Context Navigation

← Previous Revision
Next Revision →
Blame
Revision Log

timing.F90

Last change on this file was 14834, checked in by hadcv, 3 years ago
#2600: Merge in dev_r14273_HPC-02_Daley_Tiling
Property svn:keywords set to `Id`
File size: 35.3 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_mpi_off
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	IF( .NOT. l_istiled .OR. ntile == 1 ) s_timer%niter = s_timer%niter + 1 ! All tiles count as one iteration
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_mpi_off
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_mpi_off
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_mpi_off
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( clname )
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=*), INTENT(in), OPTIONAL :: clname
224	CHARACTER(len=20) :: cln
225
226	IF( PRESENT(clname) ) THEN ; cln = clname
227	ELSE ; cln = 'timing.output'
228	ENDIF
229
230	IF( ln_onefile ) THEN
231	IF( lwp) CALL ctl_opn( numtime, cln, 'REPLACE', 'FORMATTED', 'SEQUENTIAL', -1, numout,.TRUE., narea )
232	lwriter = lwp
233	ELSE
234	CALL ctl_opn( numtime, cln, 'REPLACE', 'FORMATTED', 'SEQUENTIAL', -1, numout,.FALSE., narea )
235	lwriter = .TRUE.
236	ENDIF
237
238	IF( lwriter) THEN
239	WRITE(numtime,*)
240	WRITE(numtime,*) ' CNRS - NERC - Met OFFICE - MERCATOR-ocean - CMCC - INGV'
241	WRITE(numtime,*) ' NEMO team'
242	WRITE(numtime,*) ' Ocean General Circulation Model'
243	WRITE(numtime,*) ' version 4.0 (2019) '
244	WRITE(numtime,*)
245	WRITE(numtime,*) ' Timing Informations '
246	WRITE(numtime,*)
247	WRITE(numtime,*)
248	ENDIF
249
250	! Compute clock function overhead
251	#if ! defined key_mpi_off
252	t_overclock = MPI_WTIME()
253	t_overclock = MPI_WTIME() - t_overclock
254	#else
255	CALL SYSTEM_CLOCK(COUNT_RATE=ncount_rate, COUNT_MAX=ncount_max)
256	CALL SYSTEM_CLOCK(COUNT = istart_count)
257	CALL SYSTEM_CLOCK(COUNT = ifinal_count)
258	iperiods = ifinal_count - istart_count
259	IF( ifinal_count < istart_count ) &
260	iperiods = iperiods + ncount_max
261	t_overclock = REAL(iperiods) / ncount_rate
262	#endif
263
264	! Compute cpu_time function overhead
265	CALL CPU_TIME(zdum)
266	CALL CPU_TIME(t_overcpu)
267
268	! End overhead omputation
269	t_overcpu = t_overcpu - zdum
270	t_overclock = t_overcpu + t_overclock
271
272	! Timing on date and time
273	CALL DATE_AND_TIME(cdate(1),ctime(1),czone,nvalues)
274
275	CALL CPU_TIME(t_cpu(1))
276	#if ! defined key_mpi_off
277	! Start elapsed and CPU time counters
278	t_elaps(1) = MPI_WTIME()
279	#else
280	CALL SYSTEM_CLOCK(COUNT_RATE=ncount_rate, COUNT_MAX=ncount_max)
281	CALL SYSTEM_CLOCK(COUNT = ncount)
282	#endif
283	!
284	END SUBROUTINE timing_init
285
286
287	SUBROUTINE timing_finalize
288	!!----------------------------------------------------------------------
289	!! * ROUTINE timing_finalize *
290	!! ** Purpose : compute average time
291	!! write timing output file
292	!!----------------------------------------------------------------------
293	TYPE(timer), POINTER :: s_temp
294	INTEGER :: idum, iperiods, icode
295	INTEGER :: ji
296	LOGICAL :: ll_ord, ll_averep
297	CHARACTER(len=120) :: clfmt
298	REAL(wp), DIMENSION(:), ALLOCATABLE :: timing_glob
299	REAL(wp) :: zsypd ! simulated years per day (Balaji 2017)
300	REAL(wp) :: zperc, ztot
301
302	ll_averep = .TRUE.
303
304	! total CPU and elapse
305	CALL CPU_TIME(t_cpu(2))
306	t_cpu(2) = t_cpu(2) - t_cpu(1) - t_overcpu
307	#if ! defined key_mpi_off
308	t_elaps(2) = MPI_WTIME() - t_elaps(1) - t_overclock
309	#else
310	CALL SYSTEM_CLOCK(COUNT = nfinal_count)
311	iperiods = nfinal_count - ncount
312	IF( nfinal_count < ncount ) &
313	iperiods = iperiods + ncount_max
314	t_elaps(2) = REAL(iperiods) / ncount_rate - t_overclock
315	#endif
316
317	! End of timings on date & time
318	CALL DATE_AND_TIME(cdate(2),ctime(2),czone,nvalues)
319
320	! Compute the numer of routines
321	nsize = 0
322	s_timer => s_timer_root
323	DO WHILE( ASSOCIATED(s_timer) )
324	nsize = nsize + 1
325	s_timer => s_timer%next
326	END DO
327	idum = nsize
328	CALL mpp_sum('timing', idum)
329	IF( idum/jpnij /= nsize ) THEN
330	IF( lwriter ) WRITE(numtime,*) ' ===> W A R N I N G: '
331	IF( lwriter ) WRITE(numtime,*) ' Some CPU have different number of routines instrumented for timing'
332	IF( lwriter ) WRITE(numtime,*) ' No detailed report on averaged timing can be provided'
333	IF( lwriter ) WRITE(numtime,*) ' The following detailed report only deals with the current processor'
334	IF( lwriter ) WRITE(numtime,*)
335	ll_averep = .FALSE.
336	ENDIF
337
338	#if ! defined key_mpi_off
339	! in MPI gather some info
340	ALLOCATE( all_etime(jpnij), all_ctime(jpnij) )
341	CALL MPI_ALLGATHER(t_elaps(2), 1, MPI_DOUBLE_PRECISION, &
342	all_etime , 1, MPI_DOUBLE_PRECISION, &
343	MPI_COMM_OCE, icode)
344	CALL MPI_ALLGATHER(t_cpu(2) , 1, MPI_DOUBLE_PRECISION, &
345	all_ctime, 1, MPI_DOUBLE_PRECISION, &
346	MPI_COMM_OCE, icode)
347	tot_etime = SUM(all_etime(:))
348	tot_ctime = SUM(all_ctime(:))
349	#else
350	tot_etime = t_elaps(2)
351	tot_ctime = t_cpu (2)
352	#endif
353
354	! write output file
355	IF( lwriter ) WRITE(numtime,*)
356	IF( lwriter ) WRITE(numtime,*)
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_mpi_off
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_mpi_off
380	ALLOCATE(timing_glob(4*jpnij), stat=icode)
381	CALL MPI_GATHER( (/compute_time, waiting_time(1), waiting_time(2), elapsed_time/), &
382	& 4, MPI_DOUBLE_PRECISION, timing_glob, 4, 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	ztot = SUM( timing_glob(4ji-3:4ji-1) )
391	WRITE(numtime,'(A28,F11.6, A34,I8)') 'Computing time : ',timing_glob(4*ji-3), ' on MPI rank : ', ji
392	IF ( ztot /= 0. ) zperc = timing_glob(4ji-2) / ztot 100.
393	WRITE(numtime,'(A28,F11.6,A2, F4.1,A3,A25,I8)') 'Waiting lbc_lnk time : ',timing_glob(4*ji-2) &
394	& , ' (', zperc,' %)', ' on MPI rank : ', ji
395	IF ( ztot /= 0. ) zperc = timing_glob(4ji-1) / ztot 100.
396	WRITE(numtime,'(A28,F11.6,A2, F4.1,A3,A25,I8)') 'Waiting global time : ',timing_glob(4*ji-1) &
397	& , ' (', zperc,' %)', ' on MPI rank : ', ji
398	zsypd = rn_Dt * REAL(nitend-nit000-1, wp) / (timing_glob(4ji) 365.)
399	WRITE(numtime,'(A28,F11.6,A7,F10.3,A2,A15,I8)') 'Total time : ',timing_glob(4*ji ) &
400	& , ' (SYPD: ', zsypd, ')', ' on MPI rank : ', ji
401	END DO
402	ENDIF
403	DEALLOCATE(timing_glob)
404	#endif
405
406	IF( lwriter ) CLOSE(numtime)
407	!
408	END SUBROUTINE timing_finalize
409
410
411	SUBROUTINE wcurrent_info
412	!!----------------------------------------------------------------------
413	!! * ROUTINE wcurrent_info *
414	!! ** Purpose : compute and write timing output file
415	!!----------------------------------------------------------------------
416	LOGICAL :: ll_ord
417	CHARACTER(len=2048) :: clfmt
418
419	! reorder the current list by elapse time
420	s_wrk => NULL()
421	s_timer => s_timer_root
422	DO
423	ll_ord = .TRUE.
424	s_timer => s_timer_root
425	DO WHILE ( ASSOCIATED( s_timer%next ) )
426	IF (.NOT. ASSOCIATED(s_timer%next)) EXIT
427	IF ( s_timer%tsum_clock < s_timer%next%tsum_clock ) THEN
428	ALLOCATE(s_wrk)
429	s_wrk = s_timer%next
430	CALL insert (s_timer, s_timer_root, s_wrk)
431	CALL suppress(s_timer%next)
432	ll_ord = .FALSE.
433	CYCLE
434	ENDIF
435	IF( ASSOCIATED(s_timer%next) ) s_timer => s_timer%next
436	END DO
437	IF( ll_ord ) EXIT
438	END DO
439
440	! write current info
441	WRITE(numtime,*) 'Detailed timing for proc :', narea-1
442	WRITE(numtime,*) '--------------------------'
443	WRITE(numtime,*) 'Section ', &
444	& 'Elapsed Time (s) ','Elapsed Time (%) ', &
445	& 'CPU Time(s) ','CPU Time (%) ','CPU/Elapsed ','Frequency'
446	s_timer => s_timer_root
447	clfmt = '(1x,a,4x,f12.3,6x,f12.3,x,f12.3,2x,f12.3,6x,f7.3,2x,i9)'
448	DO WHILE ( ASSOCIATED(s_timer) )
449	IF( s_timer%tsum_clock > 0._wp ) &
450	WRITE(numtime,TRIM(clfmt)) s_timer%cname, &
451	& s_timer%tsum_clock,s_timer%tsum_clock*100./t_elaps(2), &
452	& s_timer%tsum_cpu ,s_timer%tsum_cpu*100./t_cpu(2) , &
453	& s_timer%tsum_cpu/s_timer%tsum_clock, s_timer%niter
454	s_timer => s_timer%next
455	END DO
456	WRITE(numtime,*)
457	!
458	END SUBROUTINE wcurrent_info
459
460	#if ! defined key_mpi_off
461	SUBROUTINE waver_info
462	!!----------------------------------------------------------------------
463	!! * ROUTINE wcurrent_info *
464	!! ** Purpose : compute and write averaged timing informations
465	!!----------------------------------------------------------------------
466	TYPE(alltimer), POINTER :: sl_timer_glob_root => NULL()
467	TYPE(alltimer), POINTER :: sl_timer_glob => NULL()
468	TYPE(timer), POINTER :: sl_timer_ave_root => NULL()
469	TYPE(timer), POINTER :: sl_timer_ave => NULL()
470	INTEGER :: icode
471	INTEGER :: ierr
472	LOGICAL :: ll_ord
473	CHARACTER(len=200) :: clfmt
474
475	! Initialised the global strucutre
476	ALLOCATE(sl_timer_glob_root, Stat=ierr)
477	IF(ierr /= 0)THEN
478	WRITE(numtime,*) 'Failed to allocate global timing structure in waver_info'
479	RETURN
480	END IF
481
482	ALLOCATE(sl_timer_glob_root%cname (jpnij), &
483	sl_timer_glob_root%tsum_cpu (jpnij), &
484	sl_timer_glob_root%tsum_clock(jpnij), &
485	sl_timer_glob_root%niter (jpnij), Stat=ierr)
486	IF(ierr /= 0)THEN
487	WRITE(numtime,*) 'Failed to allocate global timing structure in waver_info'
488	RETURN
489	END IF
490	sl_timer_glob_root%cname(:) = ''
491	sl_timer_glob_root%tsum_cpu(:) = 0._wp
492	sl_timer_glob_root%tsum_clock(:) = 0._wp
493	sl_timer_glob_root%niter(:) = 0
494	sl_timer_glob_root%next => NULL()
495	sl_timer_glob_root%prev => NULL()
496	!ARPDBG - don't need to allocate a pointer that's immediately then
497	! set to point to some other object.
498	!ALLOCATE(sl_timer_glob)
499	!ALLOCATE(sl_timer_glob%cname (jpnij))
500	!ALLOCATE(sl_timer_glob%tsum_cpu (jpnij))
501	!ALLOCATE(sl_timer_glob%tsum_clock(jpnij))
502	!ALLOCATE(sl_timer_glob%niter (jpnij))
503	sl_timer_glob => sl_timer_glob_root
504	!
505	IF( narea .EQ. 1 ) THEN
506	ALLOCATE(sl_timer_ave_root)
507	sl_timer_ave_root%cname = ''
508	sl_timer_ave_root%t_cpu = 0._wp
509	sl_timer_ave_root%t_clock = 0._wp
510	sl_timer_ave_root%tsum_cpu = 0._wp
511	sl_timer_ave_root%tsum_clock = 0._wp
512	sl_timer_ave_root%tmax_cpu = 0._wp
513	sl_timer_ave_root%tmax_clock = 0._wp
514	sl_timer_ave_root%tmin_cpu = 0._wp
515	sl_timer_ave_root%tmin_clock = 0._wp
516	sl_timer_ave_root%tsub_cpu = 0._wp
517	sl_timer_ave_root%tsub_clock = 0._wp
518	sl_timer_ave_root%ncount = 0
519	sl_timer_ave_root%ncount_rate = 0
520	sl_timer_ave_root%ncount_max = 0
521	sl_timer_ave_root%niter = 0
522	sl_timer_ave_root%l_tdone = .FALSE.
523	sl_timer_ave_root%next => NULL()
524	sl_timer_ave_root%prev => NULL()
525	ALLOCATE(sl_timer_ave)
526	sl_timer_ave => sl_timer_ave_root
527	ENDIF
528
529	! Gather info from all processors
530	s_timer => s_timer_root
531	DO WHILE ( ASSOCIATED(s_timer) )
532	CALL MPI_GATHER(s_timer%cname , 20, MPI_CHARACTER, &
533	sl_timer_glob%cname, 20, MPI_CHARACTER, &
534	0, MPI_COMM_OCE, icode)
535	CALL MPI_GATHER(s_timer%tsum_clock , 1, MPI_DOUBLE_PRECISION, &
536	sl_timer_glob%tsum_clock, 1, MPI_DOUBLE_PRECISION, &
537	0, MPI_COMM_OCE, icode)
538	CALL MPI_GATHER(s_timer%tsum_cpu , 1, MPI_DOUBLE_PRECISION, &
539	sl_timer_glob%tsum_cpu, 1, MPI_DOUBLE_PRECISION, &
540	0, MPI_COMM_OCE, icode)
541	CALL MPI_GATHER(s_timer%niter , 1, MPI_INTEGER, &
542	sl_timer_glob%niter, 1, MPI_INTEGER, &
543	0, MPI_COMM_OCE, icode)
544
545	IF( narea == 1 .AND. ASSOCIATED(s_timer%next) ) THEN
546	ALLOCATE(sl_timer_glob%next)
547	ALLOCATE(sl_timer_glob%next%cname (jpnij))
548	ALLOCATE(sl_timer_glob%next%tsum_cpu (jpnij))
549	ALLOCATE(sl_timer_glob%next%tsum_clock(jpnij))
550	ALLOCATE(sl_timer_glob%next%niter (jpnij))
551	sl_timer_glob%next%prev => sl_timer_glob
552	sl_timer_glob%next%next => NULL()
553	sl_timer_glob => sl_timer_glob%next
554	ENDIF
555	s_timer => s_timer%next
556	END DO
557
558	IF( narea == 1 ) THEN
559	! Compute some stats
560	sl_timer_glob => sl_timer_glob_root
561	DO WHILE( ASSOCIATED(sl_timer_glob) )
562	sl_timer_ave%cname = sl_timer_glob%cname(1)
563	sl_timer_ave%tsum_cpu = SUM (sl_timer_glob%tsum_cpu (:)) / jpnij
564	sl_timer_ave%tsum_clock = SUM (sl_timer_glob%tsum_clock(:)) / jpnij
565	sl_timer_ave%tmax_cpu = MAXVAL(sl_timer_glob%tsum_cpu (:))
566	sl_timer_ave%tmax_clock = MAXVAL(sl_timer_glob%tsum_clock(:))
567	sl_timer_ave%tmin_cpu = MINVAL(sl_timer_glob%tsum_cpu (:))
568	sl_timer_ave%tmin_clock = MINVAL(sl_timer_glob%tsum_clock(:))
569	sl_timer_ave%niter = SUM (sl_timer_glob%niter (:))
570	!
571	IF( ASSOCIATED(sl_timer_glob%next) ) THEN
572	ALLOCATE(sl_timer_ave%next)
573	sl_timer_ave%next%prev => sl_timer_ave
574	sl_timer_ave%next%next => NULL()
575	sl_timer_ave => sl_timer_ave%next
576	ENDIF
577	sl_timer_glob => sl_timer_glob%next
578	END DO
579
580	! reorder the averaged list by CPU time
581	s_wrk => NULL()
582	sl_timer_ave => sl_timer_ave_root
583	DO
584	ll_ord = .TRUE.
585	sl_timer_ave => sl_timer_ave_root
586	DO WHILE( ASSOCIATED( sl_timer_ave%next ) )
587
588	IF( .NOT. ASSOCIATED(sl_timer_ave%next) ) EXIT
589
590	IF ( sl_timer_ave%tsum_clock < sl_timer_ave%next%tsum_clock ) THEN
591	ALLOCATE(s_wrk)
592	! Copy data into the new object pointed to by s_wrk
593	s_wrk = sl_timer_ave%next
594	! Insert this new timer object before our current position
595	CALL insert (sl_timer_ave, sl_timer_ave_root, s_wrk)
596	! Remove the old object from the list
597	CALL suppress(sl_timer_ave%next)
598	ll_ord = .FALSE.
599	CYCLE
600	ENDIF
601	IF( ASSOCIATED(sl_timer_ave%next) ) sl_timer_ave => sl_timer_ave%next
602	END DO
603	IF( ll_ord ) EXIT
604	END DO
605
606	! write averaged info
607	WRITE(numtime,"('Averaged timing on all processors :')")
608	WRITE(numtime,"('-----------------------------------')")
609	WRITE(numtime,"('Section',13x,'Elap. Time(s)',2x,'Elap. Time(%)',2x, &
610	& 'CPU Time(s)',2x,'CPU Time(%)',2x,'CPU/Elap',1x, &
611	& 'Max elap(%)',2x,'Min elap(%)',2x, &
612	& 'Freq')")
613	sl_timer_ave => sl_timer_ave_root
614	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)'
615	DO WHILE ( ASSOCIATED(sl_timer_ave) )
616	IF( sl_timer_ave%tsum_clock > 0. ) &
617	WRITE(numtime,TRIM(clfmt)) sl_timer_ave%cname(1:18), &
618	& sl_timer_ave%tsum_clock,sl_timer_ave%tsum_clock100.jpnij/tot_etime, &
619	& sl_timer_ave%tsum_cpu ,sl_timer_ave%tsum_cpu100.jpnij/tot_ctime , &
620	& sl_timer_ave%tsum_cpu/sl_timer_ave%tsum_clock, &
621	& sl_timer_ave%tmax_clock100.jpnij/tot_etime, &
622	& sl_timer_ave%tmin_clock100.jpnij/tot_etime, &
623	& sl_timer_ave%niter/REAL(jpnij)
624	sl_timer_ave => sl_timer_ave%next
625	END DO
626	WRITE(numtime,*)
627	!
628	DEALLOCATE(sl_timer_ave_root)
629	ENDIF
630	!
631	DEALLOCATE(sl_timer_glob_root)
632	!
633	END SUBROUTINE waver_info
634
635
636	SUBROUTINE wmpi_info
637	!!----------------------------------------------------------------------
638	!! * ROUTINE wmpi_time *
639	!! ** Purpose : compute and write a summary of MPI infos
640	!!----------------------------------------------------------------------
641	!
642	INTEGER :: idum, icode
643	INTEGER, ALLOCATABLE, DIMENSION(:) :: iall_rank
644	REAL(wp) :: ztot_ratio
645	REAL(wp) :: zmax_etime, zmax_ctime, zmax_ratio, zmin_etime, zmin_ctime, zmin_ratio
646	REAL(wp) :: zavg_etime, zavg_ctime, zavg_ratio
647	REAL(wp), ALLOCATABLE, DIMENSION(:) :: zall_ratio
648	CHARACTER(LEN=128), dimension(8) :: cllignes
649	CHARACTER(LEN=128) :: clhline, clstart_date, clfinal_date
650	CHARACTER(LEN=2048) :: clfmt
651
652	! Gather all times
653	ALLOCATE( zall_ratio(jpnij), iall_rank(jpnij) )
654	IF( narea == 1 ) THEN
655	iall_rank(:) = (/ (idum,idum=0,jpnij-1) /)
656
657	! Compute elapse user time
658	zavg_etime = tot_etime/REAL(jpnij,wp)
659	zmax_etime = MAXVAL(all_etime(:))
660	zmin_etime = MINVAL(all_etime(:))
661
662	! Compute CPU user time
663	zavg_ctime = tot_ctime/REAL(jpnij,wp)
664	zmax_ctime = MAXVAL(all_ctime(:))
665	zmin_ctime = MINVAL(all_ctime(:))
666
667	! Compute cpu/elapsed ratio
668	zall_ratio(:) = all_ctime(:) / all_etime(:)
669	ztot_ratio = SUM(all_ctime(:))/SUM(all_etime(:))
670	zavg_ratio = SUM(zall_ratio(:))/REAL(jpnij,wp)
671	zmax_ratio = MAXVAL(zall_ratio(:))
672	zmin_ratio = MINVAL(zall_ratio(:))
673
674	! Output Format
675	clhline ='1x,13("-"),"\|",18("-"),"\|",14("-"),"\|",18("-"),/,'
676	cllignes(1)='(1x,"MPI summary report :",/,'
677	cllignes(2)='1x,"--------------------",//,'
678	cllignes(3)='1x,"Process Rank \|"," Elapsed Time (s) \|"," CPU Time (s) \|"," Ratio CPU/Elapsed",/,'
679	cllignes(4)=' (4x,i6,4x,"\|",f12.3,6x,"\|",f12.3,2x,"\|",4x,f7.3,/),'
680	WRITE(cllignes(4)(1:6),'(I6)') jpnij
681	cllignes(5)='1x,"Total \|",f12.3,6x,"\|",F12.3,2x,"\|",4x,f7.3,/,'
682	cllignes(6)='1x,"Minimum \|",f12.3,6x,"\|",F12.3,2x,"\|",4x,f7.3,/,'
683	cllignes(7)='1x,"Maximum \|",f12.3,6x,"\|",F12.3,2x,"\|",4x,f7.3,/,'
684	cllignes(8)='1x,"Average \|",f12.3,6x,"\|",F12.3,2x,"\|",4x,f7.3)'
685	clfmt=TRIM(cllignes(1))// TRIM(cllignes(2))//TRIM(cllignes(3))// &
686	& TRIM(clhline)//TRIM(cllignes(4))//TRIM(clhline)//TRIM(cllignes(5))// &
687	& TRIM(clhline)//TRIM(cllignes(6))//TRIM(clhline)//TRIM(cllignes(7))// &
688	& TRIM(clhline)//TRIM(cllignes(8))
689	WRITE(numtime, TRIM(clfmt)) &
690	(iall_rank(idum),all_etime(idum),all_ctime(idum),zall_ratio(idum),idum=1, jpnij), &
691	tot_etime, tot_ctime, ztot_ratio, &
692	zmin_etime, zmin_ctime, zmin_ratio, &
693	zmax_etime, zmax_ctime, zmax_ratio, &
694	zavg_etime, zavg_ctime, zavg_ratio
695	WRITE(numtime,*)
696	END IF
697	!
698	DEALLOCATE(zall_ratio, iall_rank)
699	!
700	END SUBROUTINE wmpi_info
701	#endif
702
703
704	SUBROUTINE timing_ini_var(cdinfo)
705	!!----------------------------------------------------------------------
706	!! * ROUTINE timing_ini_var *
707	!! ** Purpose : create timing structure
708	!!----------------------------------------------------------------------
709	CHARACTER(len=*), INTENT(in) :: cdinfo
710	LOGICAL :: ll_section
711
712	!
713	IF( .NOT. ASSOCIATED(s_timer_root) ) THEN
714	ALLOCATE(s_timer_root)
715	s_timer_root%cname = cdinfo
716	s_timer_root%t_cpu = 0._wp
717	s_timer_root%t_clock = 0._wp
718	s_timer_root%tsum_cpu = 0._wp
719	s_timer_root%tsum_clock = 0._wp
720	s_timer_root%tmax_cpu = 0._wp
721	s_timer_root%tmax_clock = 0._wp
722	s_timer_root%tmin_cpu = 0._wp
723	s_timer_root%tmin_clock = 0._wp
724	s_timer_root%tsub_cpu = 0._wp
725	s_timer_root%tsub_clock = 0._wp
726	s_timer_root%ncount = 0
727	s_timer_root%ncount_rate = 0
728	s_timer_root%ncount_max = 0
729	s_timer_root%niter = 0
730	s_timer_root%l_tdone = .FALSE.
731	s_timer_root%next => NULL()
732	s_timer_root%prev => NULL()
733	s_timer => s_timer_root
734	!
735	ALLOCATE(s_wrk)
736	s_wrk => NULL()
737	!
738	ALLOCATE(s_timer_old)
739	s_timer_old%cname = cdinfo
740	s_timer_old%t_cpu = 0._wp
741	s_timer_old%t_clock = 0._wp
742	s_timer_old%tsum_cpu = 0._wp
743	s_timer_old%tsum_clock = 0._wp
744	s_timer_old%tmax_cpu = 0._wp
745	s_timer_old%tmax_clock = 0._wp
746	s_timer_old%tmin_cpu = 0._wp
747	s_timer_old%tmin_clock = 0._wp
748	s_timer_old%tsub_cpu = 0._wp
749	s_timer_old%tsub_clock = 0._wp
750	s_timer_old%ncount = 0
751	s_timer_old%ncount_rate = 0
752	s_timer_old%ncount_max = 0
753	s_timer_old%niter = 0
754	s_timer_old%l_tdone = .TRUE.
755	s_timer_old%next => NULL()
756	s_timer_old%prev => NULL()
757
758	ELSE
759	s_timer => s_timer_root
760	! case of already existing area (typically inside a loop)
761	! write(,) 'in ini_var for routine : ', cdinfo
762	DO WHILE( ASSOCIATED(s_timer) )
763	IF( TRIM(s_timer%cname) .EQ. TRIM(cdinfo) ) THEN
764	! write(,) 'in ini_var for routine : ', cdinfo,' we return'
765	RETURN ! cdinfo is already in the chain
766	ENDIF
767	s_timer => s_timer%next
768	END DO
769
770	! end of the chain
771	s_timer => s_timer_root
772	DO WHILE( ASSOCIATED(s_timer%next) )
773	s_timer => s_timer%next
774	END DO
775
776	! write(,) 'after search', s_timer%cname
777	! cdinfo is not part of the chain so we add it with initialisation
778	ALLOCATE(s_timer%next)
779	! write(,) 'after allocation of next'
780
781	s_timer%next%cname = cdinfo
782	s_timer%next%t_cpu = 0._wp
783	s_timer%next%t_clock = 0._wp
784	s_timer%next%tsum_cpu = 0._wp
785	s_timer%next%tsum_clock = 0._wp
786	s_timer%next%tmax_cpu = 0._wp
787	s_timer%next%tmax_clock = 0._wp
788	s_timer%next%tmin_cpu = 0._wp
789	s_timer%next%tmin_clock = 0._wp
790	s_timer%next%tsub_cpu = 0._wp
791	s_timer%next%tsub_clock = 0._wp
792	s_timer%next%ncount = 0
793	s_timer%next%ncount_rate = 0
794	s_timer%next%ncount_max = 0
795	s_timer%next%niter = 0
796	s_timer%next%l_tdone = .FALSE.
797	s_timer%next%parent_section => NULL()
798	s_timer%next%prev => s_timer
799	s_timer%next%next => NULL()
800	s_timer => s_timer%next
801	ENDIF
802	! write(,) 'after allocation'
803	!
804	END SUBROUTINE timing_ini_var
805
806
807	SUBROUTINE timing_reset
808	!!----------------------------------------------------------------------
809	!! * ROUTINE timing_reset *
810	!! ** Purpose : go to root of timing tree
811	!!----------------------------------------------------------------------
812	l_initdone = .TRUE.
813	! IF(lwp) WRITE(numout,*)
814	! IF(lwp) WRITE(numout,*) 'timing_reset : instrumented routines for timing'
815	! IF(lwp) WRITE(numout,*) '~~~~~~~~~~~~'
816	CALL timing_list(s_timer_root)
817	! WRITE(numout,*)
818	!
819	END SUBROUTINE timing_reset
820
821
822	RECURSIVE SUBROUTINE timing_list(ptr)
823
824	TYPE(timer), POINTER, INTENT(inout) :: ptr
825	!
826	IF( ASSOCIATED(ptr%next) ) CALL timing_list(ptr%next)
827	IF(lwp) WRITE(numout,*)' ', ptr%cname
828	!
829	END SUBROUTINE timing_list
830
831
832	SUBROUTINE insert(sd_current, sd_root ,sd_ptr)
833	!!----------------------------------------------------------------------
834	!! * ROUTINE insert *
835	!! ** Purpose : insert an element in timer structure
836	!!----------------------------------------------------------------------
837	TYPE(timer), POINTER, INTENT(inout) :: sd_current, sd_root, sd_ptr
838	!
839
840	IF( ASSOCIATED( sd_current, sd_root ) ) THEN
841	! If our current element is the root element then
842	! replace it with the one being inserted
843	sd_root => sd_ptr
844	ELSE
845	sd_current%prev%next => sd_ptr
846	END IF
847	sd_ptr%next => sd_current
848	sd_ptr%prev => sd_current%prev
849	sd_current%prev => sd_ptr
850	! Nullify the pointer to the new element now that it is held
851	! within the list. If we don't do this then a subsequent call
852	! to ALLOCATE memory to this pointer will fail.
853	sd_ptr => NULL()
854	!
855	END SUBROUTINE insert
856
857
858	SUBROUTINE suppress(sd_ptr)
859	!!----------------------------------------------------------------------
860	!! * ROUTINE suppress *
861	!! ** Purpose : supress an element in timer structure
862	!!----------------------------------------------------------------------
863	TYPE(timer), POINTER, INTENT(inout) :: sd_ptr
864	!
865	TYPE(timer), POINTER :: sl_temp
866
867	sl_temp => sd_ptr
868	sd_ptr => sd_ptr%next
869	IF ( ASSOCIATED(sl_temp%next) ) sl_temp%next%prev => sl_temp%prev
870	DEALLOCATE(sl_temp)
871	sl_temp => NULL()
872	!
873	END SUBROUTINE suppress
874
875	!!=====================================================================
876	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/trunk/src/OCE/timing.F90

Download in other formats: