fxt-tool.c 23 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940
  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. /*
  18. * Default user options
  19. */
  20. static unsigned per_task_colour = 0;
  21. static unsigned generate_distrib = 0;
  22. static unsigned no_counter = 0;
  23. static unsigned no_bus = 0;
  24. /* TODO don't make that global ? */
  25. struct fxt_ev_64 ev;
  26. /* In case we are going to gather multiple traces (eg in the case of MPI
  27. * processes), we may need to prefix the name of the containers. */
  28. char *prefix = "";
  29. uint64_t offset = 0;
  30. int rank = -1;
  31. static uint64_t start_time = 0;
  32. static uint64_t end_time = 0;
  33. static int nworkers = 0;
  34. //static char *filename = NULL;
  35. /* XXX remove the 64 ... */
  36. unsigned ninputfiles = 0;
  37. static char *filenames[64];
  38. LIST_TYPE(symbol_name,
  39. char *name;
  40. );
  41. static symbol_name_list_t symbol_list;
  42. LIST_TYPE(communication,
  43. unsigned comid;
  44. float comm_start;
  45. float bandwith;
  46. unsigned node;
  47. );
  48. static communication_list_t communication_list;
  49. /*
  50. * Paje trace file tools
  51. */
  52. static char *out_paje_path = "paje.trace";
  53. static FILE *out_paje_file;
  54. static char *distrib_time_path = "distrib.data";
  55. static FILE *distrib_time;
  56. static void paje_output_file_init(void)
  57. {
  58. /* create a new file */
  59. out_paje_file = fopen(out_paje_path, "w+");
  60. write_paje_header(out_paje_file);
  61. fprintf(out_paje_file, " \n \
  62. 1 MPIP 0 \"MPI Program\" \n \
  63. 1 P MPIP \"Program\" \n \
  64. 1 Mn P \"Memory Node\" \n \
  65. 1 T Mn \"Worker\" \n \
  66. 1 Sc P \"Scheduler State\" \n \
  67. 2 event T \"event type\" \n \
  68. 3 S T \"Thread State\" \n \
  69. 3 MS Mn \"Memory Node State\" \n \
  70. 4 ntask Sc \"Number of tasks\" \n \
  71. 4 bw Mn \"Bandwith\" \n \
  72. 6 I S Initializing \"0.0 .7 1.0\" \n \
  73. 6 D S Deinitializing \"0.0 .1 .7\" \n \
  74. 6 Fi S FetchingInput \"1.0 .1 1.0\" \n \
  75. 6 Po S PushingOutput \"0.1 1.0 1.0\" \n \
  76. 6 E S Executing \".0 .6 .4\" \n \
  77. 6 C S Callback \".0 .3 .8\" \n \
  78. 6 B S Blocked \".9 .1 .0\" \n \
  79. 6 P S Progressing \".4 .1 .6\" \n \
  80. 6 A MS Allocating \".4 .1 .0\" \n \
  81. 6 Ar MS AllocatingReuse \".1 .1 .8\" \n \
  82. 6 R MS Reclaiming \".0 .1 .4\" \n \
  83. 6 Co MS DriverCopy \".3 .5 .1\" \n \
  84. 6 No MS Nothing \".0 .0 .0\" \n \
  85. 5 MPIL MPIP P P MPIL\n \
  86. 5 L P Mn Mn L\n");
  87. fprintf(out_paje_file, "7 0.0 MPIroot MPIP 0 root\n");
  88. }
  89. /*
  90. * Generic tools
  91. */
  92. static float get_event_time_stamp(void)
  93. {
  94. return (float)((ev.time-offset)/1000000.0);
  95. }
  96. static int register_worker_id(unsigned long tid)
  97. {
  98. int workerid = nworkers++;
  99. /* create a new key in the htable */
  100. char *tidstr = malloc(16*sizeof(char));
  101. sprintf(tidstr, "%ld", tid);
  102. ENTRY item;
  103. item.key = tidstr;
  104. item.data = (void *)(uintptr_t)workerid;
  105. ENTRY *res;
  106. res = hsearch(item, FIND);
  107. /* only register a thread once */
  108. STARPU_ASSERT(res == NULL);
  109. res = hsearch(item, ENTER);
  110. STARPU_ASSERT(res);
  111. return workerid;
  112. }
  113. static int find_worker_id(unsigned long tid)
  114. {
  115. char tidstr[16];
  116. sprintf(tidstr, "%ld", tid);
  117. ENTRY item;
  118. item.key = tidstr;
  119. item.data = NULL;
  120. ENTRY *res;
  121. res = hsearch(item, FIND);
  122. if (!res)
  123. return -1;
  124. int id = (uintptr_t)(res->data);
  125. return id;
  126. }
  127. /*
  128. * Initialization
  129. */
  130. static void handle_new_mem_node(void)
  131. {
  132. fprintf(out_paje_file, "7 %f %ld Mn %sp %sMEMNODE%ld\n", get_event_time_stamp(), ev.param[0], prefix, prefix, ev.param[0]);
  133. if (!no_bus)
  134. fprintf(out_paje_file, "13 %f bw %sMEMNODE%d 0.0\n", 0.0f, prefix, ev.param[0]);
  135. }
  136. static void handle_worker_init_start(void)
  137. {
  138. /*
  139. arg0 : type of worker (cuda, cpu ..)
  140. arg1 : memory node
  141. arg2 : thread id
  142. */
  143. fprintf(out_paje_file, "7 %f %s%ld T %sMEMNODE%ld %s%ld\n",
  144. get_event_time_stamp(), prefix, ev.param[2], prefix, ev.param[1], prefix, ev.param[2]);
  145. int workerid = register_worker_id(ev.param[2]);
  146. switch (ev.param[0]) {
  147. case FUT_APPS_KEY:
  148. set_next_other_worker_color(workerid);
  149. break;
  150. case FUT_CPU_KEY:
  151. set_next_cpu_worker_color(workerid);
  152. break;
  153. case FUT_CUDA_KEY:
  154. set_next_cuda_worker_color(workerid);
  155. break;
  156. default:
  157. STARPU_ABORT();
  158. }
  159. /* start initialization */
  160. fprintf(out_paje_file, "10 %f S %s%ld I\n",
  161. get_event_time_stamp(), prefix, ev.param[2]);
  162. }
  163. static void handle_worker_init_end(void)
  164. {
  165. fprintf(out_paje_file, "10 %f S %s%ld B\n",
  166. get_event_time_stamp(), prefix, ev.param[0]);
  167. }
  168. static void handle_worker_deinit_start(void)
  169. {
  170. fprintf(out_paje_file, "10 %f S %s%ld D\n",
  171. get_event_time_stamp(), prefix, ev.param[0]);
  172. }
  173. static void handle_worker_deinit_end(void)
  174. {
  175. fprintf(out_paje_file, "8 %f %s%ld T\n",
  176. get_event_time_stamp(), prefix, ev.param[1]);
  177. }
  178. static void create_paje_state_if_not_found(char *name)
  179. {
  180. symbol_name_itor_t itor;
  181. for (itor = symbol_name_list_begin(symbol_list);
  182. itor != symbol_name_list_end(symbol_list);
  183. itor = symbol_name_list_next(itor))
  184. {
  185. if (!strcmp(name, itor->name))
  186. {
  187. /* we found an entry */
  188. return;
  189. }
  190. }
  191. /* it's the first time ... */
  192. symbol_name_t entry = symbol_name_new();
  193. entry->name = malloc(strlen(name));
  194. strcpy(entry->name, name);
  195. symbol_name_list_push_front(symbol_list, entry);
  196. /* choose some colour ... that's disguting yes */
  197. unsigned hash_symbol_red = get_colour_symbol_red(name);
  198. unsigned hash_symbol_green = get_colour_symbol_green(name);
  199. unsigned hash_symbol_blue = get_colour_symbol_blue(name);
  200. fprintf(stderr, "name %s hash red %d green %d blue %d \n", name, hash_symbol_red, hash_symbol_green, hash_symbol_blue);
  201. uint32_t hash_sum = hash_symbol_red + hash_symbol_green + hash_symbol_blue;
  202. float red = (1.0f * hash_symbol_red) / hash_sum;
  203. float green = (1.0f * hash_symbol_green) / hash_sum;
  204. float blue = (1.0f * hash_symbol_blue) / hash_sum;
  205. /* create the Paje state */
  206. fprintf(out_paje_file, "6 %s S %s \"%f %f %f\" \n", name, red, green, blue, name);
  207. }
  208. static double last_codelet_start[MAXWORKERS];
  209. static uint64_t last_codelet_hash[MAXWORKERS];
  210. static char last_codelet_symbol[128][MAXWORKERS];
  211. static void handle_start_codelet_body(void)
  212. {
  213. int worker;
  214. worker = find_worker_id(ev.param[1]);
  215. if (worker < 0) return;
  216. unsigned long has_name = ev.param[2];
  217. char *name = has_name?(char *)&ev.param[3]:"unknown";
  218. snprintf(last_codelet_symbol[worker], 128, "%s", name);
  219. /* TODO */
  220. last_codelet_hash[worker] = 0;
  221. float start_codelet_time = get_event_time_stamp();
  222. last_codelet_start[worker] = start_codelet_time;
  223. if (per_task_colour)
  224. {
  225. create_paje_state_if_not_found(name);
  226. fprintf(out_paje_file, "101 %f S %s%ld E %s\n", start_codelet_time, prefix, ev.param[1], name);
  227. }
  228. else {
  229. fprintf(out_paje_file, "10 %f S %s%ld E\n", start_codelet_time, prefix, ev.param[1]);
  230. }
  231. end_time = STARPU_MAX(end_time, ev.time);
  232. }
  233. static void handle_end_codelet_body(void)
  234. {
  235. int worker;
  236. worker = find_worker_id(ev.param[1]);
  237. if (worker < 0) return;
  238. float end_codelet_time = get_event_time_stamp();
  239. fprintf(out_paje_file, "10 %f S %s%ld B\n", end_codelet_time, prefix, ev.param[1]);
  240. float codelet_length = (end_codelet_time - last_codelet_start[worker]);
  241. if (generate_distrib)
  242. fprintf(distrib_time, "%s\t%s%d\t%lx\t%f\n", last_codelet_symbol[worker],
  243. prefix, worker, last_codelet_hash[worker], codelet_length);
  244. end_time = STARPU_MAX(end_time, ev.time);
  245. }
  246. static void handle_user_event(void)
  247. {
  248. int worker;
  249. unsigned code;
  250. code = ev.param[2];
  251. worker = find_worker_id(ev.param[1]);
  252. if (worker < 0)
  253. {
  254. fprintf(out_paje_file, "9 %f event %sp %d\n", get_event_time_stamp(), prefix, rank);
  255. }
  256. else {
  257. fprintf(out_paje_file, "9 %f event %s%ld %d\n", get_event_time_stamp(), prefix, ev.param[1], code);
  258. }
  259. }
  260. static void handle_start_callback(void)
  261. {
  262. int worker;
  263. worker = find_worker_id(ev.param[1]);
  264. if (worker < 0) return;
  265. fprintf(out_paje_file, "10 %f S %s%ld C\n", get_event_time_stamp(), prefix, ev.param[1] );
  266. }
  267. static void handle_end_callback(void)
  268. {
  269. int worker;
  270. worker = find_worker_id(ev.param[1]);
  271. if (worker < 0) return;
  272. fprintf(out_paje_file, "10 %f S %s%ld B\n", get_event_time_stamp(), prefix, ev.param[1] );
  273. }
  274. static void handle_worker_status(const char *newstatus)
  275. {
  276. int worker;
  277. worker = find_worker_id(ev.param[1]);
  278. if (worker < 0) return;
  279. fprintf(out_paje_file, "10 %f S %s%ld %s\n",
  280. get_event_time_stamp(), prefix, ev.param[1], newstatus);
  281. end_time = STARPU_MAX(end_time, ev.time);
  282. }
  283. static void handle_data_copy(void)
  284. {
  285. }
  286. static void handle_start_driver_copy(void)
  287. {
  288. unsigned src = ev.param[0];
  289. unsigned dst = ev.param[1];
  290. unsigned size = ev.param[2];
  291. unsigned comid = ev.param[3];
  292. if (!no_bus)
  293. {
  294. fprintf(out_paje_file, "10 %f MS %sMEMNODE%d Co\n", get_event_time_stamp(), prefix, dst);
  295. fprintf(out_paje_file, "18 %f L %sp %d %sMEMNODE%d com_%d\n", get_event_time_stamp(), prefix, size, prefix, src, comid);
  296. /* create a structure to store the start of the communication, this will be matched later */
  297. communication_t com = communication_new();
  298. com->comid = comid;
  299. com->comm_start = get_event_time_stamp();
  300. /* that's a hack: either src or dst is non null */
  301. com->node = (src + dst);
  302. communication_list_push_back(communication_list, com);
  303. }
  304. }
  305. static void handle_end_driver_copy(void)
  306. {
  307. unsigned dst = ev.param[1];
  308. unsigned size = ev.param[2];
  309. unsigned comid = ev.param[3];
  310. if (!no_bus)
  311. {
  312. fprintf(out_paje_file, "10 %f MS %sMEMNODE%d No\n", get_event_time_stamp(), prefix, dst);
  313. fprintf(out_paje_file, "19 %f L %sp %d %sMEMNODE%d com_%d\n", get_event_time_stamp(), prefix, size, prefix, dst, comid);
  314. /* look for a data transfer to match */
  315. communication_itor_t itor;
  316. for (itor = communication_list_begin(communication_list);
  317. itor != communication_list_end(communication_list);
  318. itor = communication_list_next(itor))
  319. {
  320. if (itor->comid == comid)
  321. {
  322. float comm_end = get_event_time_stamp();
  323. float bandwith = (float)((0.001*size)/(comm_end - itor->comm_start));
  324. itor->bandwith = bandwith;
  325. communication_t com = communication_new();
  326. com->comid = comid;
  327. com->comm_start = get_event_time_stamp();
  328. com->bandwith = -bandwith;
  329. com->node = itor->node;
  330. communication_list_push_back(communication_list, com);
  331. break;
  332. }
  333. }
  334. }
  335. }
  336. static void display_bandwith_evolution(void)
  337. {
  338. float current_bandwith = 0.0;
  339. float current_bandwith_per_node[32] = {0.0};
  340. communication_itor_t itor;
  341. for (itor = communication_list_begin(communication_list);
  342. itor != communication_list_end(communication_list);
  343. itor = communication_list_next(itor))
  344. {
  345. current_bandwith += itor->bandwith;
  346. fprintf(out_paje_file, "13 %f bw %sMEMNODE0 %f\n",
  347. itor->comm_start, prefix, current_bandwith);
  348. current_bandwith_per_node[itor->node] += itor->bandwith;
  349. fprintf(out_paje_file, "13 %f bw %sMEMNODE%d %f\n",
  350. itor->comm_start, prefix, itor->node, current_bandwith_per_node[itor->node]);
  351. }
  352. }
  353. static void handle_memnode_event(const char *eventstr)
  354. {
  355. unsigned memnode = ev.param[0];
  356. fprintf(out_paje_file, "10 %f MS %sMEMNODE%d %s\n",
  357. get_event_time_stamp(), prefix, memnode, eventstr);
  358. }
  359. /*
  360. * Number of task submitted to the scheduler
  361. */
  362. static int curq_size = 0;
  363. static void handle_job_push(void)
  364. {
  365. curq_size++;
  366. fprintf(out_paje_file, "13 %f ntask %ssched %f\n", get_event_time_stamp(), prefix, (float)curq_size);
  367. }
  368. static void handle_job_pop(void)
  369. {
  370. curq_size--;
  371. fprintf(out_paje_file, "13 %f ntask %ssched %f\n", get_event_time_stamp(), prefix, (float)curq_size);
  372. }
  373. static void handle_codelet_tag_deps(void)
  374. {
  375. uint64_t child;
  376. uint64_t father;
  377. child = ev.param[0];
  378. father = ev.param[1];
  379. add_deps(child, father);
  380. }
  381. static void handle_task_done(void)
  382. {
  383. uint64_t tag_id;
  384. tag_id = ev.param[0];
  385. unsigned long has_name = ev.param[2];
  386. char *name = has_name?(char *)&ev.param[3]:"unknown";
  387. int worker;
  388. worker = find_worker_id(ev.param[1]);
  389. char *colour;
  390. char buffer[32];
  391. if (per_task_colour) {
  392. snprintf(buffer, 32, "%.4f,%.4f,%.4f",
  393. get_colour_symbol_red(name)/1024.0,
  394. get_colour_symbol_green(name)/1024.0,
  395. get_colour_symbol_blue(name)/1024.0);
  396. colour = &buffer[0];
  397. }
  398. else {
  399. colour= (worker < 0)?"0.0,0.0,0.0":get_worker_color(worker);
  400. }
  401. dot_set_tag_done(tag_id, colour);
  402. }
  403. static void handle_mpi_barrier(void)
  404. {
  405. rank = ev.param[0];
  406. /* Add an event in the trace */
  407. fprintf(out_paje_file, "9 %f event %sp %d\n", get_event_time_stamp(), prefix, rank);
  408. }
  409. static void handle_mpi_isend(void)
  410. {
  411. int dest = ev.param[0];
  412. int mpi_tag = ev.param[1];
  413. size_t size = ev.param[2];
  414. float date = get_event_time_stamp();
  415. add_mpi_send_transfer(rank, dest, mpi_tag, size, date);
  416. }
  417. static void handle_mpi_irecv_end(void)
  418. {
  419. int src = ev.param[0];
  420. int mpi_tag = ev.param[1];
  421. float date = get_event_time_stamp();
  422. add_mpi_recv_transfer(src, rank, mpi_tag, date);
  423. }
  424. static void parse_args(int argc, char **argv)
  425. {
  426. /* We want to support arguments such as "fxt-tool -i trace_*" */
  427. unsigned reading_input_filenames = 0;
  428. int i;
  429. for (i = 1; i < argc; i++) {
  430. if (strcmp(argv[i], "-c") == 0) {
  431. per_task_colour = 1;
  432. reading_input_filenames = 0;
  433. continue;
  434. }
  435. if (strcmp(argv[i], "-o") == 0) {
  436. out_paje_path = argv[++i];
  437. reading_input_filenames = 0;
  438. continue;
  439. }
  440. if (strcmp(argv[i], "-i") == 0) {
  441. filenames[ninputfiles++] = argv[++i];
  442. reading_input_filenames = 1;
  443. continue;
  444. }
  445. if (strcmp(argv[i], "-no-counter") == 0) {
  446. no_counter = 1;
  447. reading_input_filenames = 0;
  448. continue;
  449. }
  450. if (strcmp(argv[i], "-no-bus") == 0) {
  451. no_bus = 1;
  452. reading_input_filenames = 0;
  453. continue;
  454. }
  455. if (strcmp(argv[i], "-d") == 0) {
  456. generate_distrib = 1;
  457. reading_input_filenames = 0;
  458. continue;
  459. }
  460. if (strcmp(argv[i], "-h") == 0) {
  461. fprintf(stderr, "Usage : %s [-c] [-no-counter] [-no-bus] [-i input_filename] [-o output_filename]\n", argv[0]);
  462. fprintf(stderr, "\t-c: use a different colour for every type of task.\n");
  463. exit(-1);
  464. }
  465. /* That's pretty dirty: if the reading_input_filenames flag is
  466. * set, and that the argument does not match an option, we
  467. * assume this may be another filename */
  468. if (reading_input_filenames)
  469. {
  470. filenames[ninputfiles++] = argv[i];
  471. continue;
  472. }
  473. }
  474. }
  475. void parse_new_file(char *filename_in, char *file_prefix, uint64_t file_offset)
  476. {
  477. prefix = file_prefix;
  478. offset = file_offset;
  479. /* Open the trace file */
  480. int fd_in;
  481. fd_in = open(filename_in, O_RDONLY);
  482. if (fd_in < 0) {
  483. perror("open failed :");
  484. exit(-1);
  485. }
  486. static fxt_t fut;
  487. fut = fxt_fdopen(fd_in);
  488. if (!fut) {
  489. perror("fxt_fdopen :");
  490. exit(-1);
  491. }
  492. fxt_blockev_t block;
  493. block = fxt_blockev_enter(fut);
  494. /* create a htable to identify each worker(tid) */
  495. hcreate(MAXWORKERS);
  496. symbol_list = symbol_name_list_new();
  497. communication_list = communication_list_new();
  498. /* TODO starttime ...*/
  499. /* create the "program" container */
  500. fprintf(out_paje_file, "7 0.0 %sp P MPIroot program%s \n", prefix, prefix);
  501. /* create a variable with the number of tasks */
  502. if (!no_counter)
  503. {
  504. fprintf(out_paje_file, "7 %f %ssched Sc %sp scheduler \n", 0.0, prefix, prefix);
  505. fprintf(out_paje_file, "13 0.0 ntask %ssched 0.0\n", prefix);
  506. }
  507. unsigned first_event = 1;
  508. while(1) {
  509. int ret = fxt_next_ev(block, FXT_EV_TYPE_64, (struct fxt_ev *)&ev);
  510. if (ret != FXT_EV_OK) {
  511. fprintf(stderr, "no more block ...\n");
  512. break;
  513. }
  514. __attribute__ ((unused)) int nbparam = ev.nb_params;
  515. if (first_event)
  516. {
  517. first_event = 0;
  518. start_time = ev.time;
  519. }
  520. switch (ev.code) {
  521. case FUT_WORKER_INIT_START:
  522. handle_worker_init_start();
  523. break;
  524. case FUT_WORKER_INIT_END:
  525. handle_worker_init_end();
  526. break;
  527. case FUT_NEW_MEM_NODE:
  528. handle_new_mem_node();
  529. break;
  530. /* detect when the workers were idling or not */
  531. case FUT_START_CODELET_BODY:
  532. handle_start_codelet_body();
  533. break;
  534. case FUT_END_CODELET_BODY:
  535. handle_end_codelet_body();
  536. break;
  537. case FUT_START_CALLBACK:
  538. handle_start_callback();
  539. break;
  540. case FUT_END_CALLBACK:
  541. handle_end_callback();
  542. break;
  543. /* monitor stack size */
  544. case FUT_JOB_PUSH:
  545. if (!no_counter)
  546. handle_job_push();
  547. break;
  548. case FUT_JOB_POP:
  549. if (!no_counter)
  550. handle_job_pop();
  551. break;
  552. /* check the memory transfer overhead */
  553. case FUT_START_FETCH_INPUT:
  554. handle_worker_status("Fi");
  555. break;
  556. case FUT_START_PUSH_OUTPUT:
  557. handle_worker_status("Po");
  558. break;
  559. case FUT_START_PROGRESS:
  560. handle_worker_status("P");
  561. break;
  562. case FUT_END_FETCH_INPUT:
  563. case FUT_END_PROGRESS:
  564. case FUT_END_PUSH_OUTPUT:
  565. handle_worker_status("B");
  566. break;
  567. case FUT_CODELET_TAG:
  568. /* XXX */
  569. break;
  570. case FUT_CODELET_TAG_DEPS:
  571. handle_codelet_tag_deps();
  572. break;
  573. case FUT_TASK_DONE:
  574. handle_task_done();
  575. break;
  576. case FUT_DATA_COPY:
  577. if (!no_bus)
  578. handle_data_copy();
  579. break;
  580. case FUT_START_DRIVER_COPY:
  581. if (!no_bus)
  582. handle_start_driver_copy();
  583. break;
  584. case FUT_END_DRIVER_COPY:
  585. if (!no_bus)
  586. handle_end_driver_copy();
  587. break;
  588. case FUT_WORK_STEALING:
  589. /* XXX */
  590. break;
  591. case FUT_WORKER_DEINIT_START:
  592. handle_worker_deinit_start();
  593. break;
  594. case FUT_WORKER_DEINIT_END:
  595. handle_worker_deinit_end();
  596. break;
  597. case FUT_START_ALLOC:
  598. if (!no_bus)
  599. handle_memnode_event("A");
  600. break;
  601. case FUT_START_ALLOC_REUSE:
  602. if (!no_bus)
  603. handle_memnode_event("Ar");
  604. break;
  605. case FUT_START_MEMRECLAIM:
  606. handle_memnode_event("R");
  607. break;
  608. case FUT_END_ALLOC:
  609. case FUT_END_ALLOC_REUSE:
  610. case FUT_END_MEMRECLAIM:
  611. if (!no_bus)
  612. handle_memnode_event("No");
  613. break;
  614. case FUT_USER_EVENT:
  615. handle_user_event();
  616. break;
  617. case FUT_MPI_BARRIER:
  618. handle_mpi_barrier();
  619. break;
  620. case FUT_MPI_ISEND:
  621. handle_mpi_isend();
  622. break;
  623. case FUT_MPI_IRECV_END:
  624. handle_mpi_irecv_end();
  625. break;
  626. default:
  627. fprintf(stderr, "unknown event.. %x at time %llx WITH OFFSET %llx\n",
  628. (unsigned)ev.code, (long long unsigned)ev.time, (long long unsigned)(ev.time-offset));
  629. break;
  630. }
  631. }
  632. hdestroy();
  633. /* Close the trace file */
  634. if (close(fd_in))
  635. {
  636. perror("close failed :");
  637. exit(-1);
  638. }
  639. }
  640. /*
  641. * This program should be used to parse the log generated by FxT
  642. */
  643. int main(int argc, char **argv)
  644. {
  645. int fd_out;
  646. parse_args(argc, argv);
  647. init_dag_dot();
  648. if (generate_distrib)
  649. distrib_time = fopen(distrib_time_path, "w+");
  650. paje_output_file_init();
  651. if (ninputfiles == 1)
  652. {
  653. /* we usually only have a single trace */
  654. uint64_t file_start_time = find_start_time(filenames[0]);
  655. parse_new_file(filenames[0], "", file_start_time);
  656. }
  657. else {
  658. unsigned inputfile;
  659. uint64_t offsets[64];
  660. uint64_t found_offsets[64];
  661. uint64_t start_times[64];
  662. uint64_t max = 0;
  663. /*
  664. * Find the trace offsets:
  665. * - If there is no sync point
  666. * psi_k(x) = x - start_k
  667. * - If there is a sync point sync_k
  668. * psi_k(x) = x - sync_k + M
  669. * where M = max { sync_i - start_i | there exists sync_i}
  670. * More generally:
  671. * - psi_k(x) = x - offset_k
  672. */
  673. int unique_keys[64];
  674. int rank_k[64];
  675. uint64_t start_k[64];
  676. uint64_t sync_k[64];
  677. unsigned sync_k_exists[64];
  678. uint64_t M = 0;
  679. unsigned found_one_sync_point = 0;
  680. int key;
  681. unsigned display_mpi = 0;
  682. /* Compute all start_k */
  683. for (inputfile = 0; inputfile < ninputfiles; inputfile++)
  684. {
  685. uint64_t file_start = find_start_time(filenames[inputfile]);
  686. start_k[inputfile] = file_start;
  687. }
  688. /* Compute all sync_k if they exist */
  689. for (inputfile = 0; inputfile < ninputfiles; inputfile++)
  690. {
  691. int ret = find_sync_point(filenames[inputfile],
  692. &sync_k[inputfile],
  693. &unique_keys[inputfile],
  694. &rank_k[inputfile]);
  695. if (ret == -1)
  696. {
  697. /* There was no sync point, we assume there is no offset */
  698. sync_k_exists[inputfile] = 0;
  699. }
  700. else {
  701. if (!found_one_sync_point)
  702. {
  703. key = unique_keys[inputfile];
  704. display_mpi = 1;
  705. found_one_sync_point = 1;
  706. }
  707. else {
  708. if (key != unique_keys[inputfile])
  709. {
  710. fprintf(stderr, "Warning: traces are coming from different run so we will not try to display MPI communications.\n");
  711. display_mpi = 0;
  712. }
  713. }
  714. STARPU_ASSERT(sync_k[inputfile] >= start_k[inputfile]);
  715. sync_k_exists[inputfile] = 1;
  716. uint64_t diff = sync_k[inputfile] - start_k[inputfile];
  717. if (diff > M)
  718. M = diff;
  719. }
  720. }
  721. /* Compute the offset */
  722. for (inputfile = 0; inputfile < ninputfiles; inputfile++)
  723. {
  724. offsets[inputfile] = sync_k_exists[inputfile]?
  725. (sync_k[inputfile]-M):start_k[inputfile];
  726. }
  727. /* generate the Paje trace for the different files */
  728. for (inputfile = 0; inputfile < ninputfiles; inputfile++)
  729. {
  730. int filerank = rank_k[inputfile];
  731. fprintf(stderr, "Handle file %s (rank %d)\n", filenames[inputfile], filerank);
  732. char file_prefix[32];
  733. snprintf(file_prefix, 32, "mpi_%d_", filerank);
  734. parse_new_file(filenames[inputfile], file_prefix, offsets[inputfile]);
  735. }
  736. /* display the MPI transfers if possible */
  737. if (display_mpi)
  738. for (inputfile = 0; inputfile < ninputfiles; inputfile++)
  739. {
  740. int filerank = rank_k[inputfile];
  741. display_all_transfers_from_trace(out_paje_file, filerank);
  742. }
  743. }
  744. display_bandwith_evolution();
  745. /* close the different files */
  746. fclose(out_paje_file);
  747. if (generate_distrib)
  748. fclose(distrib_time);
  749. terminate_dat_dot();
  750. return 0;
  751. }