Context Navigation

← Previous Revision
Latest Revision
Next Revision →
Blame
Revision Log

timing.F90 @ 15742

Last change on this file since 15742 was 15742, checked in by emmafiedler, 2 years ago
Merging in R75 strength changes, diagnostics updates and latest changes
File size: 35.6 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 4.0 (2019) '
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
266	! Timing on date and time
267	CALL DATE_AND_TIME(cdate(1),ctime(1),czone,nvalues)
268
269	CALL CPU_TIME(t_cpu(1))
270	#if defined key_mpp_mpi
271	! Start elapsed and CPU time counters
272	t_elaps(1) = MPI_WTIME()
273	#else
274	CALL SYSTEM_CLOCK(COUNT_RATE=ncount_rate, COUNT_MAX=ncount_max)
275	CALL SYSTEM_CLOCK(COUNT = ncount)
276	#endif
277	!
278	END SUBROUTINE timing_init
279
280
281	SUBROUTINE timing_finalize
282	!!----------------------------------------------------------------------
283	!! * ROUTINE timing_finalize *
284	!! ** Purpose : compute average time
285	!! write timing output file
286	!!----------------------------------------------------------------------
287	TYPE(timer), POINTER :: s_temp
288	INTEGER :: idum, iperiods, icode
289	INTEGER :: ji
290	LOGICAL :: ll_ord, ll_averep
291	CHARACTER(len=120) :: clfmt
292	REAL(wp), DIMENSION(:), ALLOCATABLE :: timing_glob
293	REAL(wp) :: zsypd ! simulated years per day (Balaji 2017)
294	REAL(wp) :: zperc, ztot
295
296	ll_averep = .TRUE.
297	zperc = 0._wp
298
299	! total CPU and elapse
300	CALL CPU_TIME(t_cpu(2))
301	t_cpu(2) = t_cpu(2) - t_cpu(1) - t_overcpu
302	#if defined key_mpp_mpi
303	t_elaps(2) = MPI_WTIME() - t_elaps(1) - t_overclock
304	#else
305	CALL SYSTEM_CLOCK(COUNT = nfinal_count)
306	iperiods = nfinal_count - ncount
307	IF( nfinal_count < ncount ) &
308	iperiods = iperiods + ncount_max
309	t_elaps(2) = REAL(iperiods) / ncount_rate - t_overclock
310	#endif
311
312	! End of timings on date & time
313	CALL DATE_AND_TIME(cdate(2),ctime(2),czone,nvalues)
314
315	! Compute the numer of routines
316	nsize = 0
317	s_timer => s_timer_root
318	DO WHILE( ASSOCIATED(s_timer) )
319	nsize = nsize + 1
320	s_timer => s_timer%next
321	END DO
322	idum = nsize
323	CALL mpp_sum('timing', idum)
324	IF( idum/jpnij /= nsize ) THEN
325	IF( lwriter ) WRITE(numtime,*) ' ===> W A R N I N G: '
326	IF( lwriter ) WRITE(numtime,*) ' Some CPU have different number of routines instrumented for timing'
327	IF( lwriter ) WRITE(numtime,*) ' No detailed report on averaged timing can be provided'
328	IF( lwriter ) WRITE(numtime,*) ' The following detailed report only deals with the current processor'
329	IF( lwriter ) WRITE(numtime,*)
330	ll_averep = .FALSE.
331	ENDIF
332
333	#if defined key_mpp_mpi
334	! in MPI gather some info
335	ALLOCATE( all_etime(jpnij), all_ctime(jpnij) )
336	CALL MPI_ALLGATHER(t_elaps(2), 1, MPI_DOUBLE_PRECISION, &
337	all_etime , 1, MPI_DOUBLE_PRECISION, &
338	MPI_COMM_OCE, icode)
339	CALL MPI_ALLGATHER(t_cpu(2) , 1, MPI_DOUBLE_PRECISION, &
340	all_ctime, 1, MPI_DOUBLE_PRECISION, &
341	MPI_COMM_OCE, icode)
342	tot_etime = SUM(all_etime(:))
343	tot_ctime = SUM(all_ctime(:))
344	#else
345	tot_etime = t_elaps(2)
346	tot_ctime = t_cpu (2)
347	#endif
348
349	! write output file
350	IF( lwriter ) WRITE(numtime,*)
351	IF( lwriter ) WRITE(numtime,*)
352	IF( lwriter ) WRITE(numtime,*) 'Total timing (sum) :'
353	IF( lwriter ) WRITE(numtime,*) '--------------------'
354	IF( lwriter ) WRITE(numtime,"('Elapsed Time (s) CPU Time (s)')")
355	IF( lwriter ) WRITE(numtime,'(5x,f12.3,1x,f12.3)') tot_etime, tot_ctime
356	IF( lwriter ) WRITE(numtime,*)
357	#if defined key_mpp_mpi
358	IF( ll_averep ) CALL waver_info
359	CALL wmpi_info
360	#endif
361	IF( lwriter ) CALL wcurrent_info
362
363	clfmt='(1X,"Timing started on ",2(A2,"/"),A4," at ",2(A2,":"),A2," MET ",A3,":",A2," from GMT")'
364	IF( lwriter ) WRITE(numtime, TRIM(clfmt)) &
365	& cdate(1)(7:8), cdate(1)(5:6), cdate(1)(1:4), &
366	& ctime(1)(1:2), ctime(1)(3:4), ctime(1)(5:6), &
367	& czone(1:3), czone(4:5)
368	clfmt='(1X, "Timing ended on ",2(A2,"/"),A4," at ",2(A2,":"),A2," MET ",A3,":",A2," from GMT")'
369	IF( lwriter ) WRITE(numtime, TRIM(clfmt)) &
370	& cdate(2)(7:8), cdate(2)(5:6), cdate(2)(1:4), &
371	& ctime(2)(1:2), ctime(2)(3:4), ctime(2)(5:6), &
372	& czone(1:3), czone(4:5)
373
374	#if defined key_mpp_mpi
375	ALLOCATE(timing_glob(4*jpnij), stat=icode)
376	CALL MPI_GATHER( (/compute_time, waiting_time(1), waiting_time(2), elapsed_time/), &
377	& 4, MPI_DOUBLE_PRECISION, timing_glob, 4, MPI_DOUBLE_PRECISION, 0, MPI_COMM_OCE, icode)
378	IF( narea == 1 ) THEN
379	WRITE(numtime,*) ' '
380	WRITE(numtime,*) ' Report on time spent on waiting MPI messages '
381	WRITE(numtime,*) ' total timing measured between nit000+1 and nitend-1 '
382	WRITE(numtime,*) ' warning: includes restarts writing time if output before nitend... '
383	WRITE(numtime,*) ' '
384	DO ji = 1, jpnij
385	ztot = SUM( timing_glob(4ji-3:4ji-1) )
386	WRITE(numtime,'(A28,F11.6, A34,I8)') 'Computing time : ',timing_glob(4*ji-3), ' on MPI rank : ', ji
387	IF ( ztot /= 0. ) zperc = timing_glob(4ji-2) / ztot 100.
388	WRITE(numtime,'(A28,F11.6,A2, F4.1,A3,A25,I8)') 'Waiting lbc_lnk time : ',timing_glob(4*ji-2) &
389	& , ' (', zperc,' %)', ' on MPI rank : ', ji
390	IF ( ztot /= 0. ) zperc = timing_glob(4ji-1) / ztot 100.
391	WRITE(numtime,'(A28,F11.6,A2, F4.1,A3,A25,I8)') 'Waiting global time : ',timing_glob(4*ji-1) &
392	& , ' (', zperc,' %)', ' on MPI rank : ', ji
393	zsypd = rn_rdt * REAL(nitend-nit000-1, wp) / (timing_glob(4ji) 365.)
394	WRITE(numtime,'(A28,F11.6,A7,F10.3,A2,A15,I8)') 'Total time : ',timing_glob(4*ji ) &
395	& , ' (SYPD: ', zsypd, ')', ' on MPI rank : ', ji
396	END DO
397	ENDIF
398	DEALLOCATE(timing_glob)
399	#endif
400
401	IF( lwriter ) CLOSE(numtime)
402	!
403	END SUBROUTINE timing_finalize
404
405
406	SUBROUTINE wcurrent_info
407	!!----------------------------------------------------------------------
408	!! * ROUTINE wcurrent_info *
409	!! ** Purpose : compute and write timing output file
410	!!----------------------------------------------------------------------
411	LOGICAL :: ll_ord
412	CHARACTER(len=2048) :: clfmt
413
414	! reorder the current list by elapse time
415	s_wrk => NULL()
416	s_timer => s_timer_root
417	DO
418	ll_ord = .TRUE.
419	s_timer => s_timer_root
420	DO WHILE ( ASSOCIATED( s_timer%next ) )
421	IF (.NOT. ASSOCIATED(s_timer%next)) EXIT
422	IF ( s_timer%tsum_clock < s_timer%next%tsum_clock ) THEN
423	ALLOCATE(s_wrk)
424	s_wrk = s_timer%next
425	CALL insert (s_timer, s_timer_root, s_wrk)
426	CALL suppress(s_timer%next)
427	ll_ord = .FALSE.
428	CYCLE
429	ENDIF
430	IF( ASSOCIATED(s_timer%next) ) s_timer => s_timer%next
431	END DO
432	IF( ll_ord ) EXIT
433	END DO
434
435	! write current info
436	WRITE(numtime,*) 'Detailed timing for proc :', narea-1
437	WRITE(numtime,*) '--------------------------'
438	WRITE(numtime,*) 'Section ', &
439	& 'Elapsed Time (s) ','Elapsed Time (%) ', &
440	& 'CPU Time(s) ','CPU Time (%) ','CPU/Elapsed ','Frequency'
441	s_timer => s_timer_root
442	clfmt = '(1x,a,4x,f12.3,6x,f12.3,x,f12.3,2x,f12.3,6x,f7.3,2x,i9)'
443	DO WHILE ( ASSOCIATED(s_timer) )
444	WRITE(numtime,TRIM(clfmt)) s_timer%cname, &
445	& s_timer%tsum_clock,s_timer%tsum_clock*100./t_elaps(2), &
446	& s_timer%tsum_cpu ,s_timer%tsum_cpu*100./t_cpu(2) , &
447	& s_timer%tsum_cpu/s_timer%tsum_clock, s_timer%niter
448	s_timer => s_timer%next
449	END DO
450	WRITE(numtime,*)
451	!
452	END SUBROUTINE wcurrent_info
453
454	#if defined key_mpp_mpi
455	SUBROUTINE waver_info
456	!!----------------------------------------------------------------------
457	!! * ROUTINE wcurrent_info *
458	!! ** Purpose : compute and write averaged timing informations
459	!!----------------------------------------------------------------------
460	TYPE(alltimer), POINTER :: sl_timer_glob_root => NULL()
461	TYPE(alltimer), POINTER :: sl_timer_glob => NULL()
462	TYPE(timer), POINTER :: sl_timer_ave_root => NULL()
463	TYPE(timer), POINTER :: sl_timer_ave => NULL()
464	INTEGER :: icode
465	INTEGER :: ierr
466	LOGICAL :: ll_ord
467	CHARACTER(len=200) :: clfmt
468
469	! Initialised the global strucutre
470	ALLOCATE(sl_timer_glob_root, Stat=ierr)
471	IF(ierr /= 0)THEN
472	WRITE(numtime,*) 'Failed to allocate global timing structure in waver_info'
473	RETURN
474	END IF
475
476	ALLOCATE(sl_timer_glob_root%cname (jpnij), &
477	sl_timer_glob_root%tsum_cpu (jpnij), &
478	sl_timer_glob_root%tsum_clock(jpnij), &
479	sl_timer_glob_root%niter (jpnij), Stat=ierr)
480	IF(ierr /= 0)THEN
481	WRITE(numtime,*) 'Failed to allocate global timing structure in waver_info'
482	RETURN
483	END IF
484	sl_timer_glob_root%cname(:) = ''
485	sl_timer_glob_root%tsum_cpu(:) = 0._wp
486	sl_timer_glob_root%tsum_clock(:) = 0._wp
487	sl_timer_glob_root%niter(:) = 0
488	sl_timer_glob_root%next => NULL()
489	sl_timer_glob_root%prev => NULL()
490	!ARPDBG - don't need to allocate a pointer that's immediately then
491	! set to point to some other object.
492	!ALLOCATE(sl_timer_glob)
493	!ALLOCATE(sl_timer_glob%cname (jpnij))
494	!ALLOCATE(sl_timer_glob%tsum_cpu (jpnij))
495	!ALLOCATE(sl_timer_glob%tsum_clock(jpnij))
496	!ALLOCATE(sl_timer_glob%niter (jpnij))
497	sl_timer_glob => sl_timer_glob_root
498	!
499	IF( narea .EQ. 1 ) THEN
500	ALLOCATE(sl_timer_ave_root)
501	sl_timer_ave_root%cname = ''
502	sl_timer_ave_root%t_cpu = 0._wp
503	sl_timer_ave_root%t_clock = 0._wp
504	sl_timer_ave_root%tsum_cpu = 0._wp
505	sl_timer_ave_root%tsum_clock = 0._wp
506	sl_timer_ave_root%tmax_cpu = 0._wp
507	sl_timer_ave_root%tmax_clock = 0._wp
508	sl_timer_ave_root%tmin_cpu = 0._wp
509	sl_timer_ave_root%tmin_clock = 0._wp
510	sl_timer_ave_root%tsub_cpu = 0._wp
511	sl_timer_ave_root%tsub_clock = 0._wp
512	sl_timer_ave_root%ncount = 0
513	sl_timer_ave_root%ncount_rate = 0
514	sl_timer_ave_root%ncount_max = 0
515	sl_timer_ave_root%niter = 0
516	sl_timer_ave_root%l_tdone = .FALSE.
517	sl_timer_ave_root%next => NULL()
518	sl_timer_ave_root%prev => NULL()
519	ALLOCATE(sl_timer_ave)
520	sl_timer_ave => sl_timer_ave_root
521	ENDIF
522
523	! Gather info from all processors
524	s_timer => s_timer_root
525	DO WHILE ( ASSOCIATED(s_timer) )
526	CALL MPI_GATHER(s_timer%cname , 20, MPI_CHARACTER, &
527	sl_timer_glob%cname, 20, MPI_CHARACTER, &
528	0, MPI_COMM_OCE, icode)
529	CALL MPI_GATHER(s_timer%tsum_clock , 1, MPI_DOUBLE_PRECISION, &
530	sl_timer_glob%tsum_clock, 1, MPI_DOUBLE_PRECISION, &
531	0, MPI_COMM_OCE, icode)
532	CALL MPI_GATHER(s_timer%tsum_cpu , 1, MPI_DOUBLE_PRECISION, &
533	sl_timer_glob%tsum_cpu, 1, MPI_DOUBLE_PRECISION, &
534	0, MPI_COMM_OCE, icode)
535	CALL MPI_GATHER(s_timer%niter , 1, MPI_INTEGER, &
536	sl_timer_glob%niter, 1, MPI_INTEGER, &
537	0, MPI_COMM_OCE, icode)
538
539	IF( narea == 1 .AND. ASSOCIATED(s_timer%next) ) THEN
540	ALLOCATE(sl_timer_glob%next)
541	ALLOCATE(sl_timer_glob%next%cname (jpnij))
542	ALLOCATE(sl_timer_glob%next%tsum_cpu (jpnij))
543	ALLOCATE(sl_timer_glob%next%tsum_clock(jpnij))
544	ALLOCATE(sl_timer_glob%next%niter (jpnij))
545	sl_timer_glob%next%prev => sl_timer_glob
546	sl_timer_glob%next%next => NULL()
547	sl_timer_glob => sl_timer_glob%next
548	ENDIF
549	s_timer => s_timer%next
550	END DO
551
552	IF( narea == 1 ) THEN
553	! Compute some stats
554	sl_timer_glob => sl_timer_glob_root
555	DO WHILE( ASSOCIATED(sl_timer_glob) )
556	sl_timer_ave%cname = sl_timer_glob%cname(1)
557	sl_timer_ave%tsum_cpu = SUM (sl_timer_glob%tsum_cpu (:)) / jpnij
558	sl_timer_ave%tsum_clock = SUM (sl_timer_glob%tsum_clock(:)) / jpnij
559	sl_timer_ave%tmax_cpu = MAXVAL(sl_timer_glob%tsum_cpu (:))
560	sl_timer_ave%tmax_clock = MAXVAL(sl_timer_glob%tsum_clock(:))
561	sl_timer_ave%tmin_cpu = MINVAL(sl_timer_glob%tsum_cpu (:))
562	sl_timer_ave%tmin_clock = MINVAL(sl_timer_glob%tsum_clock(:))
563	sl_timer_ave%niter = SUM (sl_timer_glob%niter (:))
564	!
565	IF( ASSOCIATED(sl_timer_glob%next) ) THEN
566	ALLOCATE(sl_timer_ave%next)
567	sl_timer_ave%next%prev => sl_timer_ave
568	sl_timer_ave%next%next => NULL()
569	sl_timer_ave => sl_timer_ave%next
570	ENDIF
571	sl_timer_glob => sl_timer_glob%next
572	END DO
573
574	! reorder the averaged list by CPU time
575	s_wrk => NULL()
576	sl_timer_ave => sl_timer_ave_root
577	DO
578	ll_ord = .TRUE.
579	sl_timer_ave => sl_timer_ave_root
580	DO WHILE( ASSOCIATED( sl_timer_ave%next ) )
581
582	IF( .NOT. ASSOCIATED(sl_timer_ave%next) ) EXIT
583
584	IF ( sl_timer_ave%tsum_clock < sl_timer_ave%next%tsum_clock ) THEN
585	ALLOCATE(s_wrk)
586	! Copy data into the new object pointed to by s_wrk
587	s_wrk = sl_timer_ave%next
588	! Insert this new timer object before our current position
589	CALL insert (sl_timer_ave, sl_timer_ave_root, s_wrk)
590	! Remove the old object from the list
591	CALL suppress(sl_timer_ave%next)
592	ll_ord = .FALSE.
593	CYCLE
594	ENDIF
595	IF( ASSOCIATED(sl_timer_ave%next) ) sl_timer_ave => sl_timer_ave%next
596	END DO
597	IF( ll_ord ) EXIT
598	END DO
599
600	! write averaged info
601	WRITE(numtime,"('Averaged timing on all processors :')")
602	WRITE(numtime,"('-----------------------------------')")
603	WRITE(numtime,"('Section',13x,'Elap. Time(s)',2x,'Elap. Time(%)',2x, &
604	& 'CPU Time(s)',2x,'CPU Time(%)',2x,'CPU/Elap',1x, &
605	& 'Max elap(%)',2x,'Min elap(%)',2x, &
606	& 'Freq')")
607	sl_timer_ave => sl_timer_ave_root
608	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)'
609	DO WHILE ( ASSOCIATED(sl_timer_ave) )
610	WRITE(numtime,TRIM(clfmt)) sl_timer_ave%cname(1:18), &
611	& sl_timer_ave%tsum_clock,sl_timer_ave%tsum_clock100.jpnij/tot_etime, &
612	& sl_timer_ave%tsum_cpu ,sl_timer_ave%tsum_cpu100.jpnij/tot_ctime , &
613	& sl_timer_ave%tsum_cpu/sl_timer_ave%tsum_clock, &
614	& sl_timer_ave%tmax_clock100.jpnij/tot_etime, &
615	& sl_timer_ave%tmin_clock100.jpnij/tot_etime, &
616	& sl_timer_ave%niter/REAL(jpnij)
617	sl_timer_ave => sl_timer_ave%next
618	END DO
619	WRITE(numtime,*)
620	!
621	DEALLOCATE(sl_timer_ave_root)
622	ENDIF
623	!
624	DEALLOCATE(sl_timer_glob_root)
625	!
626	END SUBROUTINE waver_info
627
628
629	SUBROUTINE wmpi_info
630	!!----------------------------------------------------------------------
631	!! * ROUTINE wmpi_time *
632	!! ** Purpose : compute and write a summary of MPI infos
633	!!----------------------------------------------------------------------
634	!
635	INTEGER :: idum, icode
636	INTEGER, ALLOCATABLE, DIMENSION(:) :: iall_rank
637	REAL(wp) :: ztot_ratio
638	REAL(wp) :: zmax_etime, zmax_ctime, zmax_ratio, zmin_etime, zmin_ctime, zmin_ratio
639	REAL(wp) :: zavg_etime, zavg_ctime, zavg_ratio
640	REAL(wp), ALLOCATABLE, DIMENSION(:) :: zall_ratio
641	CHARACTER(LEN=128), dimension(8) :: cllignes
642	CHARACTER(LEN=128) :: clhline, clstart_date, clfinal_date
643	CHARACTER(LEN=2048) :: clfmt
644
645	! Gather all times
646	ALLOCATE( zall_ratio(jpnij), iall_rank(jpnij) )
647	IF( narea == 1 ) THEN
648	iall_rank(:) = (/ (idum,idum=0,jpnij-1) /)
649
650	! Compute elapse user time
651	zavg_etime = tot_etime/REAL(jpnij,wp)
652	zmax_etime = MAXVAL(all_etime(:))
653	zmin_etime = MINVAL(all_etime(:))
654
655	! Compute CPU user time
656	zavg_ctime = tot_ctime/REAL(jpnij,wp)
657	zmax_ctime = MAXVAL(all_ctime(:))
658	zmin_ctime = MINVAL(all_ctime(:))
659
660	! Compute cpu/elapsed ratio
661	zall_ratio(:) = all_ctime(:) / all_etime(:)
662	ztot_ratio = SUM(all_ctime(:))/SUM(all_etime(:))
663	zavg_ratio = SUM(zall_ratio(:))/REAL(jpnij,wp)
664	zmax_ratio = MAXVAL(zall_ratio(:))
665	zmin_ratio = MINVAL(zall_ratio(:))
666
667	! Output Format
668	clhline ='1x,13("-"),"\|",18("-"),"\|",14("-"),"\|",18("-"),/,'
669	cllignes(1)='(1x,"MPI summary report :",/,'
670	cllignes(2)='1x,"--------------------",//,'
671	cllignes(3)='1x,"Process Rank \|"," Elapsed Time (s) \|"," CPU Time (s) \|"," Ratio CPU/Elapsed",/,'
672	cllignes(4)=' (4x,i6,4x,"\|",f12.3,6x,"\|",f12.3,2x,"\|",4x,f7.3,/),'
673	WRITE(cllignes(4)(1:6),'(I6)') jpnij
674	cllignes(5)='1x,"Total \|",f12.3,6x,"\|",F12.3,2x,"\|",4x,f7.3,/,'
675	cllignes(6)='1x,"Minimum \|",f12.3,6x,"\|",F12.3,2x,"\|",4x,f7.3,/,'
676	cllignes(7)='1x,"Maximum \|",f12.3,6x,"\|",F12.3,2x,"\|",4x,f7.3,/,'
677	cllignes(8)='1x,"Average \|",f12.3,6x,"\|",F12.3,2x,"\|",4x,f7.3)'
678	clfmt=TRIM(cllignes(1))// TRIM(cllignes(2))//TRIM(cllignes(3))// &
679	& TRIM(clhline)//TRIM(cllignes(4))//TRIM(clhline)//TRIM(cllignes(5))// &
680	& TRIM(clhline)//TRIM(cllignes(6))//TRIM(clhline)//TRIM(cllignes(7))// &
681	& TRIM(clhline)//TRIM(cllignes(8))
682	WRITE(numtime, TRIM(clfmt)) &
683	(iall_rank(idum),all_etime(idum),all_ctime(idum),zall_ratio(idum),idum=1, jpnij), &
684	tot_etime, tot_ctime, ztot_ratio, &
685	zmin_etime, zmin_ctime, zmin_ratio, &
686	zmax_etime, zmax_ctime, zmax_ratio, &
687	zavg_etime, zavg_ctime, zavg_ratio
688	WRITE(numtime,*)
689	END IF
690	!
691	DEALLOCATE(zall_ratio, iall_rank)
692	!
693	END SUBROUTINE wmpi_info
694	#endif
695
696
697	SUBROUTINE timing_ini_var(cdinfo)
698	!!----------------------------------------------------------------------
699	!! * ROUTINE timing_ini_var *
700	!! ** Purpose : create timing structure
701	!!----------------------------------------------------------------------
702	CHARACTER(len=*), INTENT(in) :: cdinfo
703	LOGICAL :: ll_section
704
705	!
706	IF( .NOT. ASSOCIATED(s_timer_root) ) THEN
707	ALLOCATE(s_timer_root)
708	s_timer_root%cname = cdinfo
709	s_timer_root%t_cpu = 0._wp
710	s_timer_root%t_clock = 0._wp
711	s_timer_root%tsum_cpu = 0._wp
712	s_timer_root%tsum_clock = 0._wp
713	s_timer_root%tmax_cpu = 0._wp
714	s_timer_root%tmax_clock = 0._wp
715	s_timer_root%tmin_cpu = 0._wp
716	s_timer_root%tmin_clock = 0._wp
717	s_timer_root%tsub_cpu = 0._wp
718	s_timer_root%tsub_clock = 0._wp
719	s_timer_root%ncount = 0
720	s_timer_root%ncount_rate = 0
721	s_timer_root%ncount_max = 0
722	s_timer_root%niter = 0
723	s_timer_root%l_tdone = .FALSE.
724	s_timer_root%next => NULL()
725	s_timer_root%prev => NULL()
726	s_timer => s_timer_root
727	!
728	ALLOCATE(s_wrk)
729	s_wrk => NULL()
730	!
731	ALLOCATE(s_timer_old)
732	s_timer_old%cname = cdinfo
733	s_timer_old%t_cpu = 0._wp
734	s_timer_old%t_clock = 0._wp
735	s_timer_old%tsum_cpu = 0._wp
736	s_timer_old%tsum_clock = 0._wp
737	s_timer_old%tmax_cpu = 0._wp
738	s_timer_old%tmax_clock = 0._wp
739	s_timer_old%tmin_cpu = 0._wp
740	s_timer_old%tmin_clock = 0._wp
741	s_timer_old%tsub_cpu = 0._wp
742	s_timer_old%tsub_clock = 0._wp
743	s_timer_old%ncount = 0
744	s_timer_old%ncount_rate = 0
745	s_timer_old%ncount_max = 0
746	s_timer_old%niter = 0
747	s_timer_old%l_tdone = .TRUE.
748	s_timer_old%next => NULL()
749	s_timer_old%prev => NULL()
750
751	ELSE
752	s_timer => s_timer_root
753	! case of already existing area (typically inside a loop)
754	! write(,) 'in ini_var for routine : ', cdinfo
755	DO WHILE( ASSOCIATED(s_timer) )
756	IF( TRIM(s_timer%cname) .EQ. TRIM(cdinfo) ) THEN
757	! write(,) 'in ini_var for routine : ', cdinfo,' we return'
758	RETURN ! cdinfo is already in the chain
759	ENDIF
760	s_timer => s_timer%next
761	END DO
762
763	! end of the chain
764	s_timer => s_timer_root
765	DO WHILE( ASSOCIATED(s_timer%next) )
766	s_timer => s_timer%next
767	END DO
768
769	! write(,) 'after search', s_timer%cname
770	! cdinfo is not part of the chain so we add it with initialisation
771	ALLOCATE(s_timer%next)
772	! write(,) 'after allocation of next'
773
774	s_timer%next%cname = cdinfo
775	s_timer%next%t_cpu = 0._wp
776	s_timer%next%t_clock = 0._wp
777	s_timer%next%tsum_cpu = 0._wp
778	s_timer%next%tsum_clock = 0._wp
779	s_timer%next%tmax_cpu = 0._wp
780	s_timer%next%tmax_clock = 0._wp
781	s_timer%next%tmin_cpu = 0._wp
782	s_timer%next%tmin_clock = 0._wp
783	s_timer%next%tsub_cpu = 0._wp
784	s_timer%next%tsub_clock = 0._wp
785	s_timer%next%ncount = 0
786	s_timer%next%ncount_rate = 0
787	s_timer%next%ncount_max = 0
788	s_timer%next%niter = 0
789	s_timer%next%l_tdone = .FALSE.
790	s_timer%next%parent_section => NULL()
791	s_timer%next%prev => s_timer
792	s_timer%next%next => NULL()
793	s_timer => s_timer%next
794	ENDIF
795	! write(,) 'after allocation'
796	!
797	END SUBROUTINE timing_ini_var
798
799
800	SUBROUTINE timing_reset
801	!!----------------------------------------------------------------------
802	!! * ROUTINE timing_reset *
803	!! ** Purpose : go to root of timing tree
804	!!----------------------------------------------------------------------
805	l_initdone = .TRUE.
806	! IF(lwp) WRITE(numout,*)
807	! IF(lwp) WRITE(numout,*) 'timing_reset : instrumented routines for timing'
808	! IF(lwp) WRITE(numout,*) '~~~~~~~~~~~~'
809	CALL timing_list(s_timer_root)
810	! WRITE(numout,*)
811	!
812	END SUBROUTINE timing_reset
813
814
815	RECURSIVE SUBROUTINE timing_list(ptr)
816
817	TYPE(timer), POINTER, INTENT(inout) :: ptr
818	!
819	IF( ASSOCIATED(ptr%next) ) CALL timing_list(ptr%next)
820	IF(lwp) WRITE(numout,*)' ', ptr%cname
821	!
822	END SUBROUTINE timing_list
823
824
825	SUBROUTINE insert(sd_current, sd_root ,sd_ptr)
826	!!----------------------------------------------------------------------
827	!! * ROUTINE insert *
828	!! ** Purpose : insert an element in timer structure
829	!!----------------------------------------------------------------------
830	TYPE(timer), POINTER, INTENT(inout) :: sd_current, sd_root, sd_ptr
831	!
832
833	IF( ASSOCIATED( sd_current, sd_root ) ) THEN
834	! If our current element is the root element then
835	! replace it with the one being inserted
836	sd_root => sd_ptr
837	ELSE
838	sd_current%prev%next => sd_ptr
839	END IF
840	sd_ptr%next => sd_current
841	sd_ptr%prev => sd_current%prev
842	sd_current%prev => sd_ptr
843	! Nullify the pointer to the new element now that it is held
844	! within the list. If we don't do this then a subsequent call
845	! to ALLOCATE memory to this pointer will fail.
846	sd_ptr => NULL()
847	!
848	END SUBROUTINE insert
849
850
851	SUBROUTINE suppress(sd_ptr)
852	!!----------------------------------------------------------------------
853	!! * ROUTINE suppress *
854	!! ** Purpose : supress an element in timer structure
855	!!----------------------------------------------------------------------
856	TYPE(timer), POINTER, INTENT(inout) :: sd_ptr
857	!
858	TYPE(timer), POINTER :: sl_temp
859
860	sl_temp => sd_ptr
861	sd_ptr => sd_ptr%next
862	IF ( ASSOCIATED(sl_temp%next) ) sl_temp%next%prev => sl_temp%prev
863	DEALLOCATE(sl_temp)
864	sl_temp => NULL()
865	!
866	END SUBROUTINE suppress
867
868	!!=====================================================================
869	END MODULE timing

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

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

Context Navigation

source: NEMO/branches/UKMO/NEMO_4.0.4_EAP_rheology_fix/src/OCE/timing.F90 @ 15742

Download in other formats: