src/gromacs/timing/walltime_accounting.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2013, The GROMACS development team.
   5  * Copyright (c) 2013,2014,2015,2016,2017 by the GROMACS development team.
   6  * Copyright (c) 2018,2019,2020, by the GROMACS development team, led by
   7  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   8  * and including many others, as listed in the AUTHORS file in the
   9  * top-level source directory and at http://www.gromacs.org.
  10  *
  11  * GROMACS is free software; you can redistribute it and/or
  12  * modify it under the terms of the GNU Lesser General Public License
  13  * as published by the Free Software Foundation; either version 2.1
  14  * of the License, or (at your option) any later version.
  15  *
  16  * GROMACS is distributed in the hope that it will be useful,
  17  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  19  * Lesser General Public License for more details.
  20  *
  21  * You should have received a copy of the GNU Lesser General Public
  22  * License along with GROMACS; if not, see
  23  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  24  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  25  *
  26  * If you want to redistribute modifications to GROMACS, please
  27  * consider that scientific software is very special. Version
  28  * control is crucial - bugs must be traceable. We will be happy to
  29  * consider code for inclusion in the official distribution, but
  30  * derived work must not be called official GROMACS. Details are found
  31  * in the README & COPYING files - if they are missing, get the
  32  * official version at http://www.gromacs.org.
  33  *
  34  * To help us fund GROMACS development, we humbly ask that you cite
  35  * the research papers on the package. Check out http://www.gromacs.org.
  36  */
  37 #include "gmxpre.h"
  38
  39 #include "walltime_accounting.h"
  40
  41 #include "config.h"
  42
  43 #include <ctime>
  44
  45 #ifdef HAVE_UNISTD_H
  46 #    include <unistd.h>
  47 #endif
  48 #ifdef HAVE_SYS_TIME_H
  49 #    include <sys/time.h>
  50 #endif
  51
  52 #include "gromacs/utility/basedefinitions.h"
  53 #include "gromacs/utility/smalloc.h"
  54
  55 /* TODO in future: convert gmx_walltime_accounting to a class,
  56  * resolve who should have responsibility for recording the number of
  57  * steps done, consider whether parts of finish_time, print_perf,
  58  * wallcycle_print belong in this module.
  59  *
  60  * If/when any kind of task parallelism is implemented (even OpenMP
  61  * regions simultaneously assigned to different tasks), consider
  62  * whether this data structure (and/or cycle counters) should be
  63  * maintained on a per-OpenMP-thread basis.
  64  *
  65  * Consider also replacing this with std::chrono. */
  66
  67 /*! \brief Manages caching wall-clock time measurements for
  68  * simulations */
  69 typedef struct gmx_walltime_accounting
  70 {
  71     //! Seconds since the epoch recorded at the start of the simulation
  72     double start_time_stamp;
  73     /*! \brief Seconds since the epoch recorded at the reset of
  74      * counters for the simulation (or the start, if no reset has
  75      * occured). */
  76     double reset_time_stamp;
  77     /*! \brief Seconds since the epoch recorded at the reset of
  78      * counters for the simulation for this thread (or the start, if
  79      * no reset has occured). */
  80     double reset_time_stamp_per_thread;
  81     //! Total seconds elapsed over the simulation since counter reset
  82     double elapsed_time;
  83     //! Total seconds elapsed over the simulation since counter reset running this thread
  84     double elapsed_time_over_all_threads;
  85     /*! \brief Number of OpenMP threads that will be launched by this
  86      * MPI rank.
  87      *
  88      * This is used to scale elapsed_time_over_all_threads so
  89      * that any combination of real MPI, thread MPI and OpenMP (even
  90      * mdrun -ntomp_pme) processes/threads would (when run at maximum
  91      * efficiency) return values such that the sum of
  92      * elapsed_time_over_all_threads over all threads was constant
  93      * with respect to parallelism implementation. */
  94     int numOpenMPThreads;
  95     //! Numbers of steps done before reset of counters
  96     int64_t nsteps_done_at_reset;
  97     //! Set by integrators to report the amount of work they did
  98     int64_t nsteps_done;
  99     //! Whether the simulation has finished in a way valid for walltime reporting.
 100     bool isValidFinish;
 101 } t_gmx_walltime_accounting;
 102
 103 /*! \brief Calls system timing routines (e.g. clock_gettime) to get
 104  * the (fractional) number of seconds elapsed since the epoch when
 105  * this thread was executing.
 106  *
 107  * This can be used to measure system load. This can be unreliable if
 108  * threads migrate between sockets. If thread-specific timers are not
 109  * supported by the OS (e.g. if the OS is not POSIX-compliant), this
 110  * function is implemented by gmx_gettime. */
 111 static double gmx_gettime_per_thread();
 112
 113 // TODO In principle, all this should get protected by checks that
 114 // walltime_accounting is not nullptr. In practice, that nullptr condition
 115 // does not happen, and future refactoring will likely enforce it by
 116 // having the gmx_walltime_accounting_t object be owned by the runner
 117 // object. When these become member functions, existence will be
 118 // guaranteed.
 119
 120 gmx_walltime_accounting_t walltime_accounting_init(int numOpenMPThreads)
 121 {
 122     gmx_walltime_accounting_t walltime_accounting;
 123
 124     snew(walltime_accounting, 1);
 125     walltime_accounting->start_time_stamp            = 0;
 126     walltime_accounting->reset_time_stamp            = 0;
 127     walltime_accounting->reset_time_stamp_per_thread = 0;
 128     walltime_accounting->elapsed_time                = 0;
 129     walltime_accounting->nsteps_done_at_reset        = 0;
 130     walltime_accounting->nsteps_done                 = 0;
 131     walltime_accounting->numOpenMPThreads            = numOpenMPThreads;
 132     walltime_accounting->isValidFinish               = false;
 133
 134     return walltime_accounting;
 135 }
 136
 137 void walltime_accounting_destroy(gmx_walltime_accounting_t walltime_accounting)
 138 {
 139     sfree(walltime_accounting);
 140 }
 141
 142 void walltime_accounting_reset_time(gmx_walltime_accounting_t walltime_accounting, int64_t step)
 143 {
 144     walltime_accounting->reset_time_stamp            = gmx_gettime();
 145     walltime_accounting->reset_time_stamp_per_thread = gmx_gettime_per_thread();
 146     walltime_accounting->elapsed_time                = 0;
 147     walltime_accounting->nsteps_done                 = 0;
 148     walltime_accounting->nsteps_done_at_reset        = step;
 149 }
 150
 151 void walltime_accounting_start_time(gmx_walltime_accounting_t walltime_accounting)
 152 {
 153     walltime_accounting_reset_time(walltime_accounting, 0);
 154     walltime_accounting->start_time_stamp = walltime_accounting->reset_time_stamp;
 155 }
 156
 157 void walltime_accounting_end_time(gmx_walltime_accounting_t walltime_accounting)
 158 {
 159     double now, now_per_thread;
 160
 161     now            = gmx_gettime();
 162     now_per_thread = gmx_gettime_per_thread();
 163
 164     walltime_accounting->elapsed_time = now - walltime_accounting->reset_time_stamp;
 165     walltime_accounting->elapsed_time_over_all_threads =
 166             now_per_thread - walltime_accounting->reset_time_stamp_per_thread;
 167     /* For thread-MPI, the per-thread CPU timer makes this just
 168      * work. For OpenMP threads, the per-thread CPU timer measurement
 169      * needs to be multiplied by the number of OpenMP threads used,
 170      * under the current assumption that all regions ever opened
 171      * within a process are of the same size, and each thread should
 172      * keep one core busy.
 173      */
 174     walltime_accounting->elapsed_time_over_all_threads *= walltime_accounting->numOpenMPThreads;
 175 }
 176
 177 double walltime_accounting_get_time_since_start(gmx_walltime_accounting_t walltime_accounting)
 178 {
 179     return gmx_gettime() - walltime_accounting->start_time_stamp;
 180 }
 181
 182 double walltime_accounting_get_time_since_reset(gmx_walltime_accounting_t walltime_accounting)
 183 {
 184     return gmx_gettime() - walltime_accounting->reset_time_stamp;
 185 }
 186
 187 double walltime_accounting_get_time_since_reset_over_all_threads(gmx_walltime_accounting_t walltime_accounting)
 188 {
 189     return walltime_accounting->elapsed_time_over_all_threads;
 190 }
 191
 192 double walltime_accounting_get_start_time_stamp(gmx_walltime_accounting_t walltime_accounting)
 193 {
 194     return walltime_accounting->start_time_stamp;
 195 }
 196
 197 int64_t walltime_accounting_get_nsteps_done_since_reset(gmx_walltime_accounting_t walltime_accounting)
 198 {
 199     return walltime_accounting->nsteps_done - walltime_accounting->nsteps_done_at_reset;
 200 }
 201
 202 void walltime_accounting_set_nsteps_done(gmx_walltime_accounting_t walltime_accounting, int64_t nsteps_done)
 203 {
 204     walltime_accounting->nsteps_done = nsteps_done;
 205 }
 206
 207 double gmx_gettime()
 208 {
 209     /* Use clock_gettime only if we know linking the C run-time
 210        library will work (which is not trivial on e.g. Crays), and its
 211        headers claim sufficient support for POSIX (ie not Mac and
 212        Windows). */
 213 #if HAVE_CLOCK_GETTIME && defined(_POSIX_TIMERS) && _POSIX_TIMERS > 0
 214     struct timespec t;
 215     double          seconds;
 216
 217     clock_gettime(CLOCK_REALTIME, &t);
 218     seconds = static_cast<double>(t.tv_sec) + 1e-9 * t.tv_nsec;
 219
 220     return seconds;
 221 #elif HAVE_GETTIMEOFDAY
 222     // Note that gettimeofday() is deprecated by POSIX, but since Mac
 223     // and Windows do not yet support POSIX, we are still stuck.
 224     struct timeval t;
 225     double         seconds;
 226
 227     gettimeofday(&t, nullptr);
 228     seconds = static_cast<double>(t.tv_sec) + 1e-6 * t.tv_usec;
 229
 230     return seconds;
 231 #else
 232     double seconds;
 233
 234     seconds = time(nullptr);
 235
 236     return seconds;
 237 #endif
 238 }
 239
 240 void walltime_accounting_set_valid_finish(gmx_walltime_accounting_t walltime_accounting)
 241 {
 242     walltime_accounting->isValidFinish = true;
 243 }
 244
 245 //! Return whether the simulation finished in a way valid for reporting walltime.
 246 bool walltime_accounting_get_valid_finish(const gmx_walltime_accounting* walltime_accounting)
 247 {
 248     return walltime_accounting->isValidFinish;
 249 }
 250
 251 static double gmx_gettime_per_thread()
 252 {
 253     /* Use clock_gettime only if we know linking the C run-time
 254        library will work (which is not trivial on e.g. Crays), and its
 255        headers claim sufficient support for POSIX (ie not Mac and
 256        Windows). */
 257 #if HAVE_CLOCK_GETTIME && defined(_POSIX_THREAD_CPUTIME) && _POSIX_THREAD_CPUTIME > 0
 258     struct timespec t;
 259     double          seconds;
 260
 261     clock_gettime(CLOCK_THREAD_CPUTIME_ID, &t);
 262     seconds = static_cast<double>(t.tv_sec) + 1e-9 * t.tv_nsec;
 263
 264     return seconds;
 265 #else
 266     return gmx_gettime();
 267 #endif
 268 }