1 | SUBROUTINE cpu_statistics |
---|
2 | |
---|
3 | !------------------------------------------------------------------------------! |
---|
4 | ! Current revisions: |
---|
5 | ! ----------------- |
---|
6 | ! |
---|
7 | ! |
---|
8 | ! Former revisions: |
---|
9 | ! ----------------- |
---|
10 | ! $Id: cpu_statistics.f90 623 2010-12-10 08:52:17Z suehring $ |
---|
11 | ! |
---|
12 | ! 622 2010-12-10 08:08:13Z raasch |
---|
13 | ! output of handling of collective operations |
---|
14 | ! |
---|
15 | ! 222 2009-01-12 16:04:16Z letzel |
---|
16 | ! Bugfix for nonparallel execution |
---|
17 | ! |
---|
18 | ! 197 2008-09-16 15:29:03Z raasch |
---|
19 | ! Format adjustments in order to allow CPU# > 999, |
---|
20 | ! data are collected from PE0 in an ordered sequence which seems to avoid |
---|
21 | ! hanging of processes on SGI-ICE |
---|
22 | ! |
---|
23 | ! 82 2007-04-16 15:40:52Z raasch |
---|
24 | ! Preprocessor directives for old systems removed |
---|
25 | ! |
---|
26 | ! RCS Log replace by Id keyword, revision history cleaned up |
---|
27 | ! |
---|
28 | ! Revision 1.13 2006/04/26 12:10:51 raasch |
---|
29 | ! Output of number of threads per task, max = min in case of 1 PE |
---|
30 | ! |
---|
31 | ! Revision 1.1 1997/07/24 11:11:11 raasch |
---|
32 | ! Initial revision |
---|
33 | ! |
---|
34 | ! |
---|
35 | ! Description: |
---|
36 | ! ------------ |
---|
37 | ! Analysis and output of the cpu-times measured. All PE results are collected |
---|
38 | ! on PE0 in order to calculate the mean cpu-time over all PEs and other |
---|
39 | ! statistics. The output is sorted according to the amount of cpu-time consumed |
---|
40 | ! and output on PE0. |
---|
41 | !------------------------------------------------------------------------------! |
---|
42 | |
---|
43 | USE cpulog |
---|
44 | USE pegrid |
---|
45 | USE control_parameters |
---|
46 | |
---|
47 | IMPLICIT NONE |
---|
48 | |
---|
49 | INTEGER :: i, ii(1), iii, lp, sender |
---|
50 | REAL, SAVE :: norm = 1.0 |
---|
51 | REAL, DIMENSION(:), ALLOCATABLE :: pe_max, pe_min, pe_rms, sum |
---|
52 | REAL, DIMENSION(:,:), ALLOCATABLE :: pe_log_points |
---|
53 | |
---|
54 | |
---|
55 | ! |
---|
56 | !-- Compute cpu-times in seconds |
---|
57 | log_point%mtime = log_point%mtime / norm |
---|
58 | log_point%sum = log_point%sum / norm |
---|
59 | log_point%vector = log_point%vector / norm |
---|
60 | WHERE ( log_point%counts /= 0 ) |
---|
61 | log_point%mean = log_point%sum / log_point%counts |
---|
62 | END WHERE |
---|
63 | |
---|
64 | |
---|
65 | ! |
---|
66 | !-- Collect cpu-times from all PEs and calculate statistics |
---|
67 | IF ( myid == 0 ) THEN |
---|
68 | ! |
---|
69 | !-- Allocate and initialize temporary arrays needed for statistics |
---|
70 | ALLOCATE( pe_max( SIZE( log_point ) ), pe_min( SIZE( log_point ) ), & |
---|
71 | pe_rms( SIZE( log_point ) ), & |
---|
72 | pe_log_points( SIZE( log_point ), 0:numprocs-1 ) ) |
---|
73 | pe_min = log_point%sum |
---|
74 | pe_max = log_point%sum ! need to be set in case of 1 PE |
---|
75 | pe_rms = 0.0 |
---|
76 | |
---|
77 | #if defined( __parallel ) |
---|
78 | ! |
---|
79 | !-- Receive data from all PEs |
---|
80 | DO i = 1, numprocs-1 |
---|
81 | CALL MPI_RECV( pe_max(1), SIZE( log_point ), MPI_REAL, & |
---|
82 | i, i, comm2d, status, ierr ) |
---|
83 | sender = status(MPI_SOURCE) |
---|
84 | pe_log_points(:,sender) = pe_max |
---|
85 | ENDDO |
---|
86 | pe_log_points(:,0) = log_point%sum ! Results from PE0 |
---|
87 | ! |
---|
88 | !-- Calculate mean of all PEs, store it on log_point%sum |
---|
89 | !-- and find minimum and maximum |
---|
90 | DO iii = 1, SIZE( log_point ) |
---|
91 | DO i = 1, numprocs-1 |
---|
92 | log_point(iii)%sum = log_point(iii)%sum + pe_log_points(iii,i) |
---|
93 | pe_min(iii) = MIN( pe_min(iii), pe_log_points(iii,i) ) |
---|
94 | pe_max(iii) = MAX( pe_max(iii), pe_log_points(iii,i) ) |
---|
95 | ENDDO |
---|
96 | log_point(iii)%sum = log_point(iii)%sum / numprocs |
---|
97 | ! |
---|
98 | !-- Calculate rms |
---|
99 | DO i = 0, numprocs-1 |
---|
100 | pe_rms(iii) = pe_rms(iii) + ( & |
---|
101 | pe_log_points(iii,i) - log_point(iii)%sum & |
---|
102 | )**2 |
---|
103 | ENDDO |
---|
104 | pe_rms(iii) = SQRT( pe_rms(iii) / numprocs ) |
---|
105 | ENDDO |
---|
106 | ELSE |
---|
107 | ! |
---|
108 | !-- Send data to PE0 (pe_max is used as temporary storage to send |
---|
109 | !-- the data in order to avoid sending the data type log) |
---|
110 | ALLOCATE( pe_max( SIZE( log_point ) ) ) |
---|
111 | pe_max = log_point%sum |
---|
112 | CALL MPI_SEND( pe_max(1), SIZE( log_point ), MPI_REAL, 0, myid, comm2d, & |
---|
113 | ierr ) |
---|
114 | #endif |
---|
115 | |
---|
116 | ENDIF |
---|
117 | |
---|
118 | ! |
---|
119 | !-- Write cpu-times |
---|
120 | IF ( myid == 0 ) THEN |
---|
121 | ! |
---|
122 | !-- Re-store sums |
---|
123 | ALLOCATE( sum( SIZE( log_point ) ) ) |
---|
124 | WHERE ( log_point%counts /= 0 ) |
---|
125 | sum = log_point%sum |
---|
126 | ELSEWHERE |
---|
127 | sum = -1.0 |
---|
128 | ENDWHERE |
---|
129 | |
---|
130 | ! |
---|
131 | !-- Write cpu-times sorted by size |
---|
132 | CALL check_open( 18 ) |
---|
133 | #if defined( __parallel ) |
---|
134 | WRITE ( 18, 100 ) TRIM( run_description_header ), & |
---|
135 | numprocs * threads_per_task, pdims(1), pdims(2), & |
---|
136 | threads_per_task |
---|
137 | #else |
---|
138 | WRITE ( 18, 100 ) TRIM( run_description_header ), & |
---|
139 | numprocs * threads_per_task, 1, 1, & |
---|
140 | threads_per_task |
---|
141 | #endif |
---|
142 | DO |
---|
143 | ii = MAXLOC( sum ) |
---|
144 | i = ii(1) |
---|
145 | IF ( sum(i) /= -1.0 ) THEN |
---|
146 | WRITE ( 18, 102 ) & |
---|
147 | log_point(i)%place, log_point(i)%sum, & |
---|
148 | log_point(i)%sum / log_point(1)%sum * 100.0, & |
---|
149 | log_point(i)%counts, pe_min(i), pe_max(i), pe_rms(i) |
---|
150 | sum(i) = -1.0 |
---|
151 | ELSE |
---|
152 | EXIT |
---|
153 | ENDIF |
---|
154 | ENDDO |
---|
155 | ENDIF |
---|
156 | |
---|
157 | |
---|
158 | ! |
---|
159 | !-- The same procedure again for the individual measurements. |
---|
160 | ! |
---|
161 | !-- Compute cpu-times in seconds |
---|
162 | log_point_s%mtime = log_point_s%mtime / norm |
---|
163 | log_point_s%sum = log_point_s%sum / norm |
---|
164 | log_point_s%vector = log_point_s%vector / norm |
---|
165 | WHERE ( log_point_s%counts /= 0 ) |
---|
166 | log_point_s%mean = log_point_s%sum / log_point_s%counts |
---|
167 | END WHERE |
---|
168 | |
---|
169 | ! |
---|
170 | !-- Collect cpu-times from all PEs and calculate statistics |
---|
171 | #if defined( __parallel ) |
---|
172 | ! |
---|
173 | !-- Set barrier in order to avoid that PE0 receives log_point_s-data |
---|
174 | !-- while still busy with receiving log_point-data (see above) |
---|
175 | CALL MPI_BARRIER( comm2d, ierr ) |
---|
176 | #endif |
---|
177 | IF ( myid == 0 ) THEN |
---|
178 | ! |
---|
179 | !-- Initialize temporary arrays needed for statistics |
---|
180 | pe_min = log_point_s%sum |
---|
181 | pe_max = log_point_s%sum ! need to be set in case of 1 PE |
---|
182 | pe_rms = 0.0 |
---|
183 | |
---|
184 | #if defined( __parallel ) |
---|
185 | ! |
---|
186 | !-- Receive data from all PEs |
---|
187 | DO i = 1, numprocs-1 |
---|
188 | CALL MPI_RECV( pe_max(1), SIZE( log_point ), MPI_REAL, & |
---|
189 | MPI_ANY_SOURCE, MPI_ANY_TAG, comm2d, status, ierr ) |
---|
190 | sender = status(MPI_SOURCE) |
---|
191 | pe_log_points(:,sender) = pe_max |
---|
192 | ENDDO |
---|
193 | pe_log_points(:,0) = log_point_s%sum ! Results from PE0 |
---|
194 | ! |
---|
195 | !-- Calculate mean of all PEs, store it on log_point_s%sum |
---|
196 | !-- and find minimum and maximum |
---|
197 | DO iii = 1, SIZE( log_point ) |
---|
198 | DO i = 1, numprocs-1 |
---|
199 | log_point_s(iii)%sum = log_point_s(iii)%sum + pe_log_points(iii,i) |
---|
200 | pe_min(iii) = MIN( pe_min(iii), pe_log_points(iii,i) ) |
---|
201 | pe_max(iii) = MAX( pe_max(iii), pe_log_points(iii,i) ) |
---|
202 | ENDDO |
---|
203 | log_point_s(iii)%sum = log_point_s(iii)%sum / numprocs |
---|
204 | ! |
---|
205 | !-- Calculate rms |
---|
206 | DO i = 0, numprocs-1 |
---|
207 | pe_rms(iii) = pe_rms(iii) + ( & |
---|
208 | pe_log_points(iii,i) - log_point_s(iii)%sum & |
---|
209 | )**2 |
---|
210 | ENDDO |
---|
211 | pe_rms(iii) = SQRT( pe_rms(iii) / numprocs ) |
---|
212 | ENDDO |
---|
213 | ELSE |
---|
214 | ! |
---|
215 | !-- Send data to PE0 (pe_max is used as temporary storage to send |
---|
216 | !-- the data in order to avoid sending the data type log) |
---|
217 | pe_max = log_point_s%sum |
---|
218 | CALL MPI_SEND( pe_max(1), SIZE( log_point ), MPI_REAL, 0, 0, comm2d, & |
---|
219 | ierr ) |
---|
220 | #endif |
---|
221 | |
---|
222 | ENDIF |
---|
223 | |
---|
224 | ! |
---|
225 | !-- Write cpu-times |
---|
226 | IF ( myid == 0 ) THEN |
---|
227 | ! |
---|
228 | !-- Re-store sums |
---|
229 | WHERE ( log_point_s%counts /= 0 ) |
---|
230 | sum = log_point_s%sum |
---|
231 | ELSEWHERE |
---|
232 | sum = -1.0 |
---|
233 | ENDWHERE |
---|
234 | |
---|
235 | ! |
---|
236 | !-- Write cpu-times sorted by size |
---|
237 | WRITE ( 18, 101 ) |
---|
238 | DO |
---|
239 | ii = MAXLOC( sum ) |
---|
240 | i = ii(1) |
---|
241 | IF ( sum(i) /= -1.0 ) THEN |
---|
242 | WRITE ( 18, 102 ) & |
---|
243 | log_point_s(i)%place, log_point_s(i)%sum, & |
---|
244 | log_point_s(i)%sum / log_point(1)%sum * 100.0, & |
---|
245 | log_point_s(i)%counts, pe_min(i), pe_max(i), pe_rms(i) |
---|
246 | sum(i) = -1.0 |
---|
247 | ELSE |
---|
248 | EXIT |
---|
249 | ENDIF |
---|
250 | ENDDO |
---|
251 | |
---|
252 | ! |
---|
253 | !-- Output handling of collective operations |
---|
254 | IF ( collective_wait ) THEN |
---|
255 | WRITE ( 18, 103 ) |
---|
256 | ELSE |
---|
257 | WRITE ( 18, 104 ) |
---|
258 | ENDIF |
---|
259 | |
---|
260 | ! |
---|
261 | !-- Empty lines in order to create a gap to the results of the model |
---|
262 | !-- continuation runs |
---|
263 | WRITE ( 18, 105 ) |
---|
264 | |
---|
265 | ! |
---|
266 | !-- Unit 18 is not needed anymore |
---|
267 | CALL close_file( 18 ) |
---|
268 | |
---|
269 | ENDIF |
---|
270 | |
---|
271 | |
---|
272 | 100 FORMAT (A/11('-')//'CPU measures for ',I5,' PEs (',I5,'(x) * ',I5,'(y', & |
---|
273 | &') tasks *',I5,' threads):'/ & |
---|
274 | &'----------------------------------------------------------', & |
---|
275 | &'------------'//& |
---|
276 | &'place: mean counts min ', & |
---|
277 | &' max rms'/ & |
---|
278 | &' sec. % sec. ', & |
---|
279 | &' sec. sec.'/ & |
---|
280 | &'-----------------------------------------------------------', & |
---|
281 | &'-------------------') |
---|
282 | |
---|
283 | 101 FORMAT (/'special measures:'/ & |
---|
284 | &'-----------------------------------------------------------', & |
---|
285 | &'--------------------') |
---|
286 | |
---|
287 | 102 FORMAT (A20,2X,F9.3,2X,F7.2,1X,I7,3(1X,F9.3)) |
---|
288 | 103 FORMAT (/'Barriers are set in front of collective operations') |
---|
289 | 104 FORMAT (/'No barriers are set in front of collective operations') |
---|
290 | 105 FORMAT (//) |
---|
291 | |
---|
292 | END SUBROUTINE cpu_statistics |
---|
293 | |
---|