New URL for NEMO forge! http://forge.nemo-ocean.eu

Since March 2022 along with NEMO 4.2 release, the code development moved to a self-hosted GitLab.
This present forge is now archived and remained online for history.

timing.F90 in NEMO/branches/2018/dev_r10164_HPC09_ESIWACE_PREP_MERGE/src/OCE – NEMO

Context Navigation

source: NEMO/branches/2018/dev_r10164_HPC09_ESIWACE_PREP_MERGE/src/OCE/timing.F90 @ 10172

Last change on this file since 10172 was 10172, checked in by smasson, 6 years ago
dev_r10164_HPC09_ESIWACE_PREP_MERGE: action 2b: improve of timing, add computing and waiting time, see #2133
Property svn:keywords set to `Id`
File size: 35.5 KB

Line
1	MODULE timing
2	!!========================================================================
3	!! * MODULE timing *
4	!!========================================================================
5	!! History : 4.0 ! 2001-05 (R. Benshila)
6	!!------------------------------------------------------------------------
7
8	!!------------------------------------------------------------------------
9	!! timming_init : initialize timing process
10	!! timing_start : start Timer
11	!! timing_stop : stop Timer
12	!! timing_reset : end timing variable creation
13	!! timing_finalize : compute stats and write output in calling w*_info
14	!! timing_ini_var : create timing variables
15	!! timing_listing : print instumented subroutines in ocean.output
16	!! wcurrent_info : compute and print detailed stats on the current CPU
17	!! wave_info : compute and print averaged statson all processors
18	!! wmpi_info : compute and write global stats
19	!! supress : suppress an element of the timing linked list
20	!! insert : insert an element of the timing linked list
21	!!------------------------------------------------------------------------
22	USE in_out_manager ! I/O manager
23	USE dom_oce ! ocean domain
24	USE lib_mpp
25
26	IMPLICIT NONE
27	PRIVATE
28
29	PUBLIC timing_init, timing_finalize ! called in nemogcm module
30	PUBLIC timing_reset ! called in step module
31	PUBLIC timing_start, timing_stop ! called in each routine to time
32
33	#if defined key_mpp_mpi
34	INCLUDE 'mpif.h'
35	#endif
36
37	! Variables for fine grain timing
38	TYPE timer
39	CHARACTER(LEN=20) :: cname
40	CHARACTER(LEN=20) :: surname
41	INTEGER :: rank
42	REAL(wp) :: t_cpu, t_clock, tsum_cpu, tsum_clock, tmax_cpu, tmax_clock, tmin_cpu, tmin_clock, tsub_cpu, tsub_clock
43	INTEGER :: ncount, ncount_max, ncount_rate
44	INTEGER :: niter
45	LOGICAL :: l_tdone
46	TYPE(timer), POINTER :: next => NULL()
47	TYPE(timer), POINTER :: prev => NULL()
48	TYPE(timer), POINTER :: parent_section => NULL()
49	END TYPE timer
50
51	TYPE alltimer
52	CHARACTER(LEN=20), DIMENSION(:), POINTER :: cname => NULL()
53	REAL(wp), DIMENSION(:), POINTER :: tsum_cpu => NULL()
54	REAL(wp), DIMENSION(:), POINTER :: tsum_clock => NULL()
55	INTEGER, DIMENSION(:), POINTER :: niter => NULL()
56	TYPE(alltimer), POINTER :: next => NULL()
57	TYPE(alltimer), POINTER :: prev => NULL()
58	END TYPE alltimer
59
60	TYPE(timer), POINTER :: s_timer_root => NULL()
61	TYPE(timer), POINTER :: s_timer => NULL()
62	TYPE(timer), POINTER :: s_timer_old => NULL()
63
64	TYPE(timer), POINTER :: s_wrk => NULL()
65	REAL(wp) :: t_overclock, t_overcpu
66	LOGICAL :: l_initdone = .FALSE.
67	INTEGER :: nsize
68
69	! Variables for coarse grain timing
70	REAL(wp) :: tot_etime, tot_ctime
71	REAL(kind=wp), DIMENSION(2) :: t_elaps, t_cpu
72	REAL(wp), ALLOCATABLE, DIMENSION(:) :: all_etime, all_ctime
73	INTEGER :: nfinal_count, ncount, ncount_rate, ncount_max
74	INTEGER, DIMENSION(8) :: nvalues
75	CHARACTER(LEN=8), DIMENSION(2) :: cdate
76	CHARACTER(LEN=10), DIMENSION(2) :: ctime
77	CHARACTER(LEN=5) :: czone
78
79	! From of ouput file (1/proc or one global) !RB to put in nammpp or namctl
80	LOGICAL :: ln_onefile = .TRUE.
81	LOGICAL :: lwriter
82	!!----------------------------------------------------------------------
83	!! NEMO/OCE 4.0 , NEMO Consortium (2018)
84	!! $Id$
85	!! Software governed by the CeCILL license (see ./LICENSE)
86	!!----------------------------------------------------------------------
87	CONTAINS
88
89	SUBROUTINE timing_start(cdinfo)
90	!!----------------------------------------------------------------------
91	!! * ROUTINE timing_start *
92	!! ** Purpose : collect execution time
93	!!----------------------------------------------------------------------
94	CHARACTER(len=*), INTENT(in) :: cdinfo
95	!
96	IF(ASSOCIATED(s_timer) ) s_timer_old => s_timer
97	!
98	! Create timing structure at first call of the routine
99	CALL timing_ini_var(cdinfo)
100	! write(,) 'after inivar ', s_timer%cname
101
102	! ici timing_ini_var a soit retrouve s_timer et fait return soit ajoute un maillon
103	! maintenant on regarde si le call d'avant corrsspond a un parent ou si il est ferme
104	IF( .NOT. s_timer_old%l_tdone ) THEN
105	s_timer%parent_section => s_timer_old
106	ELSE
107	s_timer%parent_section => NULL()
108	ENDIF
109
110	s_timer%l_tdone = .FALSE.
111	s_timer%niter = s_timer%niter + 1
112	s_timer%t_cpu = 0.
113	s_timer%t_clock = 0.
114
115	! CPU time collection
116	CALL CPU_TIME( s_timer%t_cpu )
117	! clock time collection
118	#if defined key_mpp_mpi
119	s_timer%t_clock= MPI_Wtime()
120	#else
121	CALL SYSTEM_CLOCK(COUNT_RATE=s_timer%ncount_rate, COUNT_MAX=s_timer%ncount_max)
122	CALL SYSTEM_CLOCK(COUNT = s_timer%ncount)
123	#endif
124	! write(,) 'end of start ', s_timer%cname
125
126	!
127	END SUBROUTINE timing_start
128
129
130	SUBROUTINE timing_stop(cdinfo, csection)
131	!!----------------------------------------------------------------------
132	!! * ROUTINE timing_stop *
133	!! ** Purpose : finalize timing and output
134	!!----------------------------------------------------------------------
135	CHARACTER(len=*), INTENT(in) :: cdinfo
136	CHARACTER(len=*), INTENT(in), OPTIONAL :: csection
137	!
138	INTEGER :: ifinal_count, iperiods
139	REAL(wp) :: zcpu_end, zmpitime,zcpu_raw,zclock_raw
140	!
141	s_wrk => NULL()
142
143	! clock time collection
144	#if defined key_mpp_mpi
145	zmpitime = MPI_Wtime()
146	#else
147	CALL SYSTEM_CLOCK(COUNT = ifinal_count)
148	#endif
149	! CPU time collection
150	CALL CPU_TIME( zcpu_end )
151
152	!!$ IF(associated(s_timer%parent_section))then
153	!!$ write(,) s_timer%cname,' <-- ', s_timer%parent_section%cname
154	!!$ ENDIF
155
156	! No need to search ... : s_timer has the last value defined in start
157	! s_timer => s_timer_root
158	! DO WHILE( TRIM(s_timer%cname) /= TRIM(cdinfo) )
159	! IF( ASSOCIATED(s_timer%next) ) s_timer => s_timer%next
160	! END DO
161
162	! CPU time correction
163	zcpu_raw = zcpu_end - s_timer%t_cpu - t_overcpu ! total time including child
164	s_timer%t_cpu = zcpu_raw - s_timer%tsub_cpu
165	! IF(s_timer%cname==trim('lbc_lnk_2d')) write(,) s_timer%tsub_cpu,zcpu_end
166
167	! clock time correction
168	#if defined key_mpp_mpi
169	zclock_raw = zmpitime - s_timer%t_clock - t_overclock ! total time including child
170	s_timer%t_clock = zclock_raw - t_overclock - s_timer%tsub_clock
171	#else
172	iperiods = ifinal_count - s_timer%ncount
173	IF( ifinal_count < s_timer%ncount ) &
174	iperiods = iperiods + s_timer%ncount_max
175	zclock_raw = REAL(iperiods) / s_timer%ncount_rate !- t_overclock
176	s_timer%t_clock = zclock_raw - s_timer%tsub_clock
177	#endif
178	! IF(s_timer%cname==trim('lbc_lnk_2d')) write(,) zclock_raw , s_timer%tsub_clock
179
180	! Correction of parent section
181	IF( .NOT. PRESENT(csection) ) THEN
182	IF ( ASSOCIATED(s_timer%parent_section ) ) THEN
183	s_timer%parent_section%tsub_cpu = zcpu_raw + s_timer%parent_section%tsub_cpu
184	s_timer%parent_section%tsub_clock = zclock_raw + s_timer%parent_section%tsub_clock
185	ENDIF
186	ENDIF
187
188	! time diagnostics
189	s_timer%tsum_clock = s_timer%tsum_clock + s_timer%t_clock
190	s_timer%tsum_cpu = s_timer%tsum_cpu + s_timer%t_cpu
191	!RB to use to get min/max during a time integration
192	! IF( .NOT. l_initdone ) THEN
193	! s_timer%tmin_clock = s_timer%t_clock
194	! s_timer%tmin_cpu = s_timer%t_cpu
195	! ELSE
196	! s_timer%tmin_clock = MIN( s_timer%tmin_clock, s_timer%t_clock )
197	! s_timer%tmin_cpu = MIN( s_timer%tmin_cpu , s_timer%t_cpu )
198	! ENDIF
199	! s_timer%tmax_clock = MAX( s_timer%tmax_clock, s_timer%t_clock )
200	! s_timer%tmax_cpu = MAX( s_timer%tmax_cpu , s_timer%t_cpu )
201	!
202	s_timer%tsub_clock = 0.
203	s_timer%tsub_cpu = 0.
204	s_timer%l_tdone = .TRUE.
205	!
206	!
207	! we come back
208	IF ( ASSOCIATED(s_timer%parent_section ) ) s_timer => s_timer%parent_section
209
210	! write(,) 'end of stop ', s_timer%cname
211
212	END SUBROUTINE timing_stop
213
214
215	SUBROUTINE timing_init
216	!!----------------------------------------------------------------------
217	!! * ROUTINE timing_init *
218	!! ** Purpose : open timing output file
219	!!----------------------------------------------------------------------
220	INTEGER :: iperiods, istart_count, ifinal_count
221	REAL(wp) :: zdum
222	LOGICAL :: ll_f
223
224	IF( ln_onefile ) THEN
225	IF( lwp) CALL ctl_opn( numtime, 'timing.output', 'REPLACE', 'FORMATTED', 'SEQUENTIAL', -1, numout,.TRUE., narea )
226	lwriter = lwp
227	ELSE
228	CALL ctl_opn( numtime, 'timing.output', 'REPLACE', 'FORMATTED', 'SEQUENTIAL', -1, numout,.FALSE., narea )
229	lwriter = .TRUE.
230	ENDIF
231
232	IF( lwriter) THEN
233	WRITE(numtime,*)
234	WRITE(numtime,*) ' CNRS - NERC - Met OFFICE - MERCATOR-ocean - CMCC - INGV'
235	WRITE(numtime,*) ' NEMO team'
236	WRITE(numtime,*) ' Ocean General Circulation Model'
237	WRITE(numtime,*) ' version 3.6 (2015) '
238	WRITE(numtime,*)
239	WRITE(numtime,*) ' Timing Informations '
240	WRITE(numtime,*)
241	WRITE(numtime,*)
242	ENDIF
243
244	! Compute clock function overhead
245	#if defined key_mpp_mpi
246	t_overclock = MPI_WTIME()
247	t_overclock = MPI_WTIME() - t_overclock
248	#else
249	CALL SYSTEM_CLOCK(COUNT_RATE=ncount_rate, COUNT_MAX=ncount_max)
250	CALL SYSTEM_CLOCK(COUNT = istart_count)
251	CALL SYSTEM_CLOCK(COUNT = ifinal_count)
252	iperiods = ifinal_count - istart_count
253	IF( ifinal_count < istart_count ) &
254	iperiods = iperiods + ncount_max
255	t_overclock = REAL(iperiods) / ncount_rate
256	#endif
257
258	! Compute cpu_time function overhead
259	CALL CPU_TIME(zdum)
260	CALL CPU_TIME(t_overcpu)
261
262	! End overhead omputation
263	t_overcpu = t_overcpu - zdum
264	t_overclock = t_overcpu + t_overclock
265	WRITE(,) 't_overcpu ', t_overcpu
266	WRITE(,) 't_overclock ', t_overclock
267	t_overcpu = 0.
268	t_overclock = 0.
269
270	! Timing on date and time
271	CALL DATE_AND_TIME(cdate(1),ctime(1),czone,nvalues)
272
273	CALL CPU_TIME(t_cpu(1))
274	#if defined key_mpp_mpi
275	! Start elapsed and CPU time counters
276	t_elaps(1) = MPI_WTIME()
277	#else
278	CALL SYSTEM_CLOCK(COUNT_RATE=ncount_rate, COUNT_MAX=ncount_max)
279	CALL SYSTEM_CLOCK(COUNT = ncount)
280	#endif
281	!
282	END SUBROUTINE timing_init
283
284
285	SUBROUTINE timing_finalize
286	!!----------------------------------------------------------------------
287	!! * ROUTINE timing_finalize *
288	!! ** Purpose : compute average time
289	!! write timing output file
290	!!----------------------------------------------------------------------
291	TYPE(timer), POINTER :: s_temp
292	INTEGER :: idum, iperiods, icode
293	INTEGER :: ji
294	LOGICAL :: ll_ord, ll_averep
295	CHARACTER(len=120) :: clfmt
296	REAL(wp), DIMENSION(:), ALLOCATABLE :: timing_glob
297	REAL(wp) :: zsypd ! simulated years per day (Balaji 2017)
298	REAL(wp) :: zperc
299
300	ll_averep = .TRUE.
301
302	! total CPU and elapse
303	CALL CPU_TIME(t_cpu(2))
304	t_cpu(2) = t_cpu(2) - t_cpu(1) - t_overcpu
305	#if defined key_mpp_mpi
306	WRITE(,) 't_overclock ', t_overclock
307	WRITE(,) 't_elaps(1) ', t_elaps(1)
308	WRITE(,) 'MPI_WTIME() ', MPI_WTIME()
309	t_elaps(2) = MPI_WTIME() - t_elaps(1) - t_overclock
310	WRITE(,) 't_elaps(2) ', t_elaps(2)
311	#else
312	CALL SYSTEM_CLOCK(COUNT = nfinal_count)
313	iperiods = nfinal_count - ncount
314	IF( nfinal_count < ncount ) &
315	iperiods = iperiods + ncount_max
316	t_elaps(2) = REAL(iperiods) / ncount_rate - t_overclock
317	#endif
318
319	! End of timings on date & time
320	CALL DATE_AND_TIME(cdate(2),ctime(2),czone,nvalues)
321
322	! Compute the numer of routines
323	nsize = 0
324	s_timer => s_timer_root
325	DO WHILE( ASSOCIATED(s_timer) )
326	nsize = nsize + 1
327	s_timer => s_timer%next
328	END DO
329	idum = nsize
330	IF(lk_mpp) CALL mpp_sum(idum)
331	IF( idum/jpnij /= nsize ) THEN
332	IF( lwriter ) WRITE(numtime,*) ' ===> W A R N I N G: '
333	IF( lwriter ) WRITE(numtime,*) ' Some CPU have different number of routines instrumented for timing'
334	IF( lwriter ) WRITE(numtime,*) ' No detailed report on averaged timing can be provided'
335	IF( lwriter ) WRITE(numtime,*) ' The following detailed report only deals with the current processor'
336	IF( lwriter ) WRITE(numtime,*)
337	ll_averep = .FALSE.
338	ENDIF
339
340	#if defined key_mpp_mpi
341	! in MPI gather some info
342	ALLOCATE( all_etime(jpnij), all_ctime(jpnij) )
343	CALL MPI_ALLGATHER(t_elaps(2), 1, MPI_DOUBLE_PRECISION, &
344	all_etime , 1, MPI_DOUBLE_PRECISION, &
345	MPI_COMM_OCE, icode)
346	CALL MPI_ALLGATHER(t_cpu(2) , 1, MPI_DOUBLE_PRECISION, &
347	all_ctime, 1, MPI_DOUBLE_PRECISION, &
348	MPI_COMM_OCE, icode)
349	tot_etime = SUM(all_etime(:))
350	tot_ctime = SUM(all_ctime(:))
351	#else
352	tot_etime = t_elaps(2)
353	tot_ctime = t_cpu (2)
354	#endif
355
356	! write output file
357	IF( lwriter ) WRITE(numtime,*) 'Total timing (sum) :'
358	IF( lwriter ) WRITE(numtime,*) '--------------------'
359	IF( lwriter ) WRITE(numtime,"('Elapsed Time (s) CPU Time (s)')")
360	IF( lwriter ) WRITE(numtime,'(5x,f12.3,1x,f12.3)') tot_etime, tot_ctime
361	IF( lwriter ) WRITE(numtime,*)
362	#if defined key_mpp_mpi
363	IF( ll_averep ) CALL waver_info
364	CALL wmpi_info
365	#endif
366	IF( lwriter ) CALL wcurrent_info
367
368	clfmt='(1X,"Timing started on ",2(A2,"/"),A4," at ",2(A2,":"),A2," MET ",A3,":",A2," from GMT")'
369	IF( lwriter ) WRITE(numtime, TRIM(clfmt)) &
370	& cdate(1)(7:8), cdate(1)(5:6), cdate(1)(1:4), &
371	& ctime(1)(1:2), ctime(1)(3:4), ctime(1)(5:6), &
372	& czone(1:3), czone(4:5)
373	clfmt='(1X, "Timing ended on ",2(A2,"/"),A4," at ",2(A2,":"),A2," MET ",A3,":",A2," from GMT")'
374	IF( lwriter ) WRITE(numtime, TRIM(clfmt)) &
375	& cdate(2)(7:8), cdate(2)(5:6), cdate(2)(1:4), &
376	& ctime(2)(1:2), ctime(2)(3:4), ctime(2)(5:6), &
377	& czone(1:3), czone(4:5)
378
379	#if defined key_mpp_mpi
380	ALLOCATE(timing_glob(3*jpnij), stat=icode)
381	CALL MPI_GATHER( (/compute_time, waiting_time, elapsed_time/), 3, MPI_DOUBLE_PRECISION, &
382	& timing_glob, 3, MPI_DOUBLE_PRECISION, 0, MPI_COMM_OCE, icode)
383	IF( narea == 1 ) THEN
384	WRITE(numtime,*) ' '
385	WRITE(numtime,*) ' Report on time spent on waiting MPI messages '
386	WRITE(numtime,*) ' total timing measured between nit000+1 and nitend-1 '
387	WRITE(numtime,*) ' warning: includes restarts writing time if output before nitend... '
388	WRITE(numtime,*) ' '
389	DO ji = 1, jpnij
390	zperc = timing_glob(3ji-1) + timing_glob(3ji-2)
391	IF (zperc /= 0. ) zperc = timing_glob(3ji-1) / zperc 100.
392	WRITE(numtime,'(A20,F11.6, A34,I8)') 'Computing time : ',timing_glob(3*ji-2), ' on MPI rank : ', ji
393	WRITE(numtime,'(A20,F11.6,A2, F4.1,A3,A25,I8)') 'Waiting time : ',timing_glob(3*ji-1) &
394	& , ' (', zperc,' %)', ' on MPI rank : ', ji
395	zsypd = rn_rdt * REAL(nitend-nit000-1, wp) / (timing_glob(3ji) 365.)
396	WRITE(numtime,'(A20,F11.6,A7,F10.3,A2,A15,I8)') 'Total time : ',timing_glob(3*ji ) &
397	& , ' (SYPD: ', zsypd, ')', ' on MPI rank : ', ji
398	END DO
399	ENDIF
400	DEALLOCATE(timing_glob)
401	#endif
402
403	IF( lwriter ) CLOSE(numtime)
404	!
405	END SUBROUTINE timing_finalize
406
407
408	SUBROUTINE wcurrent_info
409	!!----------------------------------------------------------------------
410	!! * ROUTINE wcurrent_info *
411	!! ** Purpose : compute and write timing output file
412	!!----------------------------------------------------------------------
413	LOGICAL :: ll_ord
414	CHARACTER(len=2048) :: clfmt
415
416	! reorder the current list by elapse time
417	s_wrk => NULL()
418	s_timer => s_timer_root
419	DO
420	ll_ord = .TRUE.
421	s_timer => s_timer_root
422	DO WHILE ( ASSOCIATED( s_timer%next ) )
423	IF (.NOT. ASSOCIATED(s_timer%next)) EXIT
424	IF ( s_timer%tsum_clock < s_timer%next%tsum_clock ) THEN
425	ALLOCATE(s_wrk)
426	s_wrk = s_timer%next
427	CALL insert (s_timer, s_timer_root, s_wrk)
428	CALL suppress(s_timer%next)
429	ll_ord = .FALSE.
430	CYCLE
431	ENDIF
432	IF( ASSOCIATED(s_timer%next) ) s_timer => s_timer%next
433	END DO
434	IF( ll_ord ) EXIT
435	END DO
436
437	! write current info
438	WRITE(numtime,*) 'Detailed timing for proc :', narea-1
439	WRITE(numtime,*) '--------------------------'
440	WRITE(numtime,*) 'Section ', &
441	& 'Elapsed Time (s) ','Elapsed Time (%) ', &
442	& 'CPU Time(s) ','CPU Time (%) ','CPU/Elapsed ','Frequency'
443	s_timer => s_timer_root
444	clfmt = '(1x,a,4x,f12.3,6x,f12.3,x,f12.3,2x,f12.3,6x,f7.3,2x,i9)'
445	DO WHILE ( ASSOCIATED(s_timer) )
446	WRITE(numtime,TRIM(clfmt)) s_timer%cname, &
447	& s_timer%tsum_clock,s_timer%tsum_clock*100./t_elaps(2), &
448	& s_timer%tsum_cpu ,s_timer%tsum_cpu*100./t_cpu(2) , &
449	& s_timer%tsum_cpu/s_timer%tsum_clock, s_timer%niter
450	s_timer => s_timer%next
451	END DO
452	WRITE(numtime,*)
453	!
454	END SUBROUTINE wcurrent_info
455
456	#if defined key_mpp_mpi
457	SUBROUTINE waver_info
458	!!----------------------------------------------------------------------
459	!! * ROUTINE wcurrent_info *
460	!! ** Purpose : compute and write averaged timing informations
461	!!----------------------------------------------------------------------
462	TYPE(alltimer), POINTER :: sl_timer_glob_root => NULL()
463	TYPE(alltimer), POINTER :: sl_timer_glob => NULL()
464	TYPE(timer), POINTER :: sl_timer_ave_root => NULL()
465	TYPE(timer), POINTER :: sl_timer_ave => NULL()
466	INTEGER :: icode
467	INTEGER :: ierr
468	LOGICAL :: ll_ord
469	CHARACTER(len=200) :: clfmt
470
471	! Initialised the global strucutre
472	ALLOCATE(sl_timer_glob_root, Stat=ierr)
473	IF(ierr /= 0)THEN
474	WRITE(numtime,*) 'Failed to allocate global timing structure in waver_info'
475	RETURN
476	END IF
477
478	ALLOCATE(sl_timer_glob_root%cname (jpnij), &
479	sl_timer_glob_root%tsum_cpu (jpnij), &
480	sl_timer_glob_root%tsum_clock(jpnij), &
481	sl_timer_glob_root%niter (jpnij), Stat=ierr)
482	IF(ierr /= 0)THEN
483	WRITE(numtime,*) 'Failed to allocate global timing structure in waver_info'
484	RETURN
485	END IF
486	sl_timer_glob_root%cname(:) = ''
487	sl_timer_glob_root%tsum_cpu(:) = 0._wp
488	sl_timer_glob_root%tsum_clock(:) = 0._wp
489	sl_timer_glob_root%niter(:) = 0
490	sl_timer_glob_root%next => NULL()
491	sl_timer_glob_root%prev => NULL()
492	!ARPDBG - don't need to allocate a pointer that's immediately then
493	! set to point to some other object.
494	!ALLOCATE(sl_timer_glob)
495	!ALLOCATE(sl_timer_glob%cname (jpnij))
496	!ALLOCATE(sl_timer_glob%tsum_cpu (jpnij))
497	!ALLOCATE(sl_timer_glob%tsum_clock(jpnij))
498	!ALLOCATE(sl_timer_glob%niter (jpnij))
499	sl_timer_glob => sl_timer_glob_root
500	!
501	IF( narea .EQ. 1 ) THEN
502	ALLOCATE(sl_timer_ave_root)
503	sl_timer_ave_root%cname = ''
504	sl_timer_ave_root%t_cpu = 0._wp
505	sl_timer_ave_root%t_clock = 0._wp
506	sl_timer_ave_root%tsum_cpu = 0._wp
507	sl_timer_ave_root%tsum_clock = 0._wp
508	sl_timer_ave_root%tmax_cpu = 0._wp
509	sl_timer_ave_root%tmax_clock = 0._wp
510	sl_timer_ave_root%tmin_cpu = 0._wp
511	sl_timer_ave_root%tmin_clock = 0._wp
512	sl_timer_ave_root%tsub_cpu = 0._wp
513	sl_timer_ave_root%tsub_clock = 0._wp
514	sl_timer_ave_root%ncount = 0
515	sl_timer_ave_root%ncount_rate = 0
516	sl_timer_ave_root%ncount_max = 0
517	sl_timer_ave_root%niter = 0
518	sl_timer_ave_root%l_tdone = .FALSE.
519	sl_timer_ave_root%next => NULL()
520	sl_timer_ave_root%prev => NULL()
521	ALLOCATE(sl_timer_ave)
522	sl_timer_ave => sl_timer_ave_root
523	ENDIF
524
525	! Gather info from all processors
526	s_timer => s_timer_root
527	DO WHILE ( ASSOCIATED(s_timer) )
528	CALL MPI_GATHER(s_timer%cname , 20, MPI_CHARACTER, &
529	sl_timer_glob%cname, 20, MPI_CHARACTER, &
530	0, MPI_COMM_OCE, icode)
531	CALL MPI_GATHER(s_timer%tsum_clock , 1, MPI_DOUBLE_PRECISION, &
532	sl_timer_glob%tsum_clock, 1, MPI_DOUBLE_PRECISION, &
533	0, MPI_COMM_OCE, icode)
534	CALL MPI_GATHER(s_timer%tsum_cpu , 1, MPI_DOUBLE_PRECISION, &
535	sl_timer_glob%tsum_cpu, 1, MPI_DOUBLE_PRECISION, &
536	0, MPI_COMM_OCE, icode)
537	CALL MPI_GATHER(s_timer%niter , 1, MPI_INTEGER, &
538	sl_timer_glob%niter, 1, MPI_INTEGER, &
539	0, MPI_COMM_OCE, icode)
540
541	IF( narea == 1 .AND. ASSOCIATED(s_timer%next) ) THEN
542	ALLOCATE(sl_timer_glob%next)
543	ALLOCATE(sl_timer_glob%next%cname (jpnij))
544	ALLOCATE(sl_timer_glob%next%tsum_cpu (jpnij))
545	ALLOCATE(sl_timer_glob%next%tsum_clock(jpnij))
546	ALLOCATE(sl_timer_glob%next%niter (jpnij))
547	sl_timer_glob%next%prev => sl_timer_glob
548	sl_timer_glob%next%next => NULL()
549	sl_timer_glob => sl_timer_glob%next
550	ENDIF
551	s_timer => s_timer%next
552	END DO
553
554	IF( narea == 1 ) THEN
555	! Compute some stats
556	sl_timer_glob => sl_timer_glob_root
557	DO WHILE( ASSOCIATED(sl_timer_glob) )
558	sl_timer_ave%cname = sl_timer_glob%cname(1)
559	sl_timer_ave%tsum_cpu = SUM (sl_timer_glob%tsum_cpu (:)) / jpnij
560	sl_timer_ave%tsum_clock = SUM (sl_timer_glob%tsum_clock(:)) / jpnij
561	sl_timer_ave%tmax_cpu = MAXVAL(sl_timer_glob%tsum_cpu (:))
562	sl_timer_ave%tmax_clock = MAXVAL(sl_timer_glob%tsum_clock(:))
563	sl_timer_ave%tmin_cpu = MINVAL(sl_timer_glob%tsum_cpu (:))
564	sl_timer_ave%tmin_clock = MINVAL(sl_timer_glob%tsum_clock(:))
565	sl_timer_ave%niter = SUM (sl_timer_glob%niter (:))
566	!
567	IF( ASSOCIATED(sl_timer_glob%next) ) THEN
568	ALLOCATE(sl_timer_ave%next)
569	sl_timer_ave%next%prev => sl_timer_ave
570	sl_timer_ave%next%next => NULL()
571	sl_timer_ave => sl_timer_ave%next
572	ENDIF
573	sl_timer_glob => sl_timer_glob%next
574	END DO
575
576	! reorder the averaged list by CPU time
577	s_wrk => NULL()
578	sl_timer_ave => sl_timer_ave_root
579	DO
580	ll_ord = .TRUE.
581	sl_timer_ave => sl_timer_ave_root
582	DO WHILE( ASSOCIATED( sl_timer_ave%next ) )
583
584	IF( .NOT. ASSOCIATED(sl_timer_ave%next) ) EXIT
585
586	IF ( sl_timer_ave%tsum_clock < sl_timer_ave%next%tsum_clock ) THEN
587	ALLOCATE(s_wrk)
588	! Copy data into the new object pointed to by s_wrk
589	s_wrk = sl_timer_ave%next
590	! Insert this new timer object before our current position
591	CALL insert (sl_timer_ave, sl_timer_ave_root, s_wrk)
592	! Remove the old object from the list
593	CALL suppress(sl_timer_ave%next)
594	ll_ord = .FALSE.
595	CYCLE
596	ENDIF
597	IF( ASSOCIATED(sl_timer_ave%next) ) sl_timer_ave => sl_timer_ave%next
598	END DO
599	IF( ll_ord ) EXIT
600	END DO
601
602	! write averaged info
603	WRITE(numtime,"('Averaged timing on all processors :')")
604	WRITE(numtime,"('-----------------------------------')")
605	WRITE(numtime,"('Section',13x,'Elap. Time(s)',2x,'Elap. Time(%)',2x, &
606	& 'CPU Time(s)',2x,'CPU Time(%)',2x,'CPU/Elap',1x, &
607	& 'Max elap(%)',2x,'Min elap(%)',2x, &
608	& 'Freq')")
609	sl_timer_ave => sl_timer_ave_root
610	clfmt = '((A),E15.7,2x,f6.2,5x,f12.2,5x,f6.2,5x,f7.2,2x,f12.2,4x,f6.2,2x,f9.2)'
611	DO WHILE ( ASSOCIATED(sl_timer_ave) )
612	WRITE(numtime,TRIM(clfmt)) sl_timer_ave%cname(1:18), &
613	& sl_timer_ave%tsum_clock,sl_timer_ave%tsum_clock100.jpnij/tot_etime, &
614	& sl_timer_ave%tsum_cpu ,sl_timer_ave%tsum_cpu100.jpnij/tot_ctime , &
615	& sl_timer_ave%tsum_cpu/sl_timer_ave%tsum_clock, &
616	& sl_timer_ave%tmax_clock100.jpnij/tot_etime, &
617	& sl_timer_ave%tmin_clock100.jpnij/tot_etime, &
618	& sl_timer_ave%niter/REAL(jpnij)
619	sl_timer_ave => sl_timer_ave%next
620	END DO
621	WRITE(numtime,*)
622	!
623	DEALLOCATE(sl_timer_ave_root)
624	ENDIF
625	!
626	DEALLOCATE(sl_timer_glob_root)
627	!
628	END SUBROUTINE waver_info
629
630
631	SUBROUTINE wmpi_info
632	!!----------------------------------------------------------------------
633	!! * ROUTINE wmpi_time *
634	!! ** Purpose : compute and write a summary of MPI infos
635	!!----------------------------------------------------------------------
636	!
637	INTEGER :: idum, icode
638	INTEGER, ALLOCATABLE, DIMENSION(:) :: iall_rank
639	REAL(wp) :: ztot_ratio
640	REAL(wp) :: zmax_etime, zmax_ctime, zmax_ratio, zmin_etime, zmin_ctime, zmin_ratio
641	REAL(wp) :: zavg_etime, zavg_ctime, zavg_ratio
642	REAL(wp), ALLOCATABLE, DIMENSION(:) :: zall_ratio
643	CHARACTER(LEN=128), dimension(8) :: cllignes
644	CHARACTER(LEN=128) :: clhline, clstart_date, clfinal_date
645	CHARACTER(LEN=2048) :: clfmt
646
647	! Gather all times
648	ALLOCATE( zall_ratio(jpnij), iall_rank(jpnij) )
649	IF( narea == 1 ) THEN
650	iall_rank(:) = (/ (idum,idum=0,jpnij-1) /)
651
652	! Compute elapse user time
653	zavg_etime = tot_etime/REAL(jpnij,wp)
654	zmax_etime = MAXVAL(all_etime(:))
655	zmin_etime = MINVAL(all_etime(:))
656
657	! Compute CPU user time
658	zavg_ctime = tot_ctime/REAL(jpnij,wp)
659	zmax_ctime = MAXVAL(all_ctime(:))
660	zmin_ctime = MINVAL(all_ctime(:))
661
662	! Compute cpu/elapsed ratio
663	zall_ratio(:) = all_ctime(:) / all_etime(:)
664	ztot_ratio = SUM(zall_ratio(:))
665	zavg_ratio = ztot_ratio/REAL(jpnij,wp)
666	zmax_ratio = MAXVAL(zall_ratio(:))
667	zmin_ratio = MINVAL(zall_ratio(:))
668
669	! Output Format
670	clhline ='1x,13("-"),"\|",18("-"),"\|",14("-"),"\|",18("-"),/,'
671	cllignes(1)='(1x,"MPI summary report :",/,'
672	cllignes(2)='1x,"--------------------",//,'
673	cllignes(3)='1x,"Process Rank \|"," Elapsed Time (s) \|"," CPU Time (s) \|"," Ratio CPU/Elapsed",/,'
674	cllignes(4)=' (1x,i4,9x,"\|",f12.3,6x,"\|",f12.3,2x,"\|",4x,f7.3,/),'
675	WRITE(cllignes(4)(1:4),'(I4)') jpnij
676	cllignes(5)='1x,"Total \|",f12.3,6x,"\|",F12.3,2x,"\|",4x,f7.3,/,'
677	cllignes(6)='1x,"Minimum \|",f12.3,6x,"\|",F12.3,2x,"\|",4x,f7.3,/,'
678	cllignes(7)='1x,"Maximum \|",f12.3,6x,"\|",F12.3,2x,"\|",4x,f7.3,/,'
679	cllignes(8)='1x,"Average \|",f12.3,6x,"\|",F12.3,2x,"\|",4x,f7.3)'
680	clfmt=TRIM(cllignes(1))// TRIM(cllignes(2))//TRIM(cllignes(3))// &
681	& TRIM(clhline)//TRIM(cllignes(4))//TRIM(clhline)//TRIM(cllignes(5))// &
682	& TRIM(clhline)//TRIM(cllignes(6))//TRIM(clhline)//TRIM(cllignes(7))// &
683	& TRIM(clhline)//TRIM(cllignes(8))
684	WRITE(numtime, TRIM(clfmt)) &
685	(iall_rank(idum),all_etime(idum),all_ctime(idum),zall_ratio(idum),idum=1, jpnij), &
686	tot_etime, tot_ctime, ztot_ratio, &
687	zmin_etime, zmin_ctime, zmin_ratio, &
688	zmax_etime, zmax_ctime, zmax_ratio, &
689	zavg_etime, zavg_ctime, zavg_ratio
690	WRITE(numtime,*)
691	END IF
692	!
693	DEALLOCATE(zall_ratio, iall_rank)
694	!
695	END SUBROUTINE wmpi_info
696	#endif
697
698
699	SUBROUTINE timing_ini_var(cdinfo)
700	!!----------------------------------------------------------------------
701	!! * ROUTINE timing_ini_var *
702	!! ** Purpose : create timing structure
703	!!----------------------------------------------------------------------
704	CHARACTER(len=*), INTENT(in) :: cdinfo
705	LOGICAL :: ll_section
706
707	!
708	IF( .NOT. ASSOCIATED(s_timer_root) ) THEN
709	ALLOCATE(s_timer_root)
710	s_timer_root%cname = cdinfo
711	s_timer_root%t_cpu = 0._wp
712	s_timer_root%t_clock = 0._wp
713	s_timer_root%tsum_cpu = 0._wp
714	s_timer_root%tsum_clock = 0._wp
715	s_timer_root%tmax_cpu = 0._wp
716	s_timer_root%tmax_clock = 0._wp
717	s_timer_root%tmin_cpu = 0._wp
718	s_timer_root%tmin_clock = 0._wp
719	s_timer_root%tsub_cpu = 0._wp
720	s_timer_root%tsub_clock = 0._wp
721	s_timer_root%ncount = 0
722	s_timer_root%ncount_rate = 0
723	s_timer_root%ncount_max = 0
724	s_timer_root%niter = 0
725	s_timer_root%l_tdone = .FALSE.
726	s_timer_root%next => NULL()
727	s_timer_root%prev => NULL()
728	s_timer => s_timer_root
729	!
730	ALLOCATE(s_wrk)
731	s_wrk => NULL()
732	!
733	ALLOCATE(s_timer_old)
734	s_timer_old%cname = cdinfo
735	s_timer_old%t_cpu = 0._wp
736	s_timer_old%t_clock = 0._wp
737	s_timer_old%tsum_cpu = 0._wp
738	s_timer_old%tsum_clock = 0._wp
739	s_timer_old%tmax_cpu = 0._wp
740	s_timer_old%tmax_clock = 0._wp
741	s_timer_old%tmin_cpu = 0._wp
742	s_timer_old%tmin_clock = 0._wp
743	s_timer_old%tsub_cpu = 0._wp
744	s_timer_old%tsub_clock = 0._wp
745	s_timer_old%ncount = 0
746	s_timer_old%ncount_rate = 0
747	s_timer_old%ncount_max = 0
748	s_timer_old%niter = 0
749	s_timer_old%l_tdone = .TRUE.
750	s_timer_old%next => NULL()
751	s_timer_old%prev => NULL()
752
753	ELSE
754	s_timer => s_timer_root
755	! case of already existing area (typically inside a loop)
756	! write(,) 'in ini_var for routine : ', cdinfo
757	DO WHILE( ASSOCIATED(s_timer) )
758	IF( TRIM(s_timer%cname) .EQ. TRIM(cdinfo) ) THEN
759	! write(,) 'in ini_var for routine : ', cdinfo,' we return'
760	RETURN ! cdinfo is already in the chain
761	ENDIF
762	s_timer => s_timer%next
763	END DO
764
765	! end of the chain
766	s_timer => s_timer_root
767	DO WHILE( ASSOCIATED(s_timer%next) )
768	s_timer => s_timer%next
769	END DO
770
771	! write(,) 'after search', s_timer%cname
772	! cdinfo is not part of the chain so we add it with initialisation
773	ALLOCATE(s_timer%next)
774	! write(,) 'after allocation of next'
775
776	s_timer%next%cname = cdinfo
777	s_timer%next%t_cpu = 0._wp
778	s_timer%next%t_clock = 0._wp
779	s_timer%next%tsum_cpu = 0._wp
780	s_timer%next%tsum_clock = 0._wp
781	s_timer%next%tmax_cpu = 0._wp
782	s_timer%next%tmax_clock = 0._wp
783	s_timer%next%tmin_cpu = 0._wp
784	s_timer%next%tmin_clock = 0._wp
785	s_timer%next%tsub_cpu = 0._wp
786	s_timer%next%tsub_clock = 0._wp
787	s_timer%next%ncount = 0
788	s_timer%next%ncount_rate = 0
789	s_timer%next%ncount_max = 0
790	s_timer%next%niter = 0
791	s_timer%next%l_tdone = .FALSE.
792	s_timer%next%parent_section => NULL()
793	s_timer%next%prev => s_timer
794	s_timer%next%next => NULL()
795	s_timer => s_timer%next
796	ENDIF
797	! write(,) 'after allocation'
798	!
799	END SUBROUTINE timing_ini_var
800
801
802	SUBROUTINE timing_reset
803	!!----------------------------------------------------------------------
804	!! * ROUTINE timing_reset *
805	!! ** Purpose : go to root of timing tree
806	!!----------------------------------------------------------------------
807	l_initdone = .TRUE.
808	! IF(lwp) WRITE(numout,*)
809	! IF(lwp) WRITE(numout,*) 'timing_reset : instrumented routines for timing'
810	! IF(lwp) WRITE(numout,*) '~~~~~~~~~~~~'
811	CALL timing_list(s_timer_root)
812	! WRITE(numout,*)
813	!
814	END SUBROUTINE timing_reset
815
816
817	RECURSIVE SUBROUTINE timing_list(ptr)
818
819	TYPE(timer), POINTER, INTENT(inout) :: ptr
820	!
821	IF( ASSOCIATED(ptr%next) ) CALL timing_list(ptr%next)
822	IF(lwp) WRITE(numout,*)' ', ptr%cname
823	!
824	END SUBROUTINE timing_list
825
826
827	SUBROUTINE insert(sd_current, sd_root ,sd_ptr)
828	!!----------------------------------------------------------------------
829	!! * ROUTINE insert *
830	!! ** Purpose : insert an element in timer structure
831	!!----------------------------------------------------------------------
832	TYPE(timer), POINTER, INTENT(inout) :: sd_current, sd_root, sd_ptr
833	!
834
835	IF( ASSOCIATED( sd_current, sd_root ) ) THEN
836	! If our current element is the root element then
837	! replace it with the one being inserted
838	sd_root => sd_ptr
839	ELSE
840	sd_current%prev%next => sd_ptr
841	END IF
842	sd_ptr%next => sd_current
843	sd_ptr%prev => sd_current%prev
844	sd_current%prev => sd_ptr
845	! Nullify the pointer to the new element now that it is held
846	! within the list. If we don't do this then a subsequent call
847	! to ALLOCATE memory to this pointer will fail.
848	sd_ptr => NULL()
849	!
850	END SUBROUTINE insert
851
852
853	SUBROUTINE suppress(sd_ptr)
854	!!----------------------------------------------------------------------
855	!! * ROUTINE suppress *
856	!! ** Purpose : supress an element in timer structure
857	!!----------------------------------------------------------------------
858	TYPE(timer), POINTER, INTENT(inout) :: sd_ptr
859	!
860	TYPE(timer), POINTER :: sl_temp
861
862	sl_temp => sd_ptr
863	sd_ptr => sd_ptr%next
864	IF ( ASSOCIATED(sl_temp%next) ) sl_temp%next%prev => sl_temp%prev
865	DEALLOCATE(sl_temp)
866	sl_temp => NULL()
867	!
868	END SUBROUTINE suppress
869
870	!!=====================================================================
871	END MODULE timing

Note: See TracBrowser for help on using the repository browser.

Download in other formats: