.time;
  520. }
  521. switch (ev.code) {
  522. case STARPU_FUT_WORKER_INIT_START:
  523. handle_worker_init_start();
  524. break;
  525. case STARPU_FUT_WORKER_INIT_END:
  526. handle_worker_init_end();
  527. break;
  528. case STARPU_FUT_NEW_MEM_NODE:
  529. handle_new_mem_node();
  530. break;
  531. /* detect when the workers were idling or not */
  532. case STARPU_FUT_START_CODELET_BODY:
  533. handle_start_codelet_body();
  534. break;
  535. case STARPU_FUT_END_CODELET_BODY:
  536. handle_end_codelet_body();
  537. break;
  538. case STARPU_FUT_START_CALLBACK:
  539. handle_start_callback();
  540. break;
  541. case STARPU_FUT_END_CALLBACK:
  542. handle_end_callback();
  543. break;
  544. /* monitor stack size */
  545. case STARPU_FUT_JOB_PUSH:
  546. if (!no_counter)
  547. handle_job_push();
  548. break;
  549. case STARPU_FUT_JOB_POP:
  550. if (!no_counter)
  551. handle_job_pop();
  552. break;
  553. /* check the memory transfer overhead */
  554. case STARPU_FUT_START_FETCH_INPUT:
  555. handle_worker_status("Fi");
  556. break;
  557. case STARPU_FUT_START_PUSH_OUTPUT:
  558. handle_worker_status("Po");
  559. break;
  560. case STARPU_FUT_START_PROGRESS:
  561. handle_worker_status("P");
  562. break;
  563. case STARPU_FUT_END_FETCH_INPUT:
  564. case STARPU_FUT_END_PROGRESS:
  565. case STARPU_FUT_END_PUSH_OUTPUT:
  566. handle_worker_status("B");
  567. break;
  568. case STARPU_FUT_CODELET_TAG:
  569. /* XXX */
  570. break;
  571. case STARPU_FUT_CODELET_TAG_DEPS:
  572. handle_codelet_tag_deps();
  573. break;
  574. case STARPU_FUT_TASK_DONE:
  575. handle_task_done();
  576. break;
  577. case STARPU_FUT_DATA_COPY:
  578. if (!no_bus)
  579. handle_data_copy();
  580. break;
  581. case STARPU_FUT_START_DRIVER_COPY:
  582. if (!no_bus)
  583. handle_start_driver_copy();
  584. break;
  585. case STARPU_FUT_END_DRIVER_COPY:
  586. if (!no_bus)
  587. handle_end_driver_copy();
  588. break;
  589. case STARPU_FUT_WORK_STEALING:
  590. /* XXX */
  591. break;
  592. case STARPU_FUT_WORKER_DEINIT_START:
  593. handle_worker_deinit_start();
  594. break;
  595. case STARPU_FUT_WORKER_DEINIT_END:
  596. handle_worker_deinit_end();
  597. break;
  598. case STARPU_FUT_START_ALLOC:
  599. if (!no_bus)
  600. handle_memnode_event("A");
  601. break;
  602. case STARPU_FUT_START_ALLOC_REUSE:
  603. if (!no_bus)
  604. handle_memnode_event("Ar");
  605. break;
  606. case STARPU_FUT_START_MEMRECLAIM:
  607. handle_memnode_event("R");
  608. break;
  609. case STARPU_FUT_END_ALLOC:
  610. case STARPU_FUT_END_ALLOC_REUSE:
  611. case STARPU_FUT_END_MEMRECLAIM:
  612. if (!no_bus)
  613. handle_memnode_event("No");
  614. break;
  615. case STARPU_FUT_USER_EVENT:
  616. handle_user_event();
  617. break;
  618. case FUT_MPI_BARRIER:
  619. handle_mpi_barrier();
  620. break;
  621. case FUT_MPI_ISEND:
  622. handle_mpi_isend();
  623. break;
  624. case FUT_MPI_IRECV_END:
  625. handle_mpi_irecv_end();
  626. break;
  627. default:
  628. fprintf(stderr, "unknown event.. %x at time %llx WITH OFFSET %llx\n",
  629. (unsigned)ev.code, (long long unsigned)ev.time, (long long unsigned)(ev.time-offset));
  630. break;
  631. }
  632. }
  633. hdestroy();
  634. /* Close the trace file */
  635. if (close(fd_in))
  636. {
  637. perror("close failed :");
  638. exit(-1);
  639. }
  640. }
  641. /*
  642. * This program should be used to parse the log generated by FxT
  643. */
  644. int main(int argc, char **argv)
  645. {
  646. int fd_out;
  647. parse_args(argc, argv);
  648. init_dag_dot();
  649. if (generate_distrib)
  650. distrib_time = fopen(distrib_time_path, "w+");
  651. paje_output_file_init();
  652. if (ninputfiles == 1)
  653. {
  654. /* we usually only have a single trace */
  655. uint64_t file_start_time = find_start_time(filenames[0]);
  656. parse_new_file(filenames[0], "", file_start_time);
  657. }
  658. else {
  659. unsigned inputfile;
  660. uint64_t offsets[64];
  661. uint64_t found_offsets[64];
  662. uint64_t start_times[64];
  663. uint64_t max = 0;
  664. /*
  665. * Find the trace offsets:
  666. * - If there is no sync point
  667. * psi_k(x) = x - start_k
  668. * - If there is a sync point sync_k
  669. * psi_k(x) = x - sync_k + M
  670. * where M = max { sync_i - start_i | there exists sync_i}
  671. * More generally:
  672. * - psi_k(x) = x - offset_k
  673. */
  674. int unique_keys[64];
  675. int rank_k[64];
  676. uint64_t start_k[64];
  677. uint64_t sync_k[64];
  678. unsigned sync_k_exists[64];
  679. uint64_t M = 0;
  680. unsigned found_one_sync_point = 0;
  681. int key;
  682. unsigned display_mpi = 0;
  683. /* Compute all start_k */
  684. for (inputfile = 0; inputfile < ninputfiles; inputfile++)
  685. {
  686. uint64_t file_start = find_start_time(filenames[inputfile]);
  687. start_k[inputfile] = file_start;
  688. }
  689. /* Compute all sync_k if they exist */
  690. for (inputfile = 0; inputfile < ninputfiles; inputfile++)
  691. {
  692. int ret = find_sync_point(filenames[inputfile],
  693. &sync_k[inputfile],
  694. &unique_keys[inputfile],
  695. &rank_k[inputfile]);
  696. if (ret == -1)
  697. {
  698. /* There was no sync point, we assume there is no offset */
  699. sync_k_exists[inputfile] = 0;
  700. }
  701. else {
  702. if (!found_one_sync_point)
  703. {
  704. key = unique_keys[inputfile];
  705. display_mpi = 1;
  706. found_one_sync_point = 1;
  707. }
  708. else {
  709. if (key != unique_keys[inputfile])
  710. {
  711. fprintf(stderr, "Warning: traces are coming from different run so we will not try to display MPI communications.\n");
  712. display_mpi = 0;
  713. }
  714. }
  715. STARPU_ASSERT(sync_k[inputfile] >= start_k[inputfile]);
  716. sync_k_exists[inputfile] = 1;
  717. uint64_t diff = sync_k[inputfile] - start_k[inputfile];
  718. if (diff > M)
  719. M = diff;
  720. }
  721. }
  722. /* Compute the offset */
  723. for (inputfile = 0; inputfile < ninputfiles; inputfile++)
  724. {
  725. offsets[inputfile] = sync_k_exists[inputfile]?
  726. (sync_k[inputfile]-M):start_k[inputfile];
  727. }
  728. /* generate the Paje trace for the different files */
  729. for (inputfile = 0; inputfile < ninputfiles; inputfile++)
  730. {
  731. int filerank = rank_k[inputfile];
  732. fprintf(stderr, "Handle file %s (rank %d)\n", filenames[inputfile], filerank);
  733. char file_prefix[32];
  734. snprintf(file_prefix, 32, "mpi_%d_", filerank);
  735. parse_new_file(filenames[inputfile], file_prefix, offsets[inputfile]);
  736. }
  737. /* display the MPI transfers if possible */
  738. if (display_mpi)
  739. for (inputfile = 0; inputfile < ninputfiles; inputfile++)
  740. {
  741. int filerank = rank_k[inputfile];
  742. display_all_transfers_from_trace(out_paje_file, filerank);
  743. }
  744. }
  745. display_bandwidth_evolution();
  746. /* close the different files */
  747. fclose(out_paje_file);
  748. if (generate_distrib)
  749. fclose(distrib_time);
  750. terminate_dat_dot();
  751. return 0;
  752. }