root / synthbench / skampi / measurements / localio.c
History | View | Annotate | Download (26.2 kB)
1 |
/********************************************************************************
|
---|---|
2 |
* SKaMPI MPI-Benchmark
|
3 |
*
|
4 |
* Copyright (C) 2007-2008 Joachim Mathes
|
5 |
* Lehrstuhl Informatik fuer Ingenieure und Naturwissenschaftler
|
6 |
* Institut fuer Algorithmen und Kognitive Systeme
|
7 |
* Fakultaet fuer Informatik
|
8 |
* Universitaet Karlsruhe (TH)
|
9 |
*
|
10 |
* This program is free software; you can redistribute it and/or
|
11 |
* modify it under the terms of the GNU General Public License
|
12 |
* as published by the Free Software Foundation; either version 2
|
13 |
* of the License, or (at your option) any later version.
|
14 |
*
|
15 |
* This program is distributed in the hope that it will be useful,
|
16 |
* but WITHOUT ANY WARRANTY; without even the implied warranty of
|
17 |
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
18 |
* GNU General Public License for more details.
|
19 |
*
|
20 |
* You should have received a copy of the GNU General Public License along
|
21 |
* with this program; if not, write to the Free Software Foundation, Inc.,
|
22 |
* 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
|
23 |
*******************************************************************************/
|
24 |
|
25 |
/**
|
26 |
* \file localio.c
|
27 |
*
|
28 |
* \brief SKaMPI measurement functions for rudimentary local IO
|
29 |
*
|
30 |
* This file provides rudimentary IO measurement functions for SkaMPI.
|
31 |
*
|
32 |
* \author Joachim Mathes
|
33 |
* \date 2007-2008
|
34 |
*/
|
35 |
|
36 |
#include <mpi.h> |
37 |
#include "../mpiversiontest.h" |
38 |
|
39 |
#ifdef USE_MPI_IO
|
40 |
|
41 |
/**
|
42 |
* The <tt>_GNU_SOURCE</tt> macro must be set to provide the <tt>O_DIRECT</tt>
|
43 |
* parameter for open() routine.
|
44 |
*/
|
45 |
#define _GNU_SOURCE
|
46 |
|
47 |
#include <stdio.h> |
48 |
#include <stdlib.h> |
49 |
#include <string.h> |
50 |
#include <sys/types.h> |
51 |
#include <sys/stat.h> |
52 |
#include <fcntl.h> |
53 |
#include <unistd.h> |
54 |
#include <assert.h> |
55 |
#include <errno.h> |
56 |
#include <sys/mman.h> |
57 |
|
58 |
#include "../misc.h" |
59 |
#include "../synchronize.h" |
60 |
#include "../output.h" |
61 |
#include "../mem.h" |
62 |
|
63 |
#define IO_FILENAME "skampi_io" |
64 |
#define POSIX_API "posix" |
65 |
#define MPI_API "mpi" |
66 |
#define BUFFER_SIZE 4194304 /**< 4 MB */ |
67 |
#define MAXIMUM_POWER 62 /**< limited by MPI_Offset size (number of bits) */ |
68 |
|
69 |
#pragma weak begin_skampi_extensions
|
70 |
|
71 |
/**
|
72 |
* \brief Returns a unique filename in a dedicated path
|
73 |
*
|
74 |
* This function is defined in <tt><SKaMPI path>/measurements/io.c</tt>.
|
75 |
*/
|
76 |
extern char *get_io_filename (char *, int); |
77 |
|
78 |
/********************************************************************************
|
79 |
* Global variables
|
80 |
*******************************************************************************/
|
81 |
|
82 |
static char* io_filename; |
83 |
static MPI_File io_fh;
|
84 |
static int io_fd; |
85 |
|
86 |
/********************************************************************************
|
87 |
* Measure functions
|
88 |
*******************************************************************************/
|
89 |
|
90 |
/*@{*/
|
91 |
/** \name Measurement-Functions
|
92 |
* The following measurement function blocks (<tt>init_</tt>, <tt>measure_</tt>,
|
93 |
* <tt>finalize_</tt>) are meant to work on <em>local</em> disks. Nevertheless
|
94 |
* it depends on the user defined environment variable <tt>TMP</tt>, if a
|
95 |
* distributed RFA-file will be stored on local or global (parallel) disks.
|
96 |
*/
|
97 |
|
98 |
/**
|
99 |
* \brief Initialization function of measure function
|
100 |
* measure_MPI_IO_read_file_once().
|
101 |
*
|
102 |
* Only one process is active. It reads once from a file.
|
103 |
*
|
104 |
* Remark:<br>
|
105 |
* With the <tt>O_DIRECT</tt> flag set, cache effects are minimized, because I/O
|
106 |
* is done directly to/from user space buffers. The operation system's page
|
107 |
* cache is bypassed. Under Linux 2.6 alignment to 512-byte boundaries is
|
108 |
* required for buffer and file offset. Thus the following parameters should be
|
109 |
* set in a SKaMPI input file:
|
110 |
* - <tt>set_send_buffert_alignment (512)</tt>
|
111 |
* - <tt>set_recv_buffert_alignment (512)</tt>
|
112 |
* - <tt>switch_buffer_cycling_off ()</tt><br>
|
113 |
*
|
114 |
* <tt>O_DIRECT</tt> is only relevant if the POSIX-API is used for I/O.
|
115 |
*
|
116 |
* For more information please refer to the <tt>open ()</tt> man pages.
|
117 |
*
|
118 |
* \param[in] size size of memory buffer, i.e. number of <tt>MPI_BYTE</tt>s
|
119 |
* \param[in] api POSIX-API or MPI-API for I/O accesses
|
120 |
* \param[in] directio_flag open file with <tt>O_DIRECT</tt> flag to minimize
|
121 |
* cache effects
|
122 |
*
|
123 |
* \return void
|
124 |
*/
|
125 |
void init_MPI_IO_read_file_once (int size, char *api, int directio_flag) { |
126 |
char *send_buffer;
|
127 |
|
128 |
assert (size > 0);
|
129 |
|
130 |
io_filename = get_io_filename (IO_FILENAME, 0);
|
131 |
|
132 |
if (get_measurement_rank () == 0){ |
133 |
|
134 |
send_buffer = mpi_malloc_chars (get_extent (size, MPI_BYTE)); |
135 |
|
136 |
MPI_File_open (MPI_COMM_SELF, io_filename, |
137 |
MPI_MODE_WRONLY | MPI_MODE_CREATE | MPI_MODE_UNIQUE_OPEN, |
138 |
MPI_INFO_NULL, &io_fh); |
139 |
MPI_File_set_view (io_fh, (MPI_Offset)0,
|
140 |
MPI_BYTE, MPI_BYTE, |
141 |
"native", MPI_INFO_NULL);
|
142 |
MPI_File_write (io_fh, send_buffer, size, MPI_BYTE, MPI_STATUS_IGNORE); |
143 |
MPI_File_close (&io_fh); |
144 |
mpi_free (send_buffer); |
145 |
|
146 |
set_recv_buffer_usage (size); |
147 |
set_reported_message_size (size); |
148 |
} |
149 |
|
150 |
MPI_Barrier (get_measurement_comm ()); |
151 |
|
152 |
/* set synchronization type:
|
153 |
SYNC_BARRIER if all SKaMPI processes run on one physical processor
|
154 |
SYNC_REAL if every SKaMPI process runs on its own physical processor */
|
155 |
set_synchronization (SYNC_REAL); |
156 |
|
157 |
init_synchronization (); |
158 |
} |
159 |
|
160 |
/**
|
161 |
* \brief Measures the time to read once from a file.
|
162 |
*
|
163 |
* Only one process is active. It reads once from a file.
|
164 |
*
|
165 |
* Remark:<br>
|
166 |
* With the <tt>O_DIRECT</tt> flag set, cache effects are minimized, because I/O
|
167 |
* is done directly to/from user space buffers. The operation system's page
|
168 |
* cache is bypassed. Under Linux 2.6 alignment to 512-byte boundaries is
|
169 |
* required for buffer and file offset. Thus the following parameters should be
|
170 |
* set in a SKaMPI input file:
|
171 |
* - <tt>set_send_buffert_alignment (512)</tt>
|
172 |
* - <tt>set_recv_buffert_alignment (512)</tt>
|
173 |
* - <tt>switch_buffer_cycling_off ()</tt><br>
|
174 |
*
|
175 |
* <tt>O_DIRECT</tt> is only relevant if the POSIX-API is used for I/O.
|
176 |
*
|
177 |
* For more information please refer to the <tt>open ()</tt> man pages.
|
178 |
*
|
179 |
* \param[in] size size of memory buffer, i.e. number of <tt>MPI_BYTE</tt>s
|
180 |
* \param[in] api POSIX-API or MPI-API for I/O accesses
|
181 |
* \param[in] directio_flag open file with <tt>O_DIRECT</tt> flag to minimize
|
182 |
* cache effects
|
183 |
*
|
184 |
* \return measured time
|
185 |
*/
|
186 |
double measure_MPI_IO_read_file_once (int size, char *api, int directio_flag){ |
187 |
double start_time = 1.0, end_time = 0.0; |
188 |
int open_flags;
|
189 |
char *error_string;
|
190 |
|
191 |
if (get_measurement_rank () == 0){ |
192 |
if (strcmp (api, POSIX_API) == 0){ |
193 |
|
194 |
if (directio_flag != 0) |
195 |
open_flags = O_RDONLY | O_DIRECT; |
196 |
else
|
197 |
open_flags = O_RDONLY; |
198 |
|
199 |
printf ("flags %d,%d\n", open_flags, O_DIRECT);
|
200 |
|
201 |
errno = 0;
|
202 |
if ((io_fd = open (io_filename, open_flags)) < 0){ |
203 |
error_string = strerror (errno); |
204 |
error_with_abort (errno, |
205 |
"\nmeasure_MPI_IO_read_file_once (int %d, char * %s, int %d) failed."
|
206 |
"\nCannot open local file (read only mode)."
|
207 |
"\nError: %s\n",
|
208 |
size, api, directio_flag, error_string); |
209 |
} |
210 |
|
211 |
start_time = start_synchronization (); |
212 |
read (io_fd, get_recv_buffer (), size); |
213 |
end_time = MPI_Wtime (); |
214 |
|
215 |
close (io_fd); |
216 |
|
217 |
} |
218 |
else{
|
219 |
MPI_File_open (MPI_COMM_SELF, io_filename, MPI_MODE_RDONLY, MPI_INFO_NULL, &io_fh); |
220 |
MPI_File_set_view (io_fh, (MPI_Offset)0,
|
221 |
MPI_BYTE, MPI_BYTE, |
222 |
"native", MPI_INFO_NULL);
|
223 |
|
224 |
start_time = start_synchronization (); |
225 |
MPI_File_read (io_fh, get_recv_buffer (), size, MPI_BYTE, MPI_STATUS_IGNORE); |
226 |
end_time = stop_synchronization (); |
227 |
|
228 |
MPI_File_close (&io_fh); |
229 |
|
230 |
} |
231 |
} |
232 |
else if (get_measurement_rank () != 0) { |
233 |
start_synchronization (); |
234 |
} |
235 |
stop_synchronization (); |
236 |
|
237 |
if (get_measurement_rank () == 0) |
238 |
return end_time - start_time;
|
239 |
else
|
240 |
return -1.0; |
241 |
|
242 |
} |
243 |
|
244 |
/**
|
245 |
* \brief Finalization function of measure function
|
246 |
* measure_MPI_IO_read_file_once().
|
247 |
*
|
248 |
* Only one process is active. It reads once from a file.
|
249 |
*
|
250 |
* Remark:<br>
|
251 |
* With the <tt>O_DIRECT</tt> flag set, cache effects are minimized, because I/O
|
252 |
* is done directly to/from user space buffers. The operation system's page
|
253 |
* cache is bypassed. Under Linux 2.6 alignment to 512-byte boundaries is
|
254 |
* required for buffer and file offset. Thus the following parameters should be
|
255 |
* set in a SKaMPI input file:
|
256 |
* - <tt>set_send_buffert_alignment (512)</tt>
|
257 |
* - <tt>set_recv_buffert_alignment (512)</tt>
|
258 |
* - <tt>switch_buffer_cycling_off ()</tt><br>
|
259 |
*
|
260 |
* <tt>O_DIRECT</tt> is only relevant if the POSIX-API is used for I/O.
|
261 |
*
|
262 |
* For more information please refer to the <tt>open ()</tt> man pages.
|
263 |
*
|
264 |
* \param[in] size size of memory buffer, i.e. number of <tt>MPI_BYTE</tt>s
|
265 |
* \param[in] api POSIX-API or MPI-API for I/O accesses
|
266 |
* \param[in] directio_flag open file with <tt>O_DIRECT</tt> flag to minimize
|
267 |
* cache effects
|
268 |
*
|
269 |
* \return void
|
270 |
*/
|
271 |
void finalize_MPI_IO_read_file_once (int size, char *api, int directio_flag){ |
272 |
MPI_File_delete (io_filename, MPI_INFO_NULL); |
273 |
mpi_free (io_filename); |
274 |
} |
275 |
|
276 |
/**
|
277 |
* \brief Initialization function of measure function
|
278 |
* measure_MPI_IO_write_file_once().
|
279 |
*
|
280 |
* Only one process is active. It writes once to a file.
|
281 |
*
|
282 |
* Remark:<br>
|
283 |
* With the <tt>O_DIRECT</tt> flag set, cache effects are minimized, because I/O
|
284 |
* is done directly to/from user space buffers. The operation system's page
|
285 |
* cache is bypassed. Under Linux 2.6 alignment to 512-byte boundaries is
|
286 |
* required for buffer and file offset. Thus the following parameters should be
|
287 |
* set in a SKaMPI input file:
|
288 |
* - <tt>set_send_buffert_alignment (512)</tt>
|
289 |
* - <tt>set_recv_buffert_alignment (512)</tt>
|
290 |
* - <tt>switch_buffer_cycling_off ()</tt><br>
|
291 |
*
|
292 |
* For more information please refer to the <tt>open ()</tt> man pages.
|
293 |
*
|
294 |
* \param[in] size size of memory buffer, i.e. number of <tt>MPI_BYTE</tt>s
|
295 |
* \param[in] api POSIX-API or MPI-API for I/O accesses
|
296 |
* \param[in] create_flag write into existing file (FALSE) or create it (TRUE)
|
297 |
* \param[in] directio_flag open file with <tt>O_DIRECT</tt> flag to minimize
|
298 |
* cache effects
|
299 |
*
|
300 |
* \return void
|
301 |
*/
|
302 |
void init_MPI_IO_write_file_once (int size, char *api, int create_flag, int directio_flag) { |
303 |
char *send_buffer;
|
304 |
|
305 |
assert (size > 0);
|
306 |
|
307 |
io_filename = get_io_filename (IO_FILENAME, 0);
|
308 |
|
309 |
if (get_measurement_rank () == 0){ |
310 |
|
311 |
if (create_flag == 0){ |
312 |
send_buffer = mpi_malloc_chars (get_extent (size, MPI_BYTE)); |
313 |
|
314 |
MPI_File_open (MPI_COMM_SELF, io_filename, |
315 |
MPI_MODE_WRONLY | MPI_MODE_CREATE | MPI_MODE_UNIQUE_OPEN, |
316 |
MPI_INFO_NULL, &io_fh); |
317 |
MPI_File_set_view (io_fh, (MPI_Offset)0,
|
318 |
MPI_BYTE, MPI_BYTE, |
319 |
"native", MPI_INFO_NULL);
|
320 |
MPI_File_write (io_fh, send_buffer, size, MPI_BYTE, MPI_STATUS_IGNORE); |
321 |
MPI_File_close (&io_fh); |
322 |
mpi_free (send_buffer); |
323 |
} |
324 |
|
325 |
set_send_buffer_usage (size); |
326 |
set_reported_message_size (size); |
327 |
} |
328 |
|
329 |
MPI_Barrier (get_measurement_comm ()); |
330 |
|
331 |
/* set synchronization type:
|
332 |
SYNC_BARRIER if all SKaMPI processes run on one physical processor
|
333 |
SYNC_REAL if every SKaMPI process runs on its own physical processor */
|
334 |
set_synchronization (SYNC_REAL); |
335 |
|
336 |
init_synchronization (); |
337 |
} |
338 |
|
339 |
/**
|
340 |
* \brief Measures the time to write once to a file.
|
341 |
*
|
342 |
* Only one process is active. It writes once to a file.
|
343 |
*
|
344 |
* Remark:<br>
|
345 |
* With the <tt>O_DIRECT</tt> flag set, cache effects are minimized, because I/O
|
346 |
* is done directly to/from user space buffers. The operation system's page
|
347 |
* cache is bypassed. Under Linux 2.6 alignment to 512-byte boundaries is
|
348 |
* required for buffer and file offset. Thus the following parameters should be
|
349 |
* set in a SKaMPI input file:
|
350 |
* - <tt>set_send_buffert_alignment (512)</tt>
|
351 |
* - <tt>set_recv_buffert_alignment (512)</tt>
|
352 |
* - <tt>switch_buffer_cycling_off ()</tt><br>
|
353 |
*
|
354 |
* <tt>O_DIRECT</tt> is only relevant if the POSIX-API is used for I/O.
|
355 |
*
|
356 |
* For more information please refer to the <tt>open ()</tt> man pages.
|
357 |
*
|
358 |
* \param[in] size size of memory buffer, i.e. number of <tt>MPI_BYTE</tt>s
|
359 |
* \param[in] api POSIX-API or MPI-API for I/O accesses
|
360 |
* \param[in] create_flag write into existing file (FALSE) or create it (TRUE)
|
361 |
* \param[in] directio_flag open file with <tt>O_DIRECT</tt> flag to minimize
|
362 |
* cache effects
|
363 |
*
|
364 |
* \return measured time
|
365 |
*/
|
366 |
double measure_MPI_IO_write_file_once (int size, char *api, int create_flag, int directio_flag){ |
367 |
double start_time = 1.0, end_time = 0.0; |
368 |
int open_flags;
|
369 |
char *error_string;
|
370 |
|
371 |
if (get_measurement_rank () == 0){ |
372 |
if (strcmp (api, POSIX_API) == 0){ |
373 |
|
374 |
if (directio_flag != 0) |
375 |
open_flags = O_WRONLY | O_DIRECT; |
376 |
else
|
377 |
open_flags = O_WRONLY; |
378 |
|
379 |
errno = 0;
|
380 |
|
381 |
if (create_flag == 0){ /* open existing file */ |
382 |
|
383 |
if ((io_fd = open (io_filename, open_flags)) < 0){ |
384 |
error_string = strerror (errno); |
385 |
error_with_abort (errno, |
386 |
"\nmeasure_MPI_IO_write_file_once (int %d, char * %s, int %d, int %d) failed."
|
387 |
"\nCannot open local file (write only mode)."
|
388 |
"\nError: %s\n",
|
389 |
size, api, create_flag, directio_flag, error_string); |
390 |
} |
391 |
} |
392 |
else { /* open nonexisting file and create it */ |
393 |
|
394 |
if ((io_fd = open (io_filename, open_flags|O_CREAT, S_IRUSR|S_IWUSR|S_IRGRP|S_IROTH)) < 0){ |
395 |
error_string = strerror (errno); |
396 |
error_with_abort (errno, |
397 |
"\nmeasure_MPI_IO_write_file_once (int %d, char * %s, int %d, int %d) failed."
|
398 |
"\nCannot open local file (write only mode)."
|
399 |
"\nError: %s\n",
|
400 |
size, api, create_flag, directio_flag, error_string); |
401 |
} |
402 |
} |
403 |
|
404 |
start_time = start_synchronization (); |
405 |
write (io_fd, get_send_buffer (), size); |
406 |
fsync (io_fd); |
407 |
end_time = MPI_Wtime (); |
408 |
|
409 |
close (io_fd); |
410 |
|
411 |
} |
412 |
else{ /* if strcmp (api, POSIX_API) != 0 */ |
413 |
|
414 |
if (create_flag == 0){ |
415 |
|
416 |
MPI_File_open (MPI_COMM_SELF, io_filename, MPI_MODE_WRONLY, MPI_INFO_NULL, &io_fh); |
417 |
} |
418 |
else{ /* if create_flag != 0*/ |
419 |
|
420 |
MPI_File_open (MPI_COMM_SELF, io_filename, MPI_MODE_WRONLY|MPI_MODE_CREATE, MPI_INFO_NULL, &io_fh); |
421 |
} |
422 |
|
423 |
MPI_File_set_view (io_fh, (MPI_Offset)0,
|
424 |
MPI_BYTE, MPI_BYTE, |
425 |
"native", MPI_INFO_NULL);
|
426 |
|
427 |
start_time = start_synchronization (); |
428 |
MPI_File_write (io_fh, get_send_buffer (), size, MPI_BYTE, MPI_STATUS_IGNORE); |
429 |
MPI_File_sync (io_fh); |
430 |
end_time = MPI_Wtime (); |
431 |
|
432 |
MPI_File_close (&io_fh); |
433 |
} |
434 |
} |
435 |
else if (get_measurement_rank () != 0) { |
436 |
start_synchronization (); |
437 |
} |
438 |
stop_synchronization (); |
439 |
|
440 |
if (get_measurement_rank () == 0) |
441 |
return end_time - start_time;
|
442 |
else
|
443 |
return -1.0; |
444 |
} |
445 |
|
446 |
/**
|
447 |
* \brief Finalization function of measure function
|
448 |
* measure_MPI_IO_write_file_once().
|
449 |
*
|
450 |
* Only one process is active. It writes once to a file.
|
451 |
*
|
452 |
* Remark:<br>
|
453 |
* With the <tt>O_DIRECT</tt> flag set, cache effects are minimized, because I/O
|
454 |
* is done directly to/from user space buffers. The operation system's page
|
455 |
* cache is bypassed. Under Linux 2.6 alignment to 512-byte boundaries is
|
456 |
* required for buffer and file offset. Thus the following parameters should be
|
457 |
* set in a SKaMPI input file:
|
458 |
* - <tt>set_send_buffert_alignment (512)</tt>
|
459 |
* - <tt>set_recv_buffert_alignment (512)</tt>
|
460 |
* - <tt>switch_buffer_cycling_off ()</tt><br>
|
461 |
*
|
462 |
* <tt>O_DIRECT</tt> is only relevant if the POSIX-API is used for I/O.
|
463 |
*
|
464 |
* For more information please refer to the <tt>open ()</tt> man pages.
|
465 |
*
|
466 |
* \param[in] size size of memory buffer, i.e. number of <tt>MPI_BYTE</tt>s
|
467 |
* \param[in] api POSIX-API or MPI-API for I/O accesses
|
468 |
* \param[in] create_flag write into existing file (FALSE) or create it (TRUE)
|
469 |
* \param[in] directio_flag open file with <tt>O_DIRECT</tt> flag to minimize
|
470 |
* cache effects
|
471 |
*
|
472 |
* \return void
|
473 |
*/
|
474 |
void finalize_MPI_IO_write_file_once (int size, char *api, int create_flag, int directio_flag){ |
475 |
MPI_File_delete (io_filename, MPI_INFO_NULL); |
476 |
mpi_free (io_filename); |
477 |
} |
478 |
|
479 |
/**
|
480 |
* \brief Initialization function of measure function
|
481 |
* measure_MPI_IO_read_large_file_once().
|
482 |
*
|
483 |
* Only one process is active. It writes once to a file.
|
484 |
*
|
485 |
* Since SKaMPI measurement functions are not allowed to use MPI_Offset
|
486 |
* parameters, it is impossible to tell an init_-routine to create a file
|
487 |
* which is larger than \f$2^{\mbox{\texttt{sizeof(int)}}-1}-1\f$ directly. As
|
488 |
* a preliminary solution we use a parameter (<tt>power</tt>) which commits the
|
489 |
* power to 2 as an indicator for the file size.
|
490 |
*
|
491 |
* Remark concerning the <em>HP XC6000</em>:<br>
|
492 |
* Measurements showed that there is no significant difference between MPI-API
|
493 |
* and POSIX-API I/O accesses, if files are larger than 1MB. Thus there is no
|
494 |
* choice between these two modes like in measure_MPI_IO_read_file_once(),
|
495 |
* which makes type compatibilty problems much easier. Only MPI-API is
|
496 |
* supported.
|
497 |
*
|
498 |
* Only one process is active. It reads once from a file.
|
499 |
*
|
500 |
* \param[in] power size of memory buffer; 2 to the power of `power' <tt>MPI_BYTE</tt>s
|
501 |
*
|
502 |
* \return void
|
503 |
*/
|
504 |
void init_MPI_IO_read_large_file_once (int power) { |
505 |
MPI_Offset size; |
506 |
char *error_string;
|
507 |
|
508 |
io_filename = get_io_filename (IO_FILENAME, 0);
|
509 |
|
510 |
if (get_measurement_rank () == 0){ |
511 |
|
512 |
if (power > MAXIMUM_POWER || power < 0){ |
513 |
error_string = strerror (EINVAL); |
514 |
error_with_abort (errno, |
515 |
"\ninit_MPI_IO_read_large_file_once (int %d) failed."
|
516 |
"\nInvalid power argument."
|
517 |
"\nError: %s\n",
|
518 |
power, error_string); |
519 |
} |
520 |
|
521 |
size = ((MPI_Offset) 1) << power;
|
522 |
|
523 |
MPI_File_open (MPI_COMM_SELF, io_filename, |
524 |
MPI_MODE_WRONLY | MPI_MODE_CREATE | MPI_MODE_UNIQUE_OPEN, |
525 |
MPI_INFO_NULL, &io_fh); |
526 |
MPI_File_preallocate (io_fh, size); |
527 |
MPI_File_close (&io_fh); |
528 |
|
529 |
set_recv_buffer_usage (size); |
530 |
set_reported_message_size (size); |
531 |
} |
532 |
|
533 |
MPI_Barrier (get_measurement_comm ()); |
534 |
|
535 |
/* set synchronization type:
|
536 |
SYNC_BARRIER if all SKaMPI processes run on one physical processor
|
537 |
SYNC_REAL if every SKaMPI process runs on its own physical processor */
|
538 |
set_synchronization (SYNC_REAL); |
539 |
|
540 |
init_synchronization (); |
541 |
} |
542 |
|
543 |
/**
|
544 |
* \brief Measures the time to read once from a large file.
|
545 |
*
|
546 |
* Only one process is active. It writes once to a file.
|
547 |
*
|
548 |
* Since SKaMPI measurement functions are not allowed to use MPI_Offset
|
549 |
* parameters, it is impossible to tell an init_-routine to create a file
|
550 |
* which is larger than \f$2^{\mbox{\texttt{sizeof(int)}}-1}-1\f$ directly. As
|
551 |
* a preliminary solution we use a parameter (<tt>power</tt>) which commits the
|
552 |
* power to 2 as an indicator for the file size.
|
553 |
*
|
554 |
* Remark concerning the <em>HP XC6000</em>:<br>
|
555 |
* Measurements showed that there is no significant difference between MPI-API
|
556 |
* and POSIX-API I/O accesses, if files are larger than 1MB. Thus there is no
|
557 |
* choice between these two modes like in measure_MPI_IO_read_file_once(),
|
558 |
* which makes type compatibilty problems much easier. Only MPI-API is
|
559 |
* supported.
|
560 |
*
|
561 |
* \param[in] power size of memory buffer; 2 to the power of `power' <tt>MPI_BYTE</tt>s
|
562 |
*
|
563 |
* \return measured time
|
564 |
*/
|
565 |
double measure_MPI_IO_read_large_file_once (int power){ |
566 |
MPI_Offset size; |
567 |
double start_time = 1.0, end_time = 0.0; |
568 |
|
569 |
if (get_measurement_rank () == 0){ |
570 |
|
571 |
size = ((MPI_Offset) 1) << power;
|
572 |
|
573 |
MPI_File_open (MPI_COMM_SELF, io_filename, MPI_MODE_RDONLY, MPI_INFO_NULL, &io_fh); |
574 |
MPI_File_set_view (io_fh, (MPI_Offset)0,
|
575 |
MPI_BYTE, MPI_BYTE, |
576 |
"native", MPI_INFO_NULL);
|
577 |
|
578 |
start_time = start_synchronization (); |
579 |
MPI_File_read (io_fh, get_recv_buffer (), size, MPI_BYTE, MPI_STATUS_IGNORE); |
580 |
end_time = stop_synchronization (); |
581 |
|
582 |
MPI_File_close (&io_fh); |
583 |
} |
584 |
else if (get_measurement_rank () != 0) { |
585 |
start_synchronization (); |
586 |
} |
587 |
stop_synchronization (); |
588 |
|
589 |
if (get_measurement_rank () == 0) |
590 |
return end_time - start_time;
|
591 |
else
|
592 |
return -1.0; |
593 |
|
594 |
} |
595 |
|
596 |
/**
|
597 |
* \brief Finalization function of measure function
|
598 |
* measure_MPI_IO_read_large_file_once().
|
599 |
*
|
600 |
* Only one process is active. It writes once to a file.
|
601 |
*
|
602 |
* Since SKaMPI measurement functions are not allowed to use MPI_Offset
|
603 |
* parameters, it is impossible to tell an init_-routine to create a file
|
604 |
* which is larger than \f$2^{\mbox{\texttt{sizeof(int)}}-1}-1\f$ directly. As
|
605 |
* a preliminary solution we use a parameter (<tt>power</tt>) which commits the
|
606 |
* power to 2 as an indicator for the file size.
|
607 |
*
|
608 |
* Remark concerning the <em>HP XC6000</em>:<br>
|
609 |
* Measurements showed that there is no significant difference between MPI-API
|
610 |
* and POSIX-API I/O accesses, if files are larger than 1MB. Thus there is no
|
611 |
* choice between these two modes like in measure_MPI_IO_read_file_once(),
|
612 |
* which makes type compatibilty problems much easier. Only MPI-API is
|
613 |
* supported.
|
614 |
*
|
615 |
* \param[in] power size of memory buffer; 2 to the power of `power' <tt>MPI_BYTE</tt>s
|
616 |
*
|
617 |
* \return void
|
618 |
*/
|
619 |
void finalize_MPI_IO_read_large_file_once (int power){ |
620 |
MPI_File_delete (io_filename, MPI_INFO_NULL); |
621 |
mpi_free (io_filename); |
622 |
} |
623 |
|
624 |
/**
|
625 |
* \brief Initialization function of measure function
|
626 |
* measure_MPI_IO_write_large_file_once().
|
627 |
*
|
628 |
* Only one process is active. It writes once to a file.
|
629 |
*
|
630 |
* Since SKaMPI measurement functions are not allowed to use MPI_Offset
|
631 |
* parameters, it is impossible to tell an init_-routine to create a file
|
632 |
* which is larger than \f$2^{\mbox{\texttt{sizeof(int)}}-1}-1\f$ directly. As
|
633 |
* a preliminary solution we use a parameter (<tt>power</tt>) which commits the
|
634 |
* power to 2 as an indicator for the file size.
|
635 |
*
|
636 |
* Remark concerning the <em>HP XC6000</em>:<br>
|
637 |
* Measurements showed that there is no significant difference between MPI-API
|
638 |
* and POSIX-API I/O accesses, if files are larger than 1MB. Thus there is no
|
639 |
* choice between these two modes like in measure_MPI_IO_read_file_once(),
|
640 |
* which makes type compatibilty problems much easier. Only MPI-API is
|
641 |
* supported.
|
642 |
*
|
643 |
* \param[in] power size of memory buffer; 2 to the power of `power' <tt>MPI_BYTE</tt>s
|
644 |
* \param[in] create_flag write into existing file (FALSE) or create it (TRUE)
|
645 |
*
|
646 |
* \return void
|
647 |
*/
|
648 |
void init_MPI_IO_write_large_file_once (int power, int create_flag) { |
649 |
MPI_Offset size; |
650 |
char *error_string;
|
651 |
|
652 |
io_filename = get_io_filename (IO_FILENAME, 0);
|
653 |
|
654 |
if (get_measurement_rank () == 0){ |
655 |
|
656 |
if (power > MAXIMUM_POWER || power < 0){ |
657 |
error_string = strerror (EINVAL); |
658 |
error_with_abort (errno, |
659 |
"\ninit_MPI_IO_write_large_file_once (int %d, int %d) failed."
|
660 |
"\nInvalid power argument."
|
661 |
"\nError: %s\n",
|
662 |
power, create_flag, error_string); |
663 |
} |
664 |
|
665 |
size = ((MPI_Offset) 1) << power;
|
666 |
|
667 |
if (create_flag == 0){ |
668 |
|
669 |
MPI_File_open (MPI_COMM_SELF, io_filename, |
670 |
MPI_MODE_WRONLY | MPI_MODE_CREATE | MPI_MODE_UNIQUE_OPEN, |
671 |
MPI_INFO_NULL, &io_fh); |
672 |
MPI_File_preallocate (io_fh, size); |
673 |
MPI_File_close (&io_fh); |
674 |
} |
675 |
|
676 |
set_send_buffer_usage (size); |
677 |
set_reported_message_size (size); |
678 |
} |
679 |
|
680 |
MPI_Barrier (get_measurement_comm ()); |
681 |
|
682 |
/* set synchronization type:
|
683 |
SYNC_BARRIER if all SKaMPI processes run on one physical processor
|
684 |
SYNC_REAL if every SKaMPI process runs on its own physical processor */
|
685 |
set_synchronization (SYNC_REAL); |
686 |
|
687 |
init_synchronization (); |
688 |
} |
689 |
|
690 |
/**
|
691 |
* \brief Measures the time to write once to a large file.
|
692 |
*
|
693 |
* Only one process is active. It writes once to a file.
|
694 |
*
|
695 |
* Since SKaMPI measurement functions are not allowed to use MPI_Offset
|
696 |
* parameters, it is impossible to tell an init_-routine to create a file
|
697 |
* which is larger than \f$2^{\mbox{\texttt{sizeof(int)}}-1}-1\f$ directly. As
|
698 |
* a preliminary solution we use a parameter (<tt>power</tt>) which commits the
|
699 |
* power to 2 as an indicator for the file size.
|
700 |
*
|
701 |
* Remark concerning the <em>HP XC6000</em>:<br>
|
702 |
* Measurements showed that there is no significant difference between MPI-API
|
703 |
* and POSIX-API I/O accesses, if files are larger than 1MB. Thus there is no
|
704 |
* choice between these two modes like in measure_MPI_IO_read_file_once(),
|
705 |
* which makes type compatibilty problems much easier. Only MPI-API is
|
706 |
* supported.
|
707 |
*
|
708 |
* \param[in] power size of memory buffer; 2 to the power of `power' <tt>MPI_BYTE</tt>s
|
709 |
* \param[in] create_flag write into existing file (FALSE) or create it (TRUE)
|
710 |
*
|
711 |
* \return measured time
|
712 |
*/
|
713 |
double measure_MPI_IO_write_large_file_once (int power, int create_flag){ |
714 |
MPI_Offset size; |
715 |
double start_time = 1.0, end_time = 0.0; |
716 |
|
717 |
if (get_measurement_rank () == 0){ |
718 |
|
719 |
size = ((MPI_Offset) 1) << power;
|
720 |
|
721 |
if (create_flag == 0){ |
722 |
|
723 |
MPI_File_open (MPI_COMM_SELF, io_filename, MPI_MODE_WRONLY, MPI_INFO_NULL, &io_fh); |
724 |
} |
725 |
else{ /* if create_flag != 0*/ |
726 |
|
727 |
MPI_File_open (MPI_COMM_SELF, io_filename, MPI_MODE_WRONLY|MPI_MODE_CREATE, MPI_INFO_NULL, &io_fh); |
728 |
} |
729 |
|
730 |
MPI_File_set_view (io_fh, (MPI_Offset)0,
|
731 |
MPI_BYTE, MPI_BYTE, |
732 |
"native", MPI_INFO_NULL);
|
733 |
|
734 |
start_time = start_synchronization (); |
735 |
MPI_File_write (io_fh, get_send_buffer (), size, MPI_BYTE, MPI_STATUS_IGNORE); |
736 |
end_time = MPI_Wtime (); |
737 |
|
738 |
MPI_File_close (&io_fh); |
739 |
|
740 |
} |
741 |
else if (get_measurement_rank () != 0) { |
742 |
start_synchronization (); |
743 |
} |
744 |
stop_synchronization (); |
745 |
|
746 |
if (get_measurement_rank () == 0) |
747 |
return end_time - start_time;
|
748 |
else
|
749 |
return -1.0; |
750 |
} |
751 |
|
752 |
/**
|
753 |
* \brief Finalization function of measure function
|
754 |
* measure_MPI_IO_write_large_file_once().
|
755 |
*
|
756 |
* Only one process is active. It writes once to a file.
|
757 |
*
|
758 |
* Since SKaMPI measurement functions are not allowed to use MPI_Offset
|
759 |
* parameters, it is impossible to tell an init_-routine to create a file
|
760 |
* which is larger than \f$2^{\mbox{\texttt{sizeof(int)}}-1}-1\f$ directly. As
|
761 |
* a preliminary solution we use a parameter (<tt>power</tt>) which commits the
|
762 |
* power to 2 as an indicator for the file size.
|
763 |
*
|
764 |
* Remark concerning the <em>HP XC6000</em>:<br>
|
765 |
* Measurements showed that there is no significant difference between MPI-API
|
766 |
* and POSIX-API I/O accesses, if files are larger than 1MB. Thus there is no
|
767 |
* choice between these two modes like in measure_MPI_IO_read_file_once(),
|
768 |
* which makes type compatibilty problems much easier. Only MPI-API is
|
769 |
* supported.
|
770 |
*
|
771 |
* \param[in] power size of memory buffer; 2 to the power of `power' <tt>MPI_BYTE</tt>s
|
772 |
* \param[in] create_flag write into existing file (FALSE) or create it (TRUE)
|
773 |
*
|
774 |
* \return void
|
775 |
*/
|
776 |
void finalize_MPI_IO_write_large_file_once (int power, int create_flag){ |
777 |
MPI_File_delete (io_filename, MPI_INFO_NULL); |
778 |
mpi_free (io_filename); |
779 |
} |
780 |
|
781 |
/*@}*/
|
782 |
|
783 |
#pragma weak end_skampi_extensions
|
784 |
|
785 |
#endif /* USE_MPI_IO */ |