Logo AND Algorithmique Numérique Distribuée

Public GIT Repository
Update the MPI_Waitall for SMPI replay tool, remove unnecessary loop
[simgrid.git] / src / smpi / smpi_replay.c
1 /* Copyright (c) 2009, 2010, 2011, 2012. The SimGrid Team.
2  * All rights reserved.                                                     */
3
4 /* This program is free software; you can redistribute it and/or modify it
5  * under the terms of the license (GNU LGPL) which comes with this package. */
6
7
8 #include "private.h"
9 #include <stdio.h>
10 #include <xbt.h>
11 #include <xbt/replay.h>
12
13 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_replay,smpi,"Trace Replay with SMPI");
14
15 int communicator_size = 0;
16 static int active_processes = 0;
17 xbt_dynar_t *reqq;
18
19 static void log_timed_action (const char *const *action, double clock){
20   if (XBT_LOG_ISENABLED(smpi_replay, xbt_log_priority_verbose)){
21     char *name = xbt_str_join_array(action, " ");
22     XBT_VERB("%s %f", name, smpi_process_simulated_elapsed()-clock);
23     free(name);
24   }
25 }
26
27
28 typedef struct {
29   xbt_dynar_t isends; /* of MPI_Request */
30   xbt_dynar_t irecvs; /* of MPI_Request */
31 } s_smpi_replay_globals_t, *smpi_replay_globals_t;
32
33
34 /* Helper function */
35 static double parse_double(const char *string)
36 {
37   double value;
38   char *endptr;
39   value = strtod(string, &endptr);
40   if (*endptr != '\0')
41     THROWF(unknown_error, 0, "%s is not a double", string);
42   return value;
43 }
44
45 static void action_init(const char *const *action)
46 {
47   int i;
48   XBT_DEBUG("Initialize the counters");
49   smpi_replay_globals_t globals =  xbt_new(s_smpi_replay_globals_t, 1);
50   globals->isends = xbt_dynar_new(sizeof(MPI_Request),NULL);
51   globals->irecvs = xbt_dynar_new(sizeof(MPI_Request),NULL);
52   
53   
54   smpi_process_set_user_data((void*) globals);
55
56   /* start a simulated timer */
57   smpi_process_simulated_start();
58   /*initialize the number of active processes */
59   active_processes = smpi_process_count();
60
61   reqq=xbt_new0(xbt_dynar_t,active_processes);
62   
63   for(i=0;i<active_processes;i++){
64     reqq[i]=xbt_dynar_new(sizeof(MPI_Request),NULL);
65   }
66     
67   
68 }
69
70 static void action_finalize(const char *const *action)
71 {
72   smpi_replay_globals_t globals =
73       (smpi_replay_globals_t) smpi_process_get_user_data();
74
75   if (globals){
76     XBT_DEBUG("There are %lu isends and %lu irecvs in the dynars",
77          xbt_dynar_length(globals->isends),xbt_dynar_length(globals->irecvs));
78     xbt_dynar_free_container(&(globals->isends));
79     xbt_dynar_free_container(&(globals->irecvs));
80   }
81   free(globals);
82 }
83
84 static void action_comm_size(const char *const *action)
85 {
86   double clock = smpi_process_simulated_elapsed();
87
88   communicator_size = parse_double(action[2]);
89   log_timed_action (action, clock);
90 }
91
92 static void action_comm_split(const char *const *action)
93 {
94   double clock = smpi_process_simulated_elapsed();
95
96   log_timed_action (action, clock);
97 }
98
99 static void action_comm_dup(const char *const *action)
100 {
101   double clock = smpi_process_simulated_elapsed();
102
103   log_timed_action (action, clock);
104 }
105
106 static void action_compute(const char *const *action)
107 {
108   double clock = smpi_process_simulated_elapsed();
109   smpi_execute_flops(parse_double(action[2]));
110
111   log_timed_action (action, clock);
112 }
113
114 static void action_send(const char *const *action)
115 {
116   int to = atoi(action[2]);
117   double size=parse_double(action[3]);
118   double clock = smpi_process_simulated_elapsed();
119 #ifdef HAVE_TRACING
120   int rank = smpi_comm_rank(MPI_COMM_WORLD);
121   TRACE_smpi_computing_out(rank);
122   int dst_traced = smpi_group_rank(smpi_comm_group(MPI_COMM_WORLD), to);
123   TRACE_smpi_ptp_in(rank, rank, dst_traced, __FUNCTION__);
124   TRACE_smpi_send(rank, rank, dst_traced);
125 #endif
126
127   smpi_mpi_send(NULL, size, MPI_BYTE, to , 0, MPI_COMM_WORLD);
128
129   log_timed_action (action, clock);
130
131   #ifdef HAVE_TRACING
132   TRACE_smpi_ptp_out(rank, rank, dst_traced, __FUNCTION__);
133   TRACE_smpi_computing_in(rank);
134 #endif
135
136 }
137
138 static void action_Isend(const char *const *action)
139 {
140   int to = atoi(action[2]);
141   double size=parse_double(action[3]);
142   double clock = smpi_process_simulated_elapsed();
143   MPI_Request request;
144   smpi_replay_globals_t globals =
145      (smpi_replay_globals_t) smpi_process_get_user_data();
146 #ifdef HAVE_TRACING
147   int rank = smpi_comm_rank(MPI_COMM_WORLD);
148   TRACE_smpi_computing_out(rank);
149   int dst_traced = smpi_group_rank(smpi_comm_group(MPI_COMM_WORLD), to);
150   TRACE_smpi_ptp_in(rank, rank, dst_traced, __FUNCTION__);
151   TRACE_smpi_send(rank, rank, dst_traced);
152 #endif
153
154   request = smpi_mpi_isend(NULL, size, MPI_BYTE, to, 0,MPI_COMM_WORLD);
155   
156 #ifdef HAVE_TRACING
157   TRACE_smpi_ptp_out(rank, rank, dst_traced, __FUNCTION__);
158   request->send = 1;
159   TRACE_smpi_computing_in(rank);
160 #endif
161
162   xbt_dynar_push(globals->isends,&request);
163   xbt_dynar_push(reqq[smpi_comm_rank(MPI_COMM_WORLD)],&request);
164
165   log_timed_action (action, clock);
166 }
167
168 static void action_recv(const char *const *action) {
169   int from = atoi(action[2]);
170   double size=parse_double(action[3]);
171   double clock = smpi_process_simulated_elapsed();
172   MPI_Status status;
173 #ifdef HAVE_TRACING
174   int rank = smpi_comm_rank(MPI_COMM_WORLD);
175   int src_traced = smpi_group_rank(smpi_comm_group(MPI_COMM_WORLD), from);
176   TRACE_smpi_computing_out(rank);
177
178   TRACE_smpi_ptp_in(rank, src_traced, rank, __FUNCTION__);
179 #endif
180
181   smpi_mpi_recv(NULL, size, MPI_BYTE, from, 0, MPI_COMM_WORLD, &status);
182
183 #ifdef HAVE_TRACING
184   TRACE_smpi_ptp_out(rank, src_traced, rank, __FUNCTION__);
185   TRACE_smpi_recv(rank, src_traced, rank);
186   TRACE_smpi_computing_in(rank);
187 #endif
188
189   log_timed_action (action, clock);
190 }
191
192 static void action_Irecv(const char *const *action)
193 {
194   int from = atoi(action[2]);
195   double size=parse_double(action[3]);
196   double clock = smpi_process_simulated_elapsed();
197   MPI_Request request;
198   smpi_replay_globals_t globals =
199      (smpi_replay_globals_t) smpi_process_get_user_data();
200
201 #ifdef HAVE_TRACING
202   int rank = smpi_comm_rank(MPI_COMM_WORLD);
203   int src_traced = smpi_group_rank(smpi_comm_group(MPI_COMM_WORLD), from);
204   TRACE_smpi_ptp_in(rank, src_traced, rank, __FUNCTION__);
205 #endif
206
207   request = smpi_mpi_irecv(NULL, size, MPI_BYTE, from, 0, MPI_COMM_WORLD);
208   
209 #ifdef HAVE_TRACING
210   TRACE_smpi_ptp_out(rank, src_traced, rank, __FUNCTION__);
211   request->recv = 1;
212 #endif
213   xbt_dynar_push(globals->irecvs,&request);
214   xbt_dynar_push(reqq[smpi_comm_rank(MPI_COMM_WORLD)],&request);
215
216   log_timed_action (action, clock);
217 }
218
219 static void action_wait(const char *const *action){
220   double clock = smpi_process_simulated_elapsed();
221   MPI_Request request;
222   MPI_Status status;
223   smpi_replay_globals_t globals =
224       (smpi_replay_globals_t) smpi_process_get_user_data();
225
226   xbt_assert(xbt_dynar_length(globals->irecvs),
227       "action wait not preceded by any irecv: %s",
228       xbt_str_join_array(action," "));
229   request = xbt_dynar_pop_as(globals->irecvs,MPI_Request);
230 #ifdef HAVE_TRACING
231   int rank = request && request->comm != MPI_COMM_NULL
232       ? smpi_comm_rank(request->comm)
233       : -1;
234   TRACE_smpi_computing_out(rank);
235
236   MPI_Group group = smpi_comm_group(request->comm);
237   int src_traced = smpi_group_rank(group, request->src);
238   int dst_traced = smpi_group_rank(group, request->dst);
239   int is_wait_for_receive = request->recv;
240   TRACE_smpi_ptp_in(rank, src_traced, dst_traced, __FUNCTION__);
241 #endif
242   smpi_mpi_wait(&request, &status);
243 #ifdef HAVE_TRACING
244   TRACE_smpi_ptp_out(rank, src_traced, dst_traced, __FUNCTION__);
245   if (is_wait_for_receive) {
246     TRACE_smpi_recv(rank, src_traced, dst_traced);
247   }
248   TRACE_smpi_computing_in(rank);
249 #endif
250
251   log_timed_action (action, clock);
252 }
253
254 static void action_waitall(const char *const *action){
255   double clock = smpi_process_simulated_elapsed();
256   int count_requests=0;
257   unsigned int i=0;
258
259   count_requests=xbt_dynar_length(reqq[smpi_comm_rank(MPI_COMM_WORLD)]);
260
261   if (count_requests>0) {
262     MPI_Request requests[count_requests];
263     MPI_Status status[count_requests];
264   
265     /*  The reqq is an array of dynars. Its index corresponds to the rank.
266      Thus each rank saves its own requests to the array request. */
267     xbt_dynar_foreach(reqq[smpi_comm_rank(MPI_COMM_WORLD)],i,requests[i]); 
268     
269   #ifdef HAVE_TRACING
270    //save information from requests
271  
272    xbt_dynar_t srcs = xbt_dynar_new(sizeof(int), NULL);
273    xbt_dynar_t dsts = xbt_dynar_new(sizeof(int), NULL);
274    xbt_dynar_t recvs = xbt_dynar_new(sizeof(int), NULL);
275    for (i = 0; i < count_requests; i++) {
276     if(requests[i]){
277       int *asrc = xbt_new(int, 1);
278       int *adst = xbt_new(int, 1);
279       int *arecv = xbt_new(int, 1);
280       *asrc = requests[i]->src;
281       *adst = requests[i]->dst;
282       *arecv = requests[i]->recv;
283       xbt_dynar_insert_at(srcs, i, asrc);
284       xbt_dynar_insert_at(dsts, i, adst);
285       xbt_dynar_insert_at(recvs, i, arecv);
286       xbt_free(asrc);
287       xbt_free(adst);
288       xbt_free(arecv);
289     }else {
290       int *t = xbt_new(int, 1);
291       xbt_dynar_insert_at(srcs, i, t);
292       xbt_dynar_insert_at(dsts, i, t);
293       xbt_dynar_insert_at(recvs, i, t);
294       xbt_free(t);
295     }
296    }
297    int rank_traced = smpi_process_index();
298    TRACE_smpi_computing_out(rank_traced);
299
300    TRACE_smpi_ptp_in(rank_traced, -1, -1, __FUNCTION__);
301   #endif
302
303     smpi_mpi_waitall(count_requests, requests, status);
304
305   #ifdef HAVE_TRACING
306    for (i = 0; i < count_requests; i++) {
307     int src_traced, dst_traced, is_wait_for_receive;
308     xbt_dynar_get_cpy(srcs, i, &src_traced);
309     xbt_dynar_get_cpy(dsts, i, &dst_traced);
310     xbt_dynar_get_cpy(recvs, i, &is_wait_for_receive);
311     if (is_wait_for_receive) {
312       TRACE_smpi_recv(rank_traced, src_traced, dst_traced);
313     }
314    }
315    TRACE_smpi_ptp_out(rank_traced, -1, -1, __FUNCTION__);
316    //clean-up of dynars
317    xbt_dynar_free(&srcs);
318    xbt_dynar_free(&dsts);
319    xbt_dynar_free(&recvs);
320    TRACE_smpi_computing_in(rank_traced);
321   #endif
322    
323    xbt_dynar_reset(reqq[smpi_comm_rank(MPI_COMM_WORLD)]);
324   }
325   log_timed_action (action, clock);
326 }
327
328 static void action_barrier(const char *const *action){
329   double clock = smpi_process_simulated_elapsed();
330 #ifdef HAVE_TRACING
331   int rank = smpi_comm_rank(MPI_COMM_WORLD);
332   TRACE_smpi_computing_out(rank);
333   TRACE_smpi_collective_in(rank, -1, __FUNCTION__);
334 #endif
335   smpi_mpi_barrier(MPI_COMM_WORLD);
336 #ifdef HAVE_TRACING
337   TRACE_smpi_collective_out(rank, -1, __FUNCTION__);
338   TRACE_smpi_computing_in(rank);
339 #endif
340
341   log_timed_action (action, clock);
342 }
343
344 static void action_bcast(const char *const *action)
345 {
346   double size = parse_double(action[2]);
347   double clock = smpi_process_simulated_elapsed();
348 #ifdef HAVE_TRACING
349   int rank = smpi_comm_rank(MPI_COMM_WORLD);
350   TRACE_smpi_computing_out(rank);
351   int root_traced = smpi_group_rank(smpi_comm_group(MPI_COMM_WORLD), 0);
352   TRACE_smpi_collective_in(rank, root_traced, __FUNCTION__);
353 #endif
354
355   smpi_mpi_bcast(NULL, size, MPI_BYTE, 0, MPI_COMM_WORLD);
356 #ifdef HAVE_TRACING
357   TRACE_smpi_collective_out(rank, root_traced, __FUNCTION__);
358   TRACE_smpi_computing_in(rank);
359 #endif
360
361   log_timed_action (action, clock);
362 }
363
364 static void action_reduce(const char *const *action)
365 {
366   double size = parse_double(action[2]);
367   double clock = smpi_process_simulated_elapsed();
368 #ifdef HAVE_TRACING
369   int rank = smpi_comm_rank(MPI_COMM_WORLD);
370   TRACE_smpi_computing_out(rank);
371   int root_traced = smpi_group_rank(smpi_comm_group(MPI_COMM_WORLD), 0);
372   TRACE_smpi_collective_in(rank, root_traced, __FUNCTION__);
373 #endif
374    smpi_mpi_reduce(NULL, NULL, size, MPI_BYTE, MPI_OP_NULL, 0, MPI_COMM_WORLD);
375 #ifdef HAVE_TRACING
376   TRACE_smpi_collective_out(rank, root_traced, __FUNCTION__);
377   TRACE_smpi_computing_in(rank);
378 #endif
379
380   log_timed_action (action, clock);
381 }
382
383 static void action_allReduce(const char *const *action) {
384   double comm_size = parse_double(action[2]);
385   double comp_size = parse_double(action[3]);
386   double clock = smpi_process_simulated_elapsed();
387 #ifdef HAVE_TRACING
388   int rank = smpi_comm_rank(MPI_COMM_WORLD);
389   TRACE_smpi_computing_out(rank);
390   TRACE_smpi_collective_in(rank, -1, __FUNCTION__);
391 #endif
392   smpi_mpi_reduce(NULL, NULL, comm_size, MPI_BYTE, MPI_OP_NULL, 0, MPI_COMM_WORLD);
393   smpi_execute_flops(comp_size);
394   smpi_mpi_bcast(NULL, comm_size, MPI_BYTE, 0, MPI_COMM_WORLD);
395 #ifdef HAVE_TRACING
396   TRACE_smpi_collective_out(rank, -1, __FUNCTION__);
397   TRACE_smpi_computing_in(rank);
398 #endif
399
400   log_timed_action (action, clock);
401 }
402
403 static void action_allToAll(const char *const *action) {
404   double clock = smpi_process_simulated_elapsed();
405   int comm_size = smpi_comm_size(MPI_COMM_WORLD);
406   int send_size = atoi(action[2]);
407   int recv_size = atoi(action[3]);
408   void *send = xbt_new0(int, send_size*comm_size);  
409   void *recv = xbt_new0(int, send_size*comm_size);
410
411
412 #ifdef HAVE_TRACING
413   int rank = smpi_process_index();
414   TRACE_smpi_computing_out(rank);
415   TRACE_smpi_collective_in(rank, -1, __FUNCTION__);
416 #endif
417
418   if (send_size < 200 && comm_size > 12) {
419     smpi_coll_tuned_alltoall_bruck(send, send_size, MPI_BYTE,
420                                    recv, recv_size, MPI_BYTE,
421                                    MPI_COMM_WORLD);
422   } else if (send_size < 3000) {
423   
424     smpi_coll_tuned_alltoall_basic_linear(send, send_size, MPI_BYTE,
425                                           recv, recv_size, MPI_BYTE,
426                                           MPI_COMM_WORLD);
427   } else {
428     smpi_coll_tuned_alltoall_pairwise(send, send_size, MPI_BYTE,
429                                       recv, recv_size, MPI_BYTE,
430                                       MPI_COMM_WORLD);
431   }
432
433 #ifdef HAVE_TRACING
434   TRACE_smpi_collective_out(rank, -1, __FUNCTION__);
435   TRACE_smpi_computing_in(rank);
436 #endif
437
438   log_timed_action (action, clock);
439   xbt_free(send);
440   xbt_free(recv);
441 }
442
443 static void action_allToAllv(const char *const *action) {
444   double clock = smpi_process_simulated_elapsed();
445 //  int comm_size = smpi_comm_size(MPI_COMM_WORLD);
446
447 //  PMPI_Alltoallv(NULL, send_size, send_disp,
448 //                   MPI_BYTE, NULL, recv_size,
449 //                   recv_disp, MPI_BYTE, MPI_COMM_WORLD);
450
451    
452   log_timed_action (action, clock);
453   
454 }
455
456 void smpi_replay_init(int *argc, char***argv){
457   PMPI_Init(argc, argv);
458   if (!smpi_process_index()){
459     _xbt_replay_action_init();
460     xbt_replay_action_register("init",       action_init);
461     xbt_replay_action_register("finalize",   action_finalize);
462     xbt_replay_action_register("comm_size",  action_comm_size);
463     xbt_replay_action_register("comm_split", action_comm_split);
464     xbt_replay_action_register("comm_dup",   action_comm_dup);
465     xbt_replay_action_register("send",       action_send);
466     xbt_replay_action_register("Isend",      action_Isend);
467     xbt_replay_action_register("recv",       action_recv);
468     xbt_replay_action_register("Irecv",      action_Irecv);
469     xbt_replay_action_register("wait",       action_wait);
470     xbt_replay_action_register("waitAll",    action_waitall);
471     xbt_replay_action_register("barrier",    action_barrier);
472     xbt_replay_action_register("bcast",      action_bcast);
473     xbt_replay_action_register("reduce",     action_reduce);
474     xbt_replay_action_register("allReduce",  action_allReduce);
475     xbt_replay_action_register("allToAll",   action_allToAll);
476     xbt_replay_action_register("allToAllV",  action_allToAllv);
477     xbt_replay_action_register("compute",    action_compute);
478   }
479
480   xbt_replay_action_runner(*argc, *argv);
481 }
482
483 int smpi_replay_finalize(){
484   double sim_time= 1.;
485   /* One active process will stop. Decrease the counter*/
486   active_processes--;
487   if(!active_processes){
488     /* Last process alive speaking */
489     /* end the simulated timer */
490     xbt_dynar_free(reqq);
491     sim_time = smpi_process_simulated_elapsed();
492     XBT_INFO("Simulation time %g", sim_time);
493     _xbt_replay_action_exit();
494   }
495   return PMPI_Finalize();
496 }