Context Navigation

← Previous Revision
Latest Revision
Next Revision →
Blame
Revision Log

timing.F90 @ 13442

Last change on this file since 13442 was 11646, checked in by andmirek, 5 years ago
Ticket #2314 changes for GO8 for tests
File size: 35.8 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	CHARACTER(len=128) :: ctiming_report_name ! name of timing report file
224
225	ctiming_report_name="timing.output"
226	IF( TRIM(cxios_context) == 'sas' ) WRITE(ctiming_report_name,'(A4,A13)') 'sas_', TRIM(ctiming_report_name)
227
228	IF( ln_onefile ) THEN
229	IF( lwp) CALL ctl_opn( numtime, TRIM(ctiming_report_name), 'REPLACE', 'FORMATTED', 'SEQUENTIAL', -1, numout,.TRUE., narea )
230	lwriter = lwp
231	ELSE
232	CALL ctl_opn( numtime, TRIM(ctiming_report_name), 'REPLACE', 'FORMATTED', 'SEQUENTIAL', -1, numout,.FALSE., narea )
233	lwriter = .TRUE.
234	ENDIF
235
236	IF( lwriter) THEN
237	WRITE(numtime,*)
238	WRITE(numtime,*) ' CNRS - NERC - Met OFFICE - MERCATOR-ocean - CMCC - INGV'
239	WRITE(numtime,*) ' NEMO team'
240	WRITE(numtime,*) ' Ocean General Circulation Model'
241	WRITE(numtime,*) ' version 4.0 (2019) '
242	WRITE(numtime,*)
243	WRITE(numtime,*) ' Timing Informations '
244	WRITE(numtime,*)
245	WRITE(numtime,*)
246	ENDIF
247
248	! Compute clock function overhead
249	#if defined key_mpp_mpi
250	t_overclock = MPI_WTIME()
251	t_overclock = MPI_WTIME() - t_overclock
252	#else
253	CALL SYSTEM_CLOCK(COUNT_RATE=ncount_rate, COUNT_MAX=ncount_max)
254	CALL SYSTEM_CLOCK(COUNT = istart_count)
255	CALL SYSTEM_CLOCK(COUNT = ifinal_count)
256	iperiods = ifinal_count - istart_count
257	IF( ifinal_count < istart_count ) &
258	iperiods = iperiods + ncount_max
259	t_overclock = REAL(iperiods) / ncount_rate
260	#endif
261
262	! Compute cpu_time function overhead
263	CALL CPU_TIME(zdum)
264	CALL CPU_TIME(t_overcpu)
265
266	! End overhead omputation
267	t_overcpu = t_overcpu - zdum
268	t_overclock = t_overcpu + t_overclock
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, ztot
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	t_elaps(2) = MPI_WTIME() - t_elaps(1) - t_overclock
307	#else
308	CALL SYSTEM_CLOCK(COUNT = nfinal_count)
309	iperiods = nfinal_count - ncount
310	IF( nfinal_count < ncount ) &
311	iperiods = iperiods + ncount_max
312	t_elaps(2) = REAL(iperiods) / ncount_rate - t_overclock
313	#endif
314
315	! End of timings on date & time
316	CALL DATE_AND_TIME(cdate(2),ctime(2),czone,nvalues)
317
318	! Compute the numer of routines
319	nsize = 0
320	s_timer => s_timer_root
321	DO WHILE( ASSOCIATED(s_timer) )
322	nsize = nsize + 1
323	s_timer => s_timer%next
324	END DO
325	idum = nsize
326	CALL mpp_sum('timing', idum)
327	IF( idum/jpnij /= nsize ) THEN
328	IF( lwriter ) WRITE(numtime,*) ' ===> W A R N I N G: '
329	IF( lwriter ) WRITE(numtime,*) ' Some CPU have different number of routines instrumented for timing'
330	IF( lwriter ) WRITE(numtime,*) ' No detailed report on averaged timing can be provided'
331	IF( lwriter ) WRITE(numtime,*) ' The following detailed report only deals with the current processor'
332	IF( lwriter ) WRITE(numtime,*)
333	ll_averep = .FALSE.
334	ENDIF
335
336	#if defined key_mpp_mpi
337	! in MPI gather some info
338	ALLOCATE( all_etime(jpnij), all_ctime(jpnij) )
339	CALL MPI_ALLGATHER(t_elaps(2), 1, MPI_DOUBLE_PRECISION, &
340	all_etime , 1, MPI_DOUBLE_PRECISION, &
341	MPI_COMM_OCE, icode)
342	CALL MPI_ALLGATHER(t_cpu(2) , 1, MPI_DOUBLE_PRECISION, &
343	all_ctime, 1, MPI_DOUBLE_PRECISION, &
344	MPI_COMM_OCE, icode)
345	tot_etime = SUM(all_etime(:))
346	tot_ctime = SUM(all_ctime(:))
347	#else
348	tot_etime = t_elaps(2)
349	tot_ctime = t_cpu (2)
350	#endif
351
352	! write output file
353	IF( lwriter ) WRITE(numtime,*) 'Total timing (sum) :'
354	IF( lwriter ) WRITE(numtime,*) '--------------------'
355	IF( lwriter ) WRITE(numtime,"('Elapsed Time (s) CPU Time (s)')")
356	IF( lwriter ) WRITE(numtime,'(5x,f12.3,1x,f12.3)') tot_etime, tot_ctime
357	IF( lwriter ) WRITE(numtime,*)
358	#if defined key_mpp_mpi
359	IF( ll_averep ) CALL waver_info
360	CALL wmpi_info
361	#endif
362	IF( lwriter ) CALL wcurrent_info
363
364	clfmt='(1X,"Timing started on ",2(A2,"/"),A4," at ",2(A2,":"),A2," MET ",A3,":",A2," from GMT")'
365	IF( lwriter ) WRITE(numtime, TRIM(clfmt)) &
366	& cdate(1)(7:8), cdate(1)(5:6), cdate(1)(1:4), &
367	& ctime(1)(1:2), ctime(1)(3:4), ctime(1)(5:6), &
368	& czone(1:3), czone(4:5)
369	clfmt='(1X, "Timing ended on ",2(A2,"/"),A4," at ",2(A2,":"),A2," MET ",A3,":",A2," from GMT")'
370	IF( lwriter ) WRITE(numtime, TRIM(clfmt)) &
371	& cdate(2)(7:8), cdate(2)(5:6), cdate(2)(1:4), &
372	& ctime(2)(1:2), ctime(2)(3:4), ctime(2)(5:6), &
373	& czone(1:3), czone(4:5)
374
375	#if defined key_mpp_mpi
376	ALLOCATE(timing_glob(4*jpnij), stat=icode)
377	CALL MPI_GATHER( (/compute_time, waiting_time(1), waiting_time(2), elapsed_time/), &
378	& 4, MPI_DOUBLE_PRECISION, timing_glob, 4, MPI_DOUBLE_PRECISION, 0, MPI_COMM_OCE, icode)
379	IF( narea == 1 ) THEN
380	WRITE(numtime,*) ' '
381	WRITE(numtime,*) ' Report on time spent on waiting MPI messages '
382	WRITE(numtime,*) ' total timing measured between nit000+1 and nitend-1 '
383	WRITE(numtime,*) ' warning: includes restarts writing time if output before nitend... '
384	WRITE(numtime,*) ' '
385	DO ji = 1, jpnij
386	ztot = SUM( timing_glob(4ji-3:4ji-1) )
387	WRITE(numtime,'(A28,F11.6, A34,I8)') 'Computing time : ',timing_glob(4*ji-3), ' on MPI rank : ', ji
388	IF ( ztot /= 0. ) zperc = timing_glob(4ji-2) / ztot 100.
389	WRITE(numtime,'(A28,F11.6,A2, F4.1,A3,A25,I8)') 'Waiting lbc_lnk time : ',timing_glob(4*ji-2) &
390	& , ' (', zperc,' %)', ' on MPI rank : ', ji
391	IF ( ztot /= 0. ) zperc = timing_glob(4ji-1) / ztot 100.
392	WRITE(numtime,'(A28,F11.6,A2, F4.1,A3,A25,I8)') 'Waiting global time : ',timing_glob(4*ji-1) &
393	& , ' (', zperc,' %)', ' on MPI rank : ', ji
394	zsypd = rn_rdt * REAL(nitend-nit000-1, wp) / (timing_glob(4ji) 365.)
395	WRITE(numtime,'(A28,F11.6,A7,F10.3,A2,A15,I8)') 'Total time : ',timing_glob(4*ji ) &
396	& , ' (SYPD: ', zsypd, ')', ' on MPI rank : ', ji
397	END DO
398	ENDIF
399	DEALLOCATE(timing_glob)
400	#endif
401
402	IF( lwriter ) CLOSE(numtime)
403	!
404	END SUBROUTINE timing_finalize
405
406
407	SUBROUTINE wcurrent_info
408	!!----------------------------------------------------------------------
409	!! * ROUTINE wcurrent_info *
410	!! ** Purpose : compute and write timing output file
411	!!----------------------------------------------------------------------
412	LOGICAL :: ll_ord
413	CHARACTER(len=2048) :: clfmt
414
415	! reorder the current list by elapse time
416	s_wrk => NULL()
417	s_timer => s_timer_root
418	DO
419	ll_ord = .TRUE.
420	s_timer => s_timer_root
421	DO WHILE ( ASSOCIATED( s_timer%next ) )
422	IF (.NOT. ASSOCIATED(s_timer%next)) EXIT
423	IF ( s_timer%tsum_clock < s_timer%next%tsum_clock ) THEN
424	ALLOCATE(s_wrk)
425	s_wrk = s_timer%next
426	CALL insert (s_timer, s_timer_root, s_wrk)
427	CALL suppress(s_timer%next)
428	ll_ord = .FALSE.
429	CYCLE
430	ENDIF
431	IF( ASSOCIATED(s_timer%next) ) s_timer => s_timer%next
432	END DO
433	IF( ll_ord ) EXIT
434	END DO
435
436	! write current info
437	WRITE(numtime,*) 'Detailed timing for proc :', narea-1
438	WRITE(numtime,*) '--------------------------'
439	WRITE(numtime,*) 'Section ', &
440	& 'Elapsed Time (s) ','Elapsed Time (%) ', &
441	& 'CPU Time(s) ','CPU Time (%) ','CPU/Elapsed ','Frequency'
442	s_timer => s_timer_root
443	clfmt = '(1x,a,4x,f12.3,6x,f12.3,x,f12.3,2x,f12.3,6x,f7.3,2x,i9)'
444	DO WHILE ( ASSOCIATED(s_timer) )
445	WRITE(numtime,TRIM(clfmt)) s_timer%cname, &
446	& s_timer%tsum_clock,s_timer%tsum_clock*100./t_elaps(2), &
447	& s_timer%tsum_cpu ,s_timer%tsum_cpu*100./t_cpu(2) , &
448	& s_timer%tsum_cpu/s_timer%tsum_clock, s_timer%niter
449	s_timer => s_timer%next
450	END DO
451	WRITE(numtime,*)
452	!
453	END SUBROUTINE wcurrent_info
454
455	#if defined key_mpp_mpi
456	SUBROUTINE waver_info
457	!!----------------------------------------------------------------------
458	!! * ROUTINE wcurrent_info *
459	!! ** Purpose : compute and write averaged timing informations
460	!!----------------------------------------------------------------------
461	TYPE(alltimer), POINTER :: sl_timer_glob_root => NULL()
462	TYPE(alltimer), POINTER :: sl_timer_glob => NULL()
463	TYPE(timer), POINTER :: sl_timer_ave_root => NULL()
464	TYPE(timer), POINTER :: sl_timer_ave => NULL()
465	INTEGER :: icode
466	INTEGER :: ierr
467	LOGICAL :: ll_ord
468	CHARACTER(len=200) :: clfmt
469
470	! Initialised the global strucutre
471	ALLOCATE(sl_timer_glob_root, Stat=ierr)
472	IF(ierr /= 0)THEN
473	WRITE(numtime,*) 'Failed to allocate global timing structure in waver_info'
474	RETURN
475	END IF
476
477	ALLOCATE(sl_timer_glob_root%cname (jpnij), &
478	sl_timer_glob_root%tsum_cpu (jpnij), &
479	sl_timer_glob_root%tsum_clock(jpnij), &
480	sl_timer_glob_root%niter (jpnij), Stat=ierr)
481	IF(ierr /= 0)THEN
482	WRITE(numtime,*) 'Failed to allocate global timing structure in waver_info'
483	RETURN
484	END IF
485	sl_timer_glob_root%cname(:) = ''
486	sl_timer_glob_root%tsum_cpu(:) = 0._wp
487	sl_timer_glob_root%tsum_clock(:) = 0._wp
488	sl_timer_glob_root%niter(:) = 0
489	sl_timer_glob_root%next => NULL()
490	sl_timer_glob_root%prev => NULL()
491	!ARPDBG - don't need to allocate a pointer that's immediately then
492	! set to point to some other object.
493	!ALLOCATE(sl_timer_glob)
494	!ALLOCATE(sl_timer_glob%cname (jpnij))
495	!ALLOCATE(sl_timer_glob%tsum_cpu (jpnij))
496	!ALLOCATE(sl_timer_glob%tsum_clock(jpnij))
497	!ALLOCATE(sl_timer_glob%niter (jpnij))
498	sl_timer_glob => sl_timer_glob_root
499	!
500	IF( narea .EQ. 1 ) THEN
501	ALLOCATE(sl_timer_ave_root)
502	sl_timer_ave_root%cname = ''
503	sl_timer_ave_root%t_cpu = 0._wp
504	sl_timer_ave_root%t_clock = 0._wp
505	sl_timer_ave_root%tsum_cpu = 0._wp
506	sl_timer_ave_root%tsum_clock = 0._wp
507	sl_timer_ave_root%tmax_cpu = 0._wp
508	sl_timer_ave_root%tmax_clock = 0._wp
509	sl_timer_ave_root%tmin_cpu = 0._wp
510	sl_timer_ave_root%tmin_clock = 0._wp
511	sl_timer_ave_root%tsub_cpu = 0._wp
512	sl_timer_ave_root%tsub_clock = 0._wp
513	sl_timer_ave_root%ncount = 0
514	sl_timer_ave_root%ncount_rate = 0
515	sl_timer_ave_root%ncount_max = 0
516	sl_timer_ave_root%niter = 0
517	sl_timer_ave_root%l_tdone = .FALSE.
518	sl_timer_ave_root%next => NULL()
519	sl_timer_ave_root%prev => NULL()
520	ALLOCATE(sl_timer_ave)
521	sl_timer_ave => sl_timer_ave_root
522	ENDIF
523
524	! Gather info from all processors
525	s_timer => s_timer_root
526	DO WHILE ( ASSOCIATED(s_timer) )
527	CALL MPI_GATHER(s_timer%cname , 20, MPI_CHARACTER, &
528	sl_timer_glob%cname, 20, MPI_CHARACTER, &
529	0, MPI_COMM_OCE, icode)
530	CALL MPI_GATHER(s_timer%tsum_clock , 1, MPI_DOUBLE_PRECISION, &
531	sl_timer_glob%tsum_clock, 1, MPI_DOUBLE_PRECISION, &
532	0, MPI_COMM_OCE, icode)
533	CALL MPI_GATHER(s_timer%tsum_cpu , 1, MPI_DOUBLE_PRECISION, &
534	sl_timer_glob%tsum_cpu, 1, MPI_DOUBLE_PRECISION, &
535	0, MPI_COMM_OCE, icode)
536	CALL MPI_GATHER(s_timer%niter , 1, MPI_INTEGER, &
537	sl_timer_glob%niter, 1, MPI_INTEGER, &
538	0, MPI_COMM_OCE, icode)
539
540	IF( narea == 1 .AND. ASSOCIATED(s_timer%next) ) THEN
541	ALLOCATE(sl_timer_glob%next)
542	ALLOCATE(sl_timer_glob%next%cname (jpnij))
543	ALLOCATE(sl_timer_glob%next%tsum_cpu (jpnij))
544	ALLOCATE(sl_timer_glob%next%tsum_clock(jpnij))
545	ALLOCATE(sl_timer_glob%next%niter (jpnij))
546	sl_timer_glob%next%prev => sl_timer_glob
547	sl_timer_glob%next%next => NULL()
548	sl_timer_glob => sl_timer_glob%next
549	ENDIF
550	s_timer => s_timer%next
551	END DO
552
553	IF( narea == 1 ) THEN
554	! Compute some stats
555	sl_timer_glob => sl_timer_glob_root
556	DO WHILE( ASSOCIATED(sl_timer_glob) )
557	sl_timer_ave%cname = sl_timer_glob%cname(1)
558	sl_timer_ave%tsum_cpu = SUM (sl_timer_glob%tsum_cpu (:)) / jpnij
559	sl_timer_ave%tsum_clock = SUM (sl_timer_glob%tsum_clock(:)) / jpnij
560	sl_timer_ave%tmax_cpu = MAXVAL(sl_timer_glob%tsum_cpu (:))
561	sl_timer_ave%tmax_clock = MAXVAL(sl_timer_glob%tsum_clock(:))
562	sl_timer_ave%tmin_cpu = MINVAL(sl_timer_glob%tsum_cpu (:))
563	sl_timer_ave%tmin_clock = MINVAL(sl_timer_glob%tsum_clock(:))
564	sl_timer_ave%niter = SUM (sl_timer_glob%niter (:))
565	!
566	IF( ASSOCIATED(sl_timer_glob%next) ) THEN
567	ALLOCATE(sl_timer_ave%next)
568	sl_timer_ave%next%prev => sl_timer_ave
569	sl_timer_ave%next%next => NULL()
570	sl_timer_ave => sl_timer_ave%next
571	ENDIF
572	sl_timer_glob => sl_timer_glob%next
573	END DO
574
575	! reorder the averaged list by CPU time
576	s_wrk => NULL()
577	sl_timer_ave => sl_timer_ave_root
578	DO
579	ll_ord = .TRUE.
580	sl_timer_ave => sl_timer_ave_root
581	DO WHILE( ASSOCIATED( sl_timer_ave%next ) )
582
583	IF( .NOT. ASSOCIATED(sl_timer_ave%next) ) EXIT
584
585	IF ( sl_timer_ave%tsum_clock < sl_timer_ave%next%tsum_clock ) THEN
586	ALLOCATE(s_wrk)
587	! Copy data into the new object pointed to by s_wrk
588	s_wrk = sl_timer_ave%next
589	! Insert this new timer object before our current position
590	CALL insert (sl_timer_ave, sl_timer_ave_root, s_wrk)
591	! Remove the old object from the list
592	CALL suppress(sl_timer_ave%next)
593	ll_ord = .FALSE.
594	CYCLE
595	ENDIF
596	IF( ASSOCIATED(sl_timer_ave%next) ) sl_timer_ave => sl_timer_ave%next
597	END DO
598	IF( ll_ord ) EXIT
599	END DO
600
601	! write averaged info
602	WRITE(numtime,"('Averaged timing on all processors :')")
603	WRITE(numtime,"('-----------------------------------')")
604	WRITE(numtime,"('Section',13x,'Elap. Time(s)',2x,'Elap. Time(%)',2x, &
605	& 'CPU Time(s)',2x,'CPU Time(%)',2x,'CPU/Elap',1x, &
606	& 'Max elap(%)',2x,'Min elap(%)',2x, &
607	& 'Freq')")
608	sl_timer_ave => sl_timer_ave_root
609	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)'
610	DO WHILE ( ASSOCIATED(sl_timer_ave) )
611	WRITE(numtime,TRIM(clfmt)) sl_timer_ave%cname(1:18), &
612	& sl_timer_ave%tsum_clock,sl_timer_ave%tsum_clock100.jpnij/tot_etime, &
613	& sl_timer_ave%tsum_cpu ,sl_timer_ave%tsum_cpu100.jpnij/tot_ctime , &
614	& sl_timer_ave%tsum_cpu/sl_timer_ave%tsum_clock, &
615	& sl_timer_ave%tmax_clock100.jpnij/tot_etime, &
616	& sl_timer_ave%tmin_clock100.jpnij/tot_etime, &
617	& sl_timer_ave%niter/REAL(jpnij)
618	sl_timer_ave => sl_timer_ave%next
619	END DO
620	WRITE(numtime,*)
621	!
622	DEALLOCATE(sl_timer_ave_root)
623	ENDIF
624	!
625	DEALLOCATE(sl_timer_glob_root)
626	!
627	END SUBROUTINE waver_info
628
629
630	SUBROUTINE wmpi_info
631	!!----------------------------------------------------------------------
632	!! * ROUTINE wmpi_time *
633	!! ** Purpose : compute and write a summary of MPI infos
634	!!----------------------------------------------------------------------
635	!
636	INTEGER :: idum, icode
637	INTEGER, ALLOCATABLE, DIMENSION(:) :: iall_rank
638	REAL(wp) :: ztot_ratio
639	REAL(wp) :: zmax_etime, zmax_ctime, zmax_ratio, zmin_etime, zmin_ctime, zmin_ratio
640	REAL(wp) :: zavg_etime, zavg_ctime, zavg_ratio
641	REAL(wp), ALLOCATABLE, DIMENSION(:) :: zall_ratio
642	CHARACTER(LEN=128), dimension(8) :: cllignes
643	CHARACTER(LEN=128) :: clhline, clstart_date, clfinal_date
644	CHARACTER(LEN=2048) :: clfmt
645
646	! Gather all times
647	ALLOCATE( zall_ratio(jpnij), iall_rank(jpnij) )
648	IF( narea == 1 ) THEN
649	iall_rank(:) = (/ (idum,idum=0,jpnij-1) /)
650
651	! Compute elapse user time
652	zavg_etime = tot_etime/REAL(jpnij,wp)
653	zmax_etime = MAXVAL(all_etime(:))
654	zmin_etime = MINVAL(all_etime(:))
655
656	! Compute CPU user time
657	zavg_ctime = tot_ctime/REAL(jpnij,wp)
658	zmax_ctime = MAXVAL(all_ctime(:))
659	zmin_ctime = MINVAL(all_ctime(:))
660
661	! Compute cpu/elapsed ratio
662	zall_ratio(:) = all_ctime(:) / all_etime(:)
663	ztot_ratio = SUM(zall_ratio(:))
664	zavg_ratio = ztot_ratio/REAL(jpnij,wp)
665	zmax_ratio = MAXVAL(zall_ratio(:))
666	zmin_ratio = MINVAL(zall_ratio(:))
667
668	! Output Format
669	clhline ='1x,13("-"),"\|",18("-"),"\|",14("-"),"\|",18("-"),/,'
670	cllignes(1)='(1x,"MPI summary report :",/,'
671	cllignes(2)='1x,"--------------------",//,'
672	cllignes(3)='1x,"Process Rank \|"," Elapsed Time (s) \|"," CPU Time (s) \|"," Ratio CPU/Elapsed",/,'
673	cllignes(4)=' (1x,i4,9x,"\|",f12.3,6x,"\|",f12.3,2x,"\|",4x,f7.3,/),'
674	WRITE(cllignes(4)(1:4),'(I4)') jpnij
675	cllignes(5)='1x,"Total \|",f12.3,6x,"\|",F12.3,2x,"\|",4x,f7.3,/,'
676	cllignes(6)='1x,"Minimum \|",f12.3,6x,"\|",F12.3,2x,"\|",4x,f7.3,/,'
677	cllignes(7)='1x,"Maximum \|",f12.3,6x,"\|",F12.3,2x,"\|",4x,f7.3,/,'
678	cllignes(8)='1x,"Average \|",f12.3,6x,"\|",F12.3,2x,"\|",4x,f7.3)'
679	clfmt=TRIM(cllignes(1))// TRIM(cllignes(2))//TRIM(cllignes(3))// &
680	& TRIM(clhline)//TRIM(cllignes(4))//TRIM(clhline)//TRIM(cllignes(5))// &
681	& TRIM(clhline)//TRIM(cllignes(6))//TRIM(clhline)//TRIM(cllignes(7))// &
682	& TRIM(clhline)//TRIM(cllignes(8))
683	WRITE(numtime, TRIM(clfmt)) &
684	(iall_rank(idum),all_etime(idum),all_ctime(idum),zall_ratio(idum),idum=1, jpnij), &
685	tot_etime, tot_ctime, ztot_ratio, &
686	zmin_etime, zmin_ctime, zmin_ratio, &
687	zmax_etime, zmax_ctime, zmax_ratio, &
688	zavg_etime, zavg_ctime, zavg_ratio
689	WRITE(numtime,*)
690	END IF
691	!
692	DEALLOCATE(zall_ratio, iall_rank)
693	!
694	END SUBROUTINE wmpi_info
695	#endif
696
697
698	SUBROUTINE timing_ini_var(cdinfo)
699	!!----------------------------------------------------------------------
700	!! * ROUTINE timing_ini_var *
701	!! ** Purpose : create timing structure
702	!!----------------------------------------------------------------------
703	CHARACTER(len=*), INTENT(in) :: cdinfo
704	LOGICAL :: ll_section
705
706	!
707	IF( .NOT. ASSOCIATED(s_timer_root) ) THEN
708	ALLOCATE(s_timer_root)
709	s_timer_root%cname = cdinfo
710	s_timer_root%t_cpu = 0._wp
711	s_timer_root%t_clock = 0._wp
712	s_timer_root%tsum_cpu = 0._wp
713	s_timer_root%tsum_clock = 0._wp
714	s_timer_root%tmax_cpu = 0._wp
715	s_timer_root%tmax_clock = 0._wp
716	s_timer_root%tmin_cpu = 0._wp
717	s_timer_root%tmin_clock = 0._wp
718	s_timer_root%tsub_cpu = 0._wp
719	s_timer_root%tsub_clock = 0._wp
720	s_timer_root%ncount = 0
721	s_timer_root%ncount_rate = 0
722	s_timer_root%ncount_max = 0
723	s_timer_root%niter = 0
724	s_timer_root%l_tdone = .FALSE.
725	s_timer_root%next => NULL()
726	s_timer_root%prev => NULL()
727	s_timer => s_timer_root
728	!
729	ALLOCATE(s_wrk)
730	s_wrk => NULL()
731	!
732	ALLOCATE(s_timer_old)
733	s_timer_old%cname = cdinfo
734	s_timer_old%t_cpu = 0._wp
735	s_timer_old%t_clock = 0._wp
736	s_timer_old%tsum_cpu = 0._wp
737	s_timer_old%tsum_clock = 0._wp
738	s_timer_old%tmax_cpu = 0._wp
739	s_timer_old%tmax_clock = 0._wp
740	s_timer_old%tmin_cpu = 0._wp
741	s_timer_old%tmin_clock = 0._wp
742	s_timer_old%tsub_cpu = 0._wp
743	s_timer_old%tsub_clock = 0._wp
744	s_timer_old%ncount = 0
745	s_timer_old%ncount_rate = 0
746	s_timer_old%ncount_max = 0
747	s_timer_old%niter = 0
748	s_timer_old%l_tdone = .TRUE.
749	s_timer_old%next => NULL()
750	s_timer_old%prev => NULL()
751
752	ELSE
753	s_timer => s_timer_root
754	! case of already existing area (typically inside a loop)
755	! write(,) 'in ini_var for routine : ', cdinfo
756	DO WHILE( ASSOCIATED(s_timer) )
757	IF( TRIM(s_timer%cname) .EQ. TRIM(cdinfo) ) THEN
758	! write(,) 'in ini_var for routine : ', cdinfo,' we return'
759	RETURN ! cdinfo is already in the chain
760	ENDIF
761	s_timer => s_timer%next
762	END DO
763
764	! end of the chain
765	s_timer => s_timer_root
766	DO WHILE( ASSOCIATED(s_timer%next) )
767	s_timer => s_timer%next
768	END DO
769
770	! write(,) 'after search', s_timer%cname
771	! cdinfo is not part of the chain so we add it with initialisation
772	ALLOCATE(s_timer%next)
773	! write(,) 'after allocation of next'
774
775	s_timer%next%cname = cdinfo
776	s_timer%next%t_cpu = 0._wp
777	s_timer%next%t_clock = 0._wp
778	s_timer%next%tsum_cpu = 0._wp
779	s_timer%next%tsum_clock = 0._wp
780	s_timer%next%tmax_cpu = 0._wp
781	s_timer%next%tmax_clock = 0._wp
782	s_timer%next%tmin_cpu = 0._wp
783	s_timer%next%tmin_clock = 0._wp
784	s_timer%next%tsub_cpu = 0._wp
785	s_timer%next%tsub_clock = 0._wp
786	s_timer%next%ncount = 0
787	s_timer%next%ncount_rate = 0
788	s_timer%next%ncount_max = 0
789	s_timer%next%niter = 0
790	s_timer%next%l_tdone = .FALSE.
791	s_timer%next%parent_section => NULL()
792	s_timer%next%prev => s_timer
793	s_timer%next%next => NULL()
794	s_timer => s_timer%next
795	ENDIF
796	! write(,) 'after allocation'
797	!
798	END SUBROUTINE timing_ini_var
799
800
801	SUBROUTINE timing_reset
802	!!----------------------------------------------------------------------
803	!! * ROUTINE timing_reset *
804	!! ** Purpose : go to root of timing tree
805	!!----------------------------------------------------------------------
806	l_initdone = .TRUE.
807	! IF(lwp) WRITE(numout,*)
808	! IF(lwp) WRITE(numout,*) 'timing_reset : instrumented routines for timing'
809	! IF(lwp) WRITE(numout,*) '~~~~~~~~~~~~'
810	CALL timing_list(s_timer_root)
811	! WRITE(numout,*)
812	!
813	END SUBROUTINE timing_reset
814
815
816	RECURSIVE SUBROUTINE timing_list(ptr)
817
818	TYPE(timer), POINTER, INTENT(inout) :: ptr
819	!
820	IF( ASSOCIATED(ptr%next) ) CALL timing_list(ptr%next)
821	IF(lwp) WRITE(numout,*)' ', ptr%cname
822	!
823	END SUBROUTINE timing_list
824
825
826	SUBROUTINE insert(sd_current, sd_root ,sd_ptr)
827	!!----------------------------------------------------------------------
828	!! * ROUTINE insert *
829	!! ** Purpose : insert an element in timer structure
830	!!----------------------------------------------------------------------
831	TYPE(timer), POINTER, INTENT(inout) :: sd_current, sd_root, sd_ptr
832	!
833
834	IF( ASSOCIATED( sd_current, sd_root ) ) THEN
835	! If our current element is the root element then
836	! replace it with the one being inserted
837	sd_root => sd_ptr
838	ELSE
839	sd_current%prev%next => sd_ptr
840	END IF
841	sd_ptr%next => sd_current
842	sd_ptr%prev => sd_current%prev
843	sd_current%prev => sd_ptr
844	! Nullify the pointer to the new element now that it is held
845	! within the list. If we don't do this then a subsequent call
846	! to ALLOCATE memory to this pointer will fail.
847	sd_ptr => NULL()
848	!
849	END SUBROUTINE insert
850
851
852	SUBROUTINE suppress(sd_ptr)
853	!!----------------------------------------------------------------------
854	!! * ROUTINE suppress *
855	!! ** Purpose : supress an element in timer structure
856	!!----------------------------------------------------------------------
857	TYPE(timer), POINTER, INTENT(inout) :: sd_ptr
858	!
859	TYPE(timer), POINTER :: sl_temp
860
861	sl_temp => sd_ptr
862	sd_ptr => sd_ptr%next
863	IF ( ASSOCIATED(sl_temp%next) ) sl_temp%next%prev => sl_temp%prev
864	DEALLOCATE(sl_temp)
865	sl_temp => NULL()
866	!
867	END SUBROUTINE suppress
868
869	!!=====================================================================
870	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_mirror_SI3_GPU/src/SAS/timing.F90 @ 13442

Download in other formats: