2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2009,2010,2011,2012,2013,2014,2015, by the GROMACS development team, led by
5 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
6 * and including many others, as listed in the AUTHORS file in the
7 * top-level source directory and at http://www.gromacs.org.
9 * GROMACS is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public License
11 * as published by the Free Software Foundation; either version 2.1
12 * of the License, or (at your option) any later version.
14 * GROMACS is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with GROMACS; if not, see
21 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
22 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
24 * If you want to redistribute modifications to GROMACS, please
25 * consider that scientific software is very special. Version
26 * control is crucial - bugs must be traceable. We will be happy to
27 * consider code for inclusion in the official distribution, but
28 * derived work must not be called official GROMACS. Details are found
29 * in the README & COPYING files - if they are missing, get the
30 * official version at http://www.gromacs.org.
32 * To help us fund GROMACS development, we humbly ask that you cite
33 * the research papers on the package. Check out http://www.gromacs.org.
46 #ifdef HAVE_SYS_TIME_H
50 #include "gromacs/commandline/pargs.h"
51 #include "gromacs/fileio/tpxio.h"
52 #include "gromacs/gmxana/gmx_ana.h"
53 #include "gromacs/legacyheaders/calcgrid.h"
54 #include "gromacs/legacyheaders/checkpoint.h"
55 #include "gromacs/legacyheaders/inputrec.h"
56 #include "gromacs/legacyheaders/macros.h"
57 #include "gromacs/legacyheaders/names.h"
58 #include "gromacs/legacyheaders/perf_est.h"
59 #include "gromacs/legacyheaders/readinp.h"
60 #include "gromacs/legacyheaders/typedefs.h"
61 #include "gromacs/legacyheaders/types/commrec.h"
62 #include "gromacs/math/utilities.h"
63 #include "gromacs/math/vec.h"
64 #include "gromacs/timing/walltime_accounting.h"
65 #include "gromacs/utility/baseversion.h"
66 #include "gromacs/utility/cstringutil.h"
67 #include "gromacs/utility/fatalerror.h"
68 #include "gromacs/utility/futil.h"
69 #include "gromacs/utility/gmxassert.h"
70 #include "gromacs/utility/smalloc.h"
74 /* Enum for situations that can occur during log file parsing, the
75 * corresponding string entries can be found in do_the_tests() in
76 * const char* ParseLog[] */
77 /* TODO clean up CamelCasing of these enum names */
83 eParselogResetProblem,
87 eParselogLargePrimeFactor,
88 eParselogMismatchOfNumberOfPPRanksAndAvailableGPUs,
97 int nPMEnodes; /* number of PME-only nodes used in this test */
98 int nx, ny, nz; /* DD grid */
99 int guessPME; /* if nPMEnodes == -1, this is the guessed number of PME nodes */
100 double *Gcycles; /* This can contain more than one value if doing multiple tests */
104 float *PME_f_load; /* PME mesh/force load average*/
105 float PME_f_load_Av; /* Average average ;) ... */
106 char *mdrun_cmd_line; /* Mdrun command line used for this test */
112 int nr_inputfiles; /* The number of tpr and mdp input files */
113 gmx_int64_t orig_sim_steps; /* Number of steps to be done in the real simulation */
114 gmx_int64_t orig_init_step; /* Init step for the real simulation */
115 real *rcoulomb; /* The coulomb radii [0...nr_inputfiles] */
116 real *rvdw; /* The vdW radii */
117 real *rlist; /* Neighbourlist cutoff radius */
119 int *nkx, *nky, *nkz;
120 real *fsx, *fsy, *fsz; /* Fourierspacing in x,y,z dimension */
124 static void sep_line(FILE *fp)
126 fprintf(fp, "\n------------------------------------------------------------\n");
130 /* Wrapper for system calls */
131 static int gmx_system_call(char *command)
133 return ( system(command) );
137 /* Check if string starts with substring */
138 static gmx_bool str_starts(const char *string, const char *substring)
140 return ( std::strncmp(string, substring, std::strlen(substring)) == 0);
144 static void cleandata(t_perf *perfdata, int test_nr)
146 perfdata->Gcycles[test_nr] = 0.0;
147 perfdata->ns_per_day[test_nr] = 0.0;
148 perfdata->PME_f_load[test_nr] = 0.0;
154 static void remove_if_exists(const char *fn)
158 fprintf(stdout, "Deleting %s\n", fn);
164 static void finalize(const char *fn_out)
170 fp = fopen(fn_out, "r");
171 fprintf(stdout, "\n\n");
173 while (fgets(buf, STRLEN-1, fp) != NULL)
175 fprintf(stdout, "%s", buf);
178 fprintf(stdout, "\n\n");
183 eFoundNothing, eFoundDDStr, eFoundAccountingStr, eFoundCycleStr
186 static int parse_logfile(const char *logfile, const char *errfile,
187 t_perf *perfdata, int test_nr, int presteps, gmx_int64_t cpt_steps,
191 char line[STRLEN], dumstring[STRLEN], dumstring2[STRLEN];
192 const char matchstrdd[] = "Domain decomposition grid";
193 const char matchstrcr[] = "resetting all time and cycle counters";
194 const char matchstrbal[] = "Average PME mesh/force load:";
195 const char matchstring[] = "R E A L C Y C L E A N D T I M E A C C O U N T I N G";
196 const char errSIG[] = "signal, stopping at the next";
198 float dum1, dum2, dum3, dum4;
201 gmx_int64_t resetsteps = -1;
202 gmx_bool bFoundResetStr = FALSE;
203 gmx_bool bResetChecked = FALSE;
206 if (!gmx_fexist(logfile))
208 fprintf(stderr, "WARNING: Could not find logfile %s.\n", logfile);
209 cleandata(perfdata, test_nr);
210 return eParselogNotFound;
213 fp = fopen(logfile, "r");
214 perfdata->PME_f_load[test_nr] = -1.0;
215 perfdata->guessPME = -1;
217 iFound = eFoundNothing;
220 iFound = eFoundDDStr; /* Skip some case statements */
223 while (fgets(line, STRLEN, fp) != NULL)
225 /* Remove leading spaces */
228 /* Check for TERM and INT signals from user: */
229 if (std::strstr(line, errSIG) != NULL)
232 cleandata(perfdata, test_nr);
233 return eParselogTerm;
236 /* Check whether cycle resetting worked */
237 if (presteps > 0 && !bFoundResetStr)
239 if (std::strstr(line, matchstrcr) != NULL)
241 sprintf(dumstring, "step %s", "%" GMX_SCNd64);
242 sscanf(line, dumstring, &resetsteps);
243 bFoundResetStr = TRUE;
244 if (resetsteps == presteps+cpt_steps)
246 bResetChecked = TRUE;
250 sprintf(dumstring, "%" GMX_PRId64, resetsteps);
251 sprintf(dumstring2, "%" GMX_PRId64, presteps+cpt_steps);
252 fprintf(stderr, "WARNING: Time step counters were reset at step %s,\n"
253 " though they were supposed to be reset at step %s!\n",
254 dumstring, dumstring2);
259 /* Look for strings that appear in a certain order in the log file: */
263 /* Look for domain decomp grid and separate PME nodes: */
264 if (str_starts(line, matchstrdd))
266 sscanf(line, "Domain decomposition grid %d x %d x %d, separate PME ranks %d",
267 &(perfdata->nx), &(perfdata->ny), &(perfdata->nz), &npme);
268 if (perfdata->nPMEnodes == -1)
270 perfdata->guessPME = npme;
272 else if (perfdata->nPMEnodes != npme)
274 gmx_fatal(FARGS, "PME ranks from command line and output file are not identical");
276 iFound = eFoundDDStr;
278 /* Catch a few errors that might have occured: */
279 else if (str_starts(line, "There is no domain decomposition for"))
282 return eParselogNoDDGrid;
284 else if (str_starts(line, "The number of ranks you selected"))
287 return eParselogLargePrimeFactor;
289 else if (str_starts(line, "reading tpx file"))
292 return eParselogTPXVersion;
294 else if (str_starts(line, "The -dd or -npme option request a parallel simulation"))
297 return eParselogNotParallel;
301 /* Even after the "Domain decomposition grid" string was found,
302 * it could be that mdrun had to quit due to some error. */
303 if (str_starts(line, "Incorrect launch configuration: mismatching number of"))
306 return eParselogMismatchOfNumberOfPPRanksAndAvailableGPUs;
308 else if (str_starts(line, "Some of the requested GPUs do not exist"))
311 return eParselogGpuProblem;
313 /* Look for PME mesh/force balance (not necessarily present, though) */
314 else if (str_starts(line, matchstrbal))
316 sscanf(&line[std::strlen(matchstrbal)], "%f", &(perfdata->PME_f_load[test_nr]));
318 /* Look for matchstring */
319 else if (str_starts(line, matchstring))
321 iFound = eFoundAccountingStr;
324 case eFoundAccountingStr:
325 /* Already found matchstring - look for cycle data */
326 if (str_starts(line, "Total "))
328 sscanf(line, "Total %*f %lf", &(perfdata->Gcycles[test_nr]));
329 iFound = eFoundCycleStr;
333 /* Already found cycle data - look for remaining performance info and return */
334 if (str_starts(line, "Performance:"))
336 ndum = sscanf(line, "%s %f %f %f %f", dumstring, &dum1, &dum2, &dum3, &dum4);
337 /* (ns/day) is the second last entry, depending on whether GMX_DETAILED_PERF_STATS was set in print_perf(), nrnb.c */
338 perfdata->ns_per_day[test_nr] = (ndum == 5) ? dum3 : dum1;
340 if (bResetChecked || presteps == 0)
346 return eParselogResetProblem;
353 /* Close the log file */
356 /* Check why there is no performance data in the log file.
357 * Did a fatal errors occur? */
358 if (gmx_fexist(errfile))
360 fp = fopen(errfile, "r");
361 while (fgets(line, STRLEN, fp) != NULL)
363 if (str_starts(line, "Fatal error:") )
365 if (fgets(line, STRLEN, fp) != NULL)
367 fprintf(stderr, "\nWARNING: An error occured during this benchmark:\n"
371 cleandata(perfdata, test_nr);
372 return eParselogFatal;
379 fprintf(stderr, "WARNING: Could not find stderr file %s.\n", errfile);
382 /* Giving up ... we could not find out why there is no performance data in
384 fprintf(stdout, "No performance data in log file.\n");
385 cleandata(perfdata, test_nr);
387 return eParselogNoPerfData;
391 static gmx_bool analyze_data(
400 int *index_tpr, /* OUT: Nr of mdp file with best settings */
401 int *npme_optimal) /* OUT: Optimal number of PME nodes */
404 int line = 0, line_win = -1;
405 int k_win = -1, i_win = -1, winPME;
406 double s = 0.0; /* standard deviation */
409 char str_PME_f_load[13];
410 gmx_bool bCanUseOrigTPR;
411 gmx_bool bRefinedCoul, bRefinedVdW, bRefinedGrid;
417 fprintf(fp, "Summary of successful runs:\n");
418 fprintf(fp, "Line tpr PME ranks Gcycles Av. Std.dev. ns/day PME/f");
421 fprintf(fp, " DD grid");
427 for (k = 0; k < ntprs; k++)
429 for (i = 0; i < ntests; i++)
431 /* Select the right dataset: */
432 pd = &(perfdata[k][i]);
434 pd->Gcycles_Av = 0.0;
435 pd->PME_f_load_Av = 0.0;
436 pd->ns_per_day_Av = 0.0;
438 if (pd->nPMEnodes == -1)
440 sprintf(strbuf, "(%3d)", pd->guessPME);
444 sprintf(strbuf, " ");
447 /* Get the average run time of a setting */
448 for (j = 0; j < nrepeats; j++)
450 pd->Gcycles_Av += pd->Gcycles[j];
451 pd->PME_f_load_Av += pd->PME_f_load[j];
453 pd->Gcycles_Av /= nrepeats;
454 pd->PME_f_load_Av /= nrepeats;
456 for (j = 0; j < nrepeats; j++)
458 if (pd->ns_per_day[j] > 0.0)
460 pd->ns_per_day_Av += pd->ns_per_day[j];
464 /* Somehow the performance number was not aquired for this run,
465 * therefor set the average to some negative value: */
466 pd->ns_per_day_Av = -1.0f*nrepeats;
470 pd->ns_per_day_Av /= nrepeats;
473 if (pd->PME_f_load_Av > 0.0)
475 sprintf(str_PME_f_load, "%12.3f", pd->PME_f_load_Av);
479 sprintf(str_PME_f_load, "%s", " - ");
483 /* We assume we had a successful run if both averages are positive */
484 if (pd->Gcycles_Av > 0.0 && pd->ns_per_day_Av > 0.0)
486 /* Output statistics if repeats were done */
489 /* Calculate the standard deviation */
491 for (j = 0; j < nrepeats; j++)
493 s += std::pow( pd->Gcycles[j] - pd->Gcycles_Av, 2 );
498 fprintf(fp, "%4d %3d %4d%s %12.3f %12.3f %12.3f %s",
499 line, k, pd->nPMEnodes, strbuf, pd->Gcycles_Av, s,
500 pd->ns_per_day_Av, str_PME_f_load);
503 fprintf(fp, " %3d %3d %3d", pd->nx, pd->ny, pd->nz);
507 /* Store the index of the best run found so far in 'winner': */
508 if ( (k_win == -1) || (pd->Gcycles_Av < perfdata[k_win][i_win].Gcycles_Av) )
521 gmx_fatal(FARGS, "None of the runs was successful! Check %s for problems.", fn);
526 winPME = perfdata[k_win][i_win].nPMEnodes;
530 /* We stuck to a fixed number of PME-only nodes */
531 sprintf(strbuf, "settings No. %d", k_win);
535 /* We have optimized the number of PME-only nodes */
538 sprintf(strbuf, "%s", "the automatic number of PME ranks");
542 sprintf(strbuf, "%d PME ranks", winPME);
545 fprintf(fp, "Best performance was achieved with %s", strbuf);
546 if ((nrepeats > 1) && (ntests > 1))
548 fprintf(fp, " (see line %d)", line_win);
552 /* Only mention settings if they were modified: */
553 bRefinedCoul = !gmx_within_tol(info->rcoulomb[k_win], info->rcoulomb[0], GMX_REAL_EPS);
554 bRefinedVdW = !gmx_within_tol(info->rvdw[k_win], info->rvdw[0], GMX_REAL_EPS);
555 bRefinedGrid = !(info->nkx[k_win] == info->nkx[0] &&
556 info->nky[k_win] == info->nky[0] &&
557 info->nkz[k_win] == info->nkz[0]);
559 if (bRefinedCoul || bRefinedVdW || bRefinedGrid)
561 fprintf(fp, "Optimized PME settings:\n");
562 bCanUseOrigTPR = FALSE;
566 bCanUseOrigTPR = TRUE;
571 fprintf(fp, " New Coulomb radius: %f nm (was %f nm)\n", info->rcoulomb[k_win], info->rcoulomb[0]);
576 fprintf(fp, " New Van der Waals radius: %f nm (was %f nm)\n", info->rvdw[k_win], info->rvdw[0]);
581 fprintf(fp, " New Fourier grid xyz: %d %d %d (was %d %d %d)\n", info->nkx[k_win], info->nky[k_win], info->nkz[k_win],
582 info->nkx[0], info->nky[0], info->nkz[0]);
585 if (bCanUseOrigTPR && ntprs > 1)
587 fprintf(fp, "and original PME settings.\n");
592 /* Return the index of the mdp file that showed the highest performance
593 * and the optimal number of PME nodes */
595 *npme_optimal = winPME;
597 return bCanUseOrigTPR;
601 /* Get the commands we need to set up the runs from environment variables */
602 static void get_program_paths(gmx_bool bThreads, char *cmd_mpirun[], char *cmd_mdrun[])
605 const char def_mpirun[] = "mpirun";
606 const char def_mdrun[] = "mdrun";
608 const char empty_mpirun[] = "";
610 /* Get the commands we need to set up the runs from environment variables */
613 if ( (cp = getenv("MPIRUN")) != NULL)
615 *cmd_mpirun = gmx_strdup(cp);
619 *cmd_mpirun = gmx_strdup(def_mpirun);
624 *cmd_mpirun = gmx_strdup(empty_mpirun);
627 if ( (cp = getenv("MDRUN" )) != NULL)
629 *cmd_mdrun = gmx_strdup(cp);
633 *cmd_mdrun = gmx_strdup(def_mdrun);
637 /* Check that the commands will run mdrun (perhaps via mpirun) by
638 * running a very quick test simulation. Requires MPI environment or
639 * GPU support to be available if applicable. */
640 /* TODO implement feature to parse the log file to get the list of
641 compatible GPUs from mdrun, if the user of gmx tune-pme has not
643 static void check_mdrun_works(gmx_bool bThreads,
644 const char *cmd_mpirun,
646 const char *cmd_mdrun,
647 gmx_bool bNeedGpuSupport)
649 char *command = NULL;
653 const char filename[] = "benchtest.log";
655 /* This string should always be identical to the one in copyrite.c,
656 * gmx_print_version_info() in the defined(GMX_MPI) section */
657 const char match_mpi[] = "MPI library: MPI";
658 const char match_mdrun[] = "Executable: ";
659 const char match_gpu[] = "GPU support: enabled";
660 gmx_bool bMdrun = FALSE;
661 gmx_bool bMPI = FALSE;
662 gmx_bool bHaveGpuSupport = FALSE;
664 /* Run a small test to see whether mpirun + mdrun work */
665 fprintf(stdout, "Making sure that mdrun can be executed. ");
668 snew(command, std::strlen(cmd_mdrun) + std::strlen(cmd_np) + std::strlen(filename) + 50);
669 sprintf(command, "%s%s-version -maxh 0.001 1> %s 2>&1", cmd_mdrun, cmd_np, filename);
673 snew(command, std::strlen(cmd_mpirun) + std::strlen(cmd_np) + std::strlen(cmd_mdrun) + std::strlen(filename) + 50);
674 sprintf(command, "%s%s%s -version -maxh 0.001 1> %s 2>&1", cmd_mpirun, cmd_np, cmd_mdrun, filename);
676 fprintf(stdout, "Trying '%s' ... ", command);
677 make_backup(filename);
678 gmx_system_call(command);
680 /* Check if we find the characteristic string in the output: */
681 if (!gmx_fexist(filename))
683 gmx_fatal(FARGS, "Output from test run could not be found.");
686 fp = fopen(filename, "r");
687 /* We need to scan the whole output file, since sometimes the queuing system
688 * also writes stuff to stdout/err */
691 cp = fgets(line, STRLEN, fp);
694 if (str_starts(line, match_mdrun) )
698 if (str_starts(line, match_mpi) )
702 if (str_starts(line, match_gpu) )
704 bHaveGpuSupport = TRUE;
714 gmx_fatal(FARGS, "Need a threaded version of mdrun. This one\n"
716 "seems to have been compiled with MPI instead.",
724 gmx_fatal(FARGS, "Need an MPI-enabled version of mdrun. This one\n"
726 "seems to have been compiled without MPI support.",
733 gmx_fatal(FARGS, "Cannot execute mdrun. Please check %s for problems!",
737 if (bNeedGpuSupport && !bHaveGpuSupport)
739 gmx_fatal(FARGS, "The mdrun executable did not have the expected GPU support.");
742 fprintf(stdout, "passed.\n");
749 /*! \brief Helper struct so we can parse the string with eligible GPU
750 IDs outside do_the_tests. */
751 typedef struct eligible_gpu_ids
753 int n; /**< Length of ids */
754 int *ids; /**< Array of length n. NULL if no GPUs in use */
755 } t_eligible_gpu_ids;
757 /* Handles the no-GPU case by emitting an empty string. */
758 static char *make_gpu_id_command_line(int numRanks, int numPmeRanks, const t_eligible_gpu_ids *gpu_ids)
760 char *command_line, *ptr;
761 const char *flag = "-gpu_id ";
764 /* Reserve enough room for the option name, enough single-digit
765 GPU ids (since that is currently all that is possible to use
766 with mdrun), and a terminating NULL. */
767 flag_length = std::strlen(flag);
768 snew(command_line, flag_length + numRanks + 1);
771 /* If the user has given no eligible GPU IDs, or we're trying the
772 * default behaviour, then there is nothing for g_tune_pme to give
773 * to mdrun -gpu_id */
774 if (gpu_ids->n > 0 && numPmeRanks > -1)
776 int numPpRanks, max_num_ranks_for_each_GPU;
779 /* Write the option flag */
780 std::strcpy(ptr, flag);
783 numPpRanks = numRanks - numPmeRanks;
784 max_num_ranks_for_each_GPU = numPpRanks / gpu_ids->n;
785 if (max_num_ranks_for_each_GPU * gpu_ids->n != numPpRanks)
787 /* Some GPUs will receive more work than others, which
788 * we choose to be those with the lowest indices */
789 max_num_ranks_for_each_GPU++;
792 /* Loop over all eligible GPU ids */
793 for (gpu_id = 0, rank = 0; gpu_id < gpu_ids->n; gpu_id++)
795 int rank_for_this_GPU;
796 /* Loop over all PP ranks for GPU with ID gpu_id, building the
797 assignment string. */
798 for (rank_for_this_GPU = 0;
799 rank_for_this_GPU < max_num_ranks_for_each_GPU && rank < numPpRanks;
800 rank++, rank_for_this_GPU++)
802 *ptr = '0' + gpu_ids->ids[gpu_id];
812 static void launch_simulation(
813 gmx_bool bLaunch, /* Should the simulation be launched? */
814 FILE *fp, /* General log file */
815 gmx_bool bThreads, /* whether to use threads */
816 char *cmd_mpirun, /* Command for mpirun */
817 char *cmd_np, /* Switch for -np or -ntmpi or empty */
818 char *cmd_mdrun, /* Command for mdrun */
819 char *args_for_mdrun, /* Arguments for mdrun */
820 const char *simulation_tpr, /* This tpr will be simulated */
821 int nnodes, /* Number of ranks to use */
822 int nPMEnodes, /* Number of PME ranks to use */
823 const t_eligible_gpu_ids *gpu_ids) /* Struct containing GPU IDs for
824 * constructing mdrun command lines */
826 char *command, *cmd_gpu_ids;
829 /* Make enough space for the system call command,
830 * (200 extra chars for -npme ... etc. options should suffice): */
831 snew(command, std::strlen(cmd_mpirun)+std::strlen(cmd_mdrun)+std::strlen(cmd_np)+std::strlen(args_for_mdrun)+std::strlen(simulation_tpr)+200);
833 cmd_gpu_ids = make_gpu_id_command_line(nnodes, nPMEnodes, gpu_ids);
835 /* Note that the -passall options requires args_for_mdrun to be at the end
836 * of the command line string */
839 sprintf(command, "%s%s-npme %d -s %s %s %s",
840 cmd_mdrun, cmd_np, nPMEnodes, simulation_tpr, args_for_mdrun, cmd_gpu_ids);
844 sprintf(command, "%s%s%s -npme %d -s %s %s %s",
845 cmd_mpirun, cmd_np, cmd_mdrun, nPMEnodes, simulation_tpr, args_for_mdrun, cmd_gpu_ids);
848 fprintf(fp, "%s this command line to launch the simulation:\n\n%s", bLaunch ? "Using" : "Please use", command);
852 /* Now the real thing! */
855 fprintf(stdout, "\nLaunching simulation with best parameters now.\nExecuting '%s'", command);
858 gmx_system_call(command);
863 static void modify_PMEsettings(
864 gmx_int64_t simsteps, /* Set this value as number of time steps */
865 gmx_int64_t init_step, /* Set this value as init_step */
866 const char *fn_best_tpr, /* tpr file with the best performance */
867 const char *fn_sim_tpr) /* name of tpr file to be launched */
875 read_tpx_state(fn_best_tpr, ir, &state, NULL, &mtop);
877 /* Reset nsteps and init_step to the value of the input .tpr file */
878 ir->nsteps = simsteps;
879 ir->init_step = init_step;
881 /* Write the tpr file which will be launched */
882 sprintf(buf, "Writing optimized simulation file %s with nsteps=%s.\n", fn_sim_tpr, "%" GMX_PRId64);
883 fprintf(stdout, buf, ir->nsteps);
885 write_tpx_state(fn_sim_tpr, ir, &state, &mtop);
890 static gmx_bool can_scale_rvdw(int vdwtype)
892 return (evdwCUT == vdwtype ||
896 #define EPME_SWITCHED(e) ((e) == eelPMESWITCH || (e) == eelPMEUSERSWITCH)
898 /* Make additional TPR files with more computational load for the
899 * direct space processors: */
900 static void make_benchmark_tprs(
901 const char *fn_sim_tpr, /* READ : User-provided tpr file */
902 char *fn_bench_tprs[], /* WRITE: Names of benchmark tpr files */
903 gmx_int64_t benchsteps, /* Number of time steps for benchmark runs */
904 gmx_int64_t statesteps, /* Step counter in checkpoint file */
905 real rmin, /* Minimal Coulomb radius */
906 real rmax, /* Maximal Coulomb radius */
907 real bScaleRvdw, /* Scale rvdw along with rcoulomb */
908 int *ntprs, /* No. of TPRs to write, each with a different
909 rcoulomb and fourierspacing */
910 t_inputinfo *info, /* Contains information about mdp file options */
911 FILE *fp) /* Write the output here */
917 real nlist_buffer; /* Thickness of the buffer regions for PME-switch potentials */
920 gmx_bool bNote = FALSE;
921 real add; /* Add this to rcoul for the next test */
922 real fac = 1.0; /* Scaling factor for Coulomb radius */
923 real fourierspacing; /* Basic fourierspacing from tpr */
926 sprintf(buf, "Making benchmark tpr file%s with %s time step%s",
927 *ntprs > 1 ? "s" : "", "%" GMX_PRId64, benchsteps > 1 ? "s" : "");
928 fprintf(stdout, buf, benchsteps);
931 sprintf(buf, " (adding %s steps from checkpoint file)", "%" GMX_PRId64);
932 fprintf(stdout, buf, statesteps);
933 benchsteps += statesteps;
935 fprintf(stdout, ".\n");
939 read_tpx_state(fn_sim_tpr, ir, &state, NULL, &mtop);
941 /* Check if some kind of PME was chosen */
942 if (EEL_PME(ir->coulombtype) == FALSE)
944 gmx_fatal(FARGS, "Can only do optimizations for simulations with %s electrostatics.",
948 /* Check if rcoulomb == rlist, which is necessary for plain PME. */
949 if ( (ir->cutoff_scheme != ecutsVERLET) &&
950 (eelPME == ir->coulombtype) && !(ir->rcoulomb == ir->rlist))
952 gmx_fatal(FARGS, "%s requires rcoulomb (%f) to be equal to rlist (%f).",
953 EELTYPE(eelPME), ir->rcoulomb, ir->rlist);
955 /* For other PME types, rcoulomb is allowed to be smaller than rlist */
956 else if (ir->rcoulomb > ir->rlist)
958 gmx_fatal(FARGS, "%s requires rcoulomb (%f) to be equal to or smaller than rlist (%f)",
959 EELTYPE(ir->coulombtype), ir->rcoulomb, ir->rlist);
962 if (bScaleRvdw && ir->rvdw != ir->rcoulomb)
964 fprintf(stdout, "NOTE: input rvdw != rcoulomb, will not scale rvdw\n");
968 /* Reduce the number of steps for the benchmarks */
969 info->orig_sim_steps = ir->nsteps;
970 ir->nsteps = benchsteps;
971 /* We must not use init_step from the input tpr file for the benchmarks */
972 info->orig_init_step = ir->init_step;
975 /* For PME-switch potentials, keep the radial distance of the buffer region */
976 nlist_buffer = ir->rlist - ir->rcoulomb;
978 /* Determine length of triclinic box vectors */
979 for (d = 0; d < DIM; d++)
982 for (i = 0; i < DIM; i++)
984 box_size[d] += state.box[d][i]*state.box[d][i];
986 box_size[d] = std::sqrt(box_size[d]);
989 if (ir->fourier_spacing > 0)
991 info->fsx[0] = ir->fourier_spacing;
992 info->fsy[0] = ir->fourier_spacing;
993 info->fsz[0] = ir->fourier_spacing;
997 /* Reconstruct fourierspacing per dimension from the number of grid points and box size */
998 info->fsx[0] = box_size[XX]/ir->nkx;
999 info->fsy[0] = box_size[YY]/ir->nky;
1000 info->fsz[0] = box_size[ZZ]/ir->nkz;
1003 /* If no value for the fourierspacing was provided on the command line, we
1004 * use the reconstruction from the tpr file */
1005 if (ir->fourier_spacing > 0)
1007 /* Use the spacing from the tpr */
1008 fourierspacing = ir->fourier_spacing;
1012 /* Use the maximum observed spacing */
1013 fourierspacing = std::max(std::max(info->fsx[0], info->fsy[0]), info->fsz[0]);
1016 fprintf(stdout, "Calculating PME grid points on the basis of a fourierspacing of %f nm\n", fourierspacing);
1018 /* For performance comparisons the number of particles is useful to have */
1019 fprintf(fp, " Number of particles : %d\n", mtop.natoms);
1021 /* Print information about settings of which some are potentially modified: */
1022 fprintf(fp, " Coulomb type : %s\n", EELTYPE(ir->coulombtype));
1023 fprintf(fp, " Grid spacing x y z : %f %f %f\n",
1024 box_size[XX]/ir->nkx, box_size[YY]/ir->nky, box_size[ZZ]/ir->nkz);
1025 fprintf(fp, " Van der Waals type : %s\n", EVDWTYPE(ir->vdwtype));
1026 if (ir_vdw_switched(ir))
1028 fprintf(fp, " rvdw_switch : %f nm\n", ir->rvdw_switch);
1030 if (EPME_SWITCHED(ir->coulombtype))
1032 fprintf(fp, " rlist : %f nm\n", ir->rlist);
1034 if (ir->rlistlong != max_cutoff(ir->rvdw, ir->rcoulomb))
1036 fprintf(fp, " rlistlong : %f nm\n", ir->rlistlong);
1039 /* Print a descriptive line about the tpr settings tested */
1040 fprintf(fp, "\nWill try these real/reciprocal workload settings:\n");
1041 fprintf(fp, " No. scaling rcoulomb");
1042 fprintf(fp, " nkx nky nkz");
1043 fprintf(fp, " spacing");
1044 if (can_scale_rvdw(ir->vdwtype))
1046 fprintf(fp, " rvdw");
1048 if (EPME_SWITCHED(ir->coulombtype))
1050 fprintf(fp, " rlist");
1052 if (ir->rlistlong != max_cutoff(ir->rlist, max_cutoff(ir->rvdw, ir->rcoulomb)) )
1054 fprintf(fp, " rlistlong");
1056 fprintf(fp, " tpr file\n");
1058 /* Loop to create the requested number of tpr input files */
1059 for (j = 0; j < *ntprs; j++)
1061 /* The first .tpr is the provided one, just need to modify nsteps,
1062 * so skip the following block */
1065 /* Determine which Coulomb radii rc to use in the benchmarks */
1066 add = (rmax-rmin)/(*ntprs-1);
1067 if (gmx_within_tol(rmin, info->rcoulomb[0], GMX_REAL_EPS))
1069 ir->rcoulomb = rmin + j*add;
1071 else if (gmx_within_tol(rmax, info->rcoulomb[0], GMX_REAL_EPS))
1073 ir->rcoulomb = rmin + (j-1)*add;
1077 /* rmin != rcoul != rmax, ergo test between rmin and rmax */
1078 add = (rmax-rmin)/(*ntprs-2);
1079 ir->rcoulomb = rmin + (j-1)*add;
1082 /* Determine the scaling factor fac */
1083 fac = ir->rcoulomb/info->rcoulomb[0];
1085 /* Scale the Fourier grid spacing */
1086 ir->nkx = ir->nky = ir->nkz = 0;
1087 calc_grid(NULL, state.box, fourierspacing*fac, &ir->nkx, &ir->nky, &ir->nkz);
1089 /* Adjust other radii since various conditions need to be fulfilled */
1090 if (eelPME == ir->coulombtype)
1092 /* plain PME, rcoulomb must be equal to rlist */
1093 ir->rlist = ir->rcoulomb;
1097 /* rlist must be >= rcoulomb, we keep the size of the buffer region */
1098 ir->rlist = ir->rcoulomb + nlist_buffer;
1101 if (bScaleRvdw && can_scale_rvdw(ir->vdwtype))
1103 if (ecutsVERLET == ir->cutoff_scheme ||
1104 evdwPME == ir->vdwtype)
1106 /* With either the Verlet cutoff-scheme or LJ-PME,
1107 the van der Waals radius must always equal the
1109 ir->rvdw = ir->rcoulomb;
1113 /* For vdw cutoff, rvdw >= rlist */
1114 ir->rvdw = std::max(info->rvdw[0], ir->rlist);
1118 ir->rlistlong = max_cutoff(ir->rlist, max_cutoff(ir->rvdw, ir->rcoulomb));
1120 } /* end of "if (j != 0)" */
1122 /* for j==0: Save the original settings
1123 * for j >0: Save modified radii and Fourier grids */
1124 info->rcoulomb[j] = ir->rcoulomb;
1125 info->rvdw[j] = ir->rvdw;
1126 info->nkx[j] = ir->nkx;
1127 info->nky[j] = ir->nky;
1128 info->nkz[j] = ir->nkz;
1129 info->rlist[j] = ir->rlist;
1130 info->rlistlong[j] = ir->rlistlong;
1131 info->fsx[j] = fac*fourierspacing;
1132 info->fsy[j] = fac*fourierspacing;
1133 info->fsz[j] = fac*fourierspacing;
1135 /* Write the benchmark tpr file */
1136 std::strncpy(fn_bench_tprs[j], fn_sim_tpr, std::strlen(fn_sim_tpr)-std::strlen(".tpr"));
1137 sprintf(buf, "_bench%.2d.tpr", j);
1138 std::strcat(fn_bench_tprs[j], buf);
1139 fprintf(stdout, "Writing benchmark tpr %s with nsteps=", fn_bench_tprs[j]);
1140 fprintf(stdout, "%" GMX_PRId64, ir->nsteps);
1143 fprintf(stdout, ", scaling factor %f\n", fac);
1147 fprintf(stdout, ", unmodified settings\n");
1150 write_tpx_state(fn_bench_tprs[j], ir, &state, &mtop);
1152 /* Write information about modified tpr settings to log file */
1153 fprintf(fp, "%4d%10f%10f", j, fac, ir->rcoulomb);
1154 fprintf(fp, "%5d%5d%5d", ir->nkx, ir->nky, ir->nkz);
1155 fprintf(fp, " %9f ", info->fsx[j]);
1156 if (can_scale_rvdw(ir->vdwtype))
1158 fprintf(fp, "%10f", ir->rvdw);
1160 if (EPME_SWITCHED(ir->coulombtype))
1162 fprintf(fp, "%10f", ir->rlist);
1164 if (info->rlistlong[0] != max_cutoff(info->rlist[0], max_cutoff(info->rvdw[0], info->rcoulomb[0])) )
1166 fprintf(fp, "%10f", ir->rlistlong);
1168 fprintf(fp, " %-14s\n", fn_bench_tprs[j]);
1170 /* Make it clear to the user that some additional settings were modified */
1171 if (!gmx_within_tol(ir->rvdw, info->rvdw[0], GMX_REAL_EPS)
1172 || !gmx_within_tol(ir->rlistlong, info->rlistlong[0], GMX_REAL_EPS) )
1179 fprintf(fp, "\nNote that in addition to the Coulomb radius and the Fourier grid\n"
1180 "other input settings were also changed (see table above).\n"
1181 "Please check if the modified settings are appropriate.\n");
1189 /* Rename the files we want to keep to some meaningful filename and
1190 * delete the rest */
1191 static void cleanup(const t_filenm *fnm, int nfile, int k, int nnodes,
1192 int nPMEnodes, int nr, gmx_bool bKeepStderr)
1194 char numstring[STRLEN];
1195 char newfilename[STRLEN];
1196 const char *fn = NULL;
1201 fprintf(stdout, "Cleaning up, deleting benchmark temp files ...\n");
1203 for (i = 0; i < nfile; i++)
1205 opt = (char *)fnm[i].opt;
1206 if (std::strcmp(opt, "-p") == 0)
1208 /* do nothing; keep this file */
1211 else if (std::strcmp(opt, "-bg") == 0)
1213 /* Give the log file a nice name so one can later see which parameters were used */
1214 numstring[0] = '\0';
1217 sprintf(numstring, "_%d", nr);
1219 sprintf(newfilename, "%s_no%d_np%d_npme%d%s", opt2fn("-bg", nfile, fnm), k, nnodes, nPMEnodes, numstring);
1220 if (gmx_fexist(opt2fn("-bg", nfile, fnm)))
1222 fprintf(stdout, "renaming log file to %s\n", newfilename);
1223 make_backup(newfilename);
1224 rename(opt2fn("-bg", nfile, fnm), newfilename);
1227 else if (std::strcmp(opt, "-err") == 0)
1229 /* This file contains the output of stderr. We want to keep it in
1230 * cases where there have been problems. */
1231 fn = opt2fn(opt, nfile, fnm);
1232 numstring[0] = '\0';
1235 sprintf(numstring, "_%d", nr);
1237 sprintf(newfilename, "%s_no%d_np%d_npme%d%s", fn, k, nnodes, nPMEnodes, numstring);
1242 fprintf(stdout, "Saving stderr output in %s\n", newfilename);
1243 make_backup(newfilename);
1244 rename(fn, newfilename);
1248 fprintf(stdout, "Deleting %s\n", fn);
1253 /* Delete the files which are created for each benchmark run: (options -b*) */
1254 else if ( (0 == std::strncmp(opt, "-b", 2)) && (opt2bSet(opt, nfile, fnm) || !is_optional(&fnm[i])) )
1256 remove_if_exists(opt2fn(opt, nfile, fnm));
1263 eNpmeAuto, eNpmeAll, eNpmeReduced, eNpmeSubset, eNpmeNr
1266 /* Create a list of numbers of PME nodes to test */
1267 static void make_npme_list(
1268 const char *npmevalues_opt, /* Make a complete list with all
1269 * possibilities or a short list that keeps only
1270 * reasonable numbers of PME nodes */
1271 int *nentries, /* Number of entries we put in the nPMEnodes list */
1272 int *nPMEnodes[], /* Each entry contains the value for -npme */
1273 int nnodes, /* Total number of nodes to do the tests on */
1274 int minPMEnodes, /* Minimum number of PME nodes */
1275 int maxPMEnodes) /* Maximum number of PME nodes */
1278 int min_factor = 1; /* We request that npp and npme have this minimal
1279 * largest common factor (depends on npp) */
1280 int nlistmax; /* Max. list size */
1281 int nlist; /* Actual number of entries in list */
1285 /* Do we need to check all possible values for -npme or is a reduced list enough? */
1286 if (!std::strcmp(npmevalues_opt, "all") )
1290 else if (!std::strcmp(npmevalues_opt, "subset") )
1292 eNPME = eNpmeSubset;
1294 else /* "auto" or "range" */
1300 else if (nnodes < 128)
1302 eNPME = eNpmeReduced;
1306 eNPME = eNpmeSubset;
1310 /* Calculate how many entries we could possibly have (in case of -npme all) */
1313 nlistmax = maxPMEnodes - minPMEnodes + 3;
1314 if (0 == minPMEnodes)
1324 /* Now make the actual list which is at most of size nlist */
1325 snew(*nPMEnodes, nlistmax);
1326 nlist = 0; /* start counting again, now the real entries in the list */
1327 for (i = 0; i < nlistmax - 2; i++)
1329 npme = maxPMEnodes - i;
1340 /* For 2d PME we want a common largest factor of at least the cube
1341 * root of the number of PP nodes */
1342 min_factor = static_cast<int>(std::pow(npp, 1.0/3.0));
1345 gmx_fatal(FARGS, "Unknown option for eNPME in make_npme_list");
1348 if (gmx_greatest_common_divisor(npp, npme) >= min_factor)
1350 (*nPMEnodes)[nlist] = npme;
1354 /* We always test 0 PME nodes and the automatic number */
1355 *nentries = nlist + 2;
1356 (*nPMEnodes)[nlist ] = 0;
1357 (*nPMEnodes)[nlist+1] = -1;
1359 fprintf(stderr, "Will try the following %d different values for -npme:\n", *nentries);
1360 for (i = 0; i < *nentries-1; i++)
1362 fprintf(stderr, "%d, ", (*nPMEnodes)[i]);
1364 fprintf(stderr, "and %d (auto).\n", (*nPMEnodes)[*nentries-1]);
1368 /* Allocate memory to store the performance data */
1369 static void init_perfdata(t_perf *perfdata[], int ntprs, int datasets, int repeats)
1374 for (k = 0; k < ntprs; k++)
1376 snew(perfdata[k], datasets);
1377 for (i = 0; i < datasets; i++)
1379 for (j = 0; j < repeats; j++)
1381 snew(perfdata[k][i].Gcycles, repeats);
1382 snew(perfdata[k][i].ns_per_day, repeats);
1383 snew(perfdata[k][i].PME_f_load, repeats);
1390 /* Check for errors on mdrun -h */
1391 static void make_sure_it_runs(char *mdrun_cmd_line, int length, FILE *fp,
1392 const t_filenm *fnm, int nfile)
1394 char *command, *msg;
1397 snew(command, length + 15);
1398 snew(msg, length + 500);
1400 fprintf(stdout, "Making sure the benchmarks can be executed by running just 1 step...\n");
1401 sprintf(command, "%s -nsteps 1 -quiet", mdrun_cmd_line);
1402 fprintf(stdout, "Executing '%s' ...\n", command);
1403 ret = gmx_system_call(command);
1407 /* To prevent confusion, do not again issue a gmx_fatal here since we already
1408 * get the error message from mdrun itself */
1410 "Cannot run the first benchmark simulation! Please check the error message of\n"
1411 "mdrun for the source of the problem. Did you provide a command line\n"
1412 "argument that neither gmx tune_pme nor mdrun understands? If you're\n"
1413 "sure your command line should work, you can bypass this check with \n"
1414 "gmx tune_pme -nocheck. The failing command was:\n"
1415 "\n%s\n\n", command);
1417 fprintf(stderr, "%s", msg);
1419 fprintf(fp, "%s", msg);
1423 fprintf(stdout, "Benchmarks can be executed!\n");
1425 /* Clean up the benchmark output files we just created */
1426 fprintf(stdout, "Cleaning up ...\n");
1427 remove_if_exists(opt2fn("-bc", nfile, fnm));
1428 remove_if_exists(opt2fn("-be", nfile, fnm));
1429 remove_if_exists(opt2fn("-bcpo", nfile, fnm));
1430 remove_if_exists(opt2fn("-bg", nfile, fnm));
1436 static void do_the_tests(
1437 FILE *fp, /* General g_tune_pme output file */
1438 char **tpr_names, /* Filenames of the input files to test */
1439 int maxPMEnodes, /* Max fraction of nodes to use for PME */
1440 int minPMEnodes, /* Min fraction of nodes to use for PME */
1441 int npme_fixed, /* If >= -1, test fixed number of PME
1443 const char *npmevalues_opt, /* Which -npme values should be tested */
1444 t_perf **perfdata, /* Here the performace data is stored */
1445 int *pmeentries, /* Entries in the nPMEnodes list */
1446 int repeats, /* Repeat each test this often */
1447 int nnodes, /* Total number of nodes = nPP + nPME */
1448 int nr_tprs, /* Total number of tpr files to test */
1449 gmx_bool bThreads, /* Threads or MPI? */
1450 char *cmd_mpirun, /* mpirun command string */
1451 char *cmd_np, /* "-np", "-n", whatever mpirun needs */
1452 char *cmd_mdrun, /* mdrun command string */
1453 char *cmd_args_bench, /* arguments for mdrun in a string */
1454 const t_filenm *fnm, /* List of filenames from command line */
1455 int nfile, /* Number of files specified on the cmdl. */
1456 int presteps, /* DLB equilibration steps, is checked */
1457 gmx_int64_t cpt_steps, /* Time step counter in the checkpoint */
1458 gmx_bool bCheck, /* Check whether benchmark mdrun works */
1459 const t_eligible_gpu_ids *gpu_ids) /* Struct containing GPU IDs for
1460 * constructing mdrun command lines */
1462 int i, nr, k, ret, count = 0, totaltests;
1463 int *nPMEnodes = NULL;
1466 char *command, *cmd_stub;
1468 gmx_bool bResetProblem = FALSE;
1469 gmx_bool bFirst = TRUE;
1471 /* This string array corresponds to the eParselog enum type at the start
1473 const char* ParseLog[] = {
1475 "Logfile not found!",
1476 "No timings, logfile truncated?",
1477 "Run was terminated.",
1478 "Counters were not reset properly.",
1479 "No DD grid found for these settings.",
1480 "TPX version conflict!",
1481 "mdrun was not started in parallel!",
1482 "Number of PP ranks has a prime factor that is too large.",
1483 "The number of PP ranks did not suit the number of GPUs.",
1484 "Some GPUs were not detected or are incompatible.",
1487 char str_PME_f_load[13];
1490 /* Allocate space for the mdrun command line. 100 extra characters should
1491 be more than enough for the -npme etcetera arguments */
1492 cmdline_length = std::strlen(cmd_mpirun)
1493 + std::strlen(cmd_np)
1494 + std::strlen(cmd_mdrun)
1495 + std::strlen(cmd_args_bench)
1496 + std::strlen(tpr_names[0]) + 100;
1497 snew(command, cmdline_length);
1498 snew(cmd_stub, cmdline_length);
1500 /* Construct the part of the command line that stays the same for all tests: */
1503 sprintf(cmd_stub, "%s%s", cmd_mdrun, cmd_np);
1507 sprintf(cmd_stub, "%s%s%s ", cmd_mpirun, cmd_np, cmd_mdrun);
1510 /* Create a list of numbers of PME nodes to test */
1511 if (npme_fixed < -1)
1513 make_npme_list(npmevalues_opt, pmeentries, &nPMEnodes,
1514 nnodes, minPMEnodes, maxPMEnodes);
1520 nPMEnodes[0] = npme_fixed;
1521 fprintf(stderr, "Will use a fixed number of %d PME-only ranks.\n", nPMEnodes[0]);
1526 fprintf(fp, "\nNo benchmarks done since number of repeats (-r) is 0.\n");
1528 finalize(opt2fn("-p", nfile, fnm));
1532 /* Allocate one dataset for each tpr input file: */
1533 init_perfdata(perfdata, nr_tprs, *pmeentries, repeats);
1535 /*****************************************/
1536 /* Main loop over all tpr files to test: */
1537 /*****************************************/
1538 totaltests = nr_tprs*(*pmeentries)*repeats;
1539 for (k = 0; k < nr_tprs; k++)
1541 fprintf(fp, "\nIndividual timings for input file %d (%s):\n", k, tpr_names[k]);
1542 fprintf(fp, "PME ranks Gcycles ns/day PME/f Remark\n");
1543 /* Loop over various numbers of PME nodes: */
1544 for (i = 0; i < *pmeentries; i++)
1546 char *cmd_gpu_ids = NULL;
1548 pd = &perfdata[k][i];
1550 cmd_gpu_ids = make_gpu_id_command_line(nnodes, nPMEnodes[i], gpu_ids);
1552 /* Loop over the repeats for each scenario: */
1553 for (nr = 0; nr < repeats; nr++)
1555 pd->nPMEnodes = nPMEnodes[i];
1557 /* Add -npme and -s to the command line and save it. Note that
1558 * the -passall (if set) options requires cmd_args_bench to be
1559 * at the end of the command line string */
1560 snew(pd->mdrun_cmd_line, cmdline_length);
1561 sprintf(pd->mdrun_cmd_line, "%s-npme %d -s %s %s %s",
1562 cmd_stub, pd->nPMEnodes, tpr_names[k], cmd_args_bench, cmd_gpu_ids);
1564 /* To prevent that all benchmarks fail due to a show-stopper argument
1565 * on the mdrun command line, we make a quick check first.
1566 * This check can be turned off in cases where the automatically chosen
1567 * number of PME-only ranks leads to a number of PP ranks for which no
1568 * decomposition can be found (e.g. for large prime numbers) */
1569 if (bFirst && bCheck)
1571 /* TODO This check is really for a functional
1572 * .tpr, and if we need it, it should take place
1573 * for every .tpr, and the logic for it should be
1574 * immediately inside the loop over k, not in
1575 * this inner loop. */
1576 char *temporary_cmd_line;
1578 snew(temporary_cmd_line, cmdline_length);
1579 /* TODO -npme 0 is more likely to succeed at low
1580 parallelism than the default of -npme -1, but
1581 is more likely to fail at the scaling limit
1582 when the PP domains may be too small. "mpirun
1583 -np 1 mdrun" is probably a reasonable thing to
1584 do for this check, but it'll be easier to
1585 implement that after some refactoring of how
1586 the number of MPI ranks is managed. */
1587 sprintf(temporary_cmd_line, "%s-npme 0 -nb cpu -s %s %s",
1588 cmd_stub, tpr_names[k], cmd_args_bench);
1589 make_sure_it_runs(temporary_cmd_line, cmdline_length, fp, fnm, nfile);
1593 /* Do a benchmark simulation: */
1596 sprintf(buf, ", pass %d/%d", nr+1, repeats);
1602 fprintf(stdout, "\n=== Progress %2.0f%%, tpr %d/%d, run %d/%d%s:\n",
1603 (100.0*count)/totaltests,
1604 k+1, nr_tprs, i+1, *pmeentries, buf);
1605 make_backup(opt2fn("-err", nfile, fnm));
1606 sprintf(command, "%s 1> /dev/null 2>%s", pd->mdrun_cmd_line, opt2fn("-err", nfile, fnm));
1607 fprintf(stdout, "%s\n", pd->mdrun_cmd_line);
1608 gmx_system_call(command);
1610 /* Collect the performance data from the log file; also check stderr
1611 * for fatal errors */
1612 ret = parse_logfile(opt2fn("-bg", nfile, fnm), opt2fn("-err", nfile, fnm),
1613 pd, nr, presteps, cpt_steps, nnodes);
1614 if ((presteps > 0) && (ret == eParselogResetProblem))
1616 bResetProblem = TRUE;
1619 if (-1 == pd->nPMEnodes)
1621 sprintf(buf, "(%3d)", pd->guessPME);
1629 if (pd->PME_f_load[nr] > 0.0)
1631 sprintf(str_PME_f_load, "%12.3f", pd->PME_f_load[nr]);
1635 sprintf(str_PME_f_load, "%s", " - ");
1638 /* Write the data we got to disk */
1639 fprintf(fp, "%4d%s %12.3f %12.3f %s %s", pd->nPMEnodes,
1640 buf, pd->Gcycles[nr], pd->ns_per_day[nr], str_PME_f_load, ParseLog[ret]);
1641 if (!(ret == eParselogOK || ret == eParselogNoDDGrid || ret == eParselogNotFound) )
1643 fprintf(fp, " Check %s file for problems.", ret == eParselogFatal ? "err" : "log");
1649 /* Do some cleaning up and delete the files we do not need any more */
1650 cleanup(fnm, nfile, k, nnodes, pd->nPMEnodes, nr, ret == eParselogFatal);
1652 /* If the first run with this number of processors already failed, do not try again: */
1653 if (pd->Gcycles[0] <= 0.0 && repeats > 1)
1655 fprintf(stdout, "Skipping remaining passes of unsuccessful setting, see log file for details.\n");
1656 count += repeats-(nr+1);
1659 } /* end of repeats loop */
1661 } /* end of -npme loop */
1662 } /* end of tpr file loop */
1667 fprintf(fp, "WARNING: The cycle and time step counters could not be reset properly. ");
1675 static void check_input(
1682 real maxPMEfraction,
1683 real minPMEfraction,
1685 gmx_int64_t bench_nsteps,
1686 const t_filenm *fnm,
1696 /* Make sure the input file exists */
1697 if (!gmx_fexist(opt2fn("-s", nfile, fnm)))
1699 gmx_fatal(FARGS, "File %s not found.", opt2fn("-s", nfile, fnm));
1702 /* Make sure that the checkpoint file is not overwritten during benchmarking */
1703 if ( (0 == std::strcmp(opt2fn("-cpi", nfile, fnm), opt2fn("-bcpo", nfile, fnm)) ) && (sim_part > 1) )
1705 gmx_fatal(FARGS, "Checkpoint input (-cpi) and benchmark checkpoint output (-bcpo) files must not be identical.\n"
1706 "The checkpoint input file must not be overwritten during the benchmarks.\n");
1709 /* Make sure that repeats is >= 0 (if == 0, only write tpr files) */
1712 gmx_fatal(FARGS, "Number of repeats < 0!");
1715 /* Check number of nodes */
1718 gmx_fatal(FARGS, "Number of ranks/threads must be a positive integer.");
1721 /* Automatically choose -ntpr if not set */
1731 /* Set a reasonable scaling factor for rcoulomb */
1734 *rmax = rcoulomb * 1.2;
1737 fprintf(stderr, "Will test %d tpr file%s.\n", *ntprs, *ntprs == 1 ? "" : "s");
1743 fprintf(stderr, "Note: Choose ntpr>1 to shift PME load between real and reciprocal space.\n");
1747 /* Make shure that rmin <= rcoulomb <= rmax */
1756 if (!(*rmin <= *rmax) )
1758 gmx_fatal(FARGS, "Please choose the Coulomb radii such that rmin <= rmax.\n"
1759 "rmin = %g, rmax = %g, actual rcoul from .tpr file = %g\n", *rmin, *rmax, rcoulomb);
1761 /* Add test scenarios if rmin or rmax were set */
1764 if (!gmx_within_tol(*rmin, rcoulomb, GMX_REAL_EPS) && (*ntprs == 1) )
1767 fprintf(stderr, "NOTE: Setting -rmin to %g changed -ntpr to %d\n",
1770 if (!gmx_within_tol(*rmax, rcoulomb, GMX_REAL_EPS) && (*ntprs == 1) )
1773 fprintf(stderr, "NOTE: Setting -rmax to %g changed -ntpr to %d\n",
1778 /* If one of rmin, rmax is set, we need 2 tpr files at minimum */
1779 if (!gmx_within_tol(*rmax, rcoulomb, GMX_REAL_EPS) || !gmx_within_tol(*rmin, rcoulomb, GMX_REAL_EPS) )
1781 *ntprs = std::max(*ntprs, 2);
1784 /* If both rmin, rmax are set, we need 3 tpr files at minimum */
1785 if (!gmx_within_tol(*rmax, rcoulomb, GMX_REAL_EPS) && !gmx_within_tol(*rmin, rcoulomb, GMX_REAL_EPS) )
1787 *ntprs = std::max(*ntprs, 3);
1792 fprintf(stderr, "NOTE: Your rmin, rmax setting changed -ntpr to %d\n", *ntprs);
1797 if (gmx_within_tol(*rmin, rcoulomb, GMX_REAL_EPS) && gmx_within_tol(rcoulomb, *rmax, GMX_REAL_EPS)) /* We have just a single rc */
1799 fprintf(stderr, "WARNING: Resetting -ntpr to 1 since no Coulomb radius scaling is requested.\n"
1800 "Please set rmin < rmax to test Coulomb radii in the [rmin, rmax] interval\n"
1801 "with correspondingly adjusted PME grid settings\n");
1806 /* Check whether max and min fraction are within required values */
1807 if (maxPMEfraction > 0.5 || maxPMEfraction < 0)
1809 gmx_fatal(FARGS, "-max must be between 0 and 0.5");
1811 if (minPMEfraction > 0.5 || minPMEfraction < 0)
1813 gmx_fatal(FARGS, "-min must be between 0 and 0.5");
1815 if (maxPMEfraction < minPMEfraction)
1817 gmx_fatal(FARGS, "-max must be larger or equal to -min");
1820 /* Check whether the number of steps - if it was set - has a reasonable value */
1821 if (bench_nsteps < 0)
1823 gmx_fatal(FARGS, "Number of steps must be positive.");
1826 if (bench_nsteps > 10000 || bench_nsteps < 100)
1828 fprintf(stderr, "WARNING: steps=");
1829 fprintf(stderr, "%" GMX_PRId64, bench_nsteps);
1830 fprintf(stderr, ". Are you sure you want to perform so %s steps for each benchmark?\n", (bench_nsteps < 100) ? "few" : "many");
1835 gmx_fatal(FARGS, "Cannot have a negative number of presteps.\n");
1838 /* Check for rcoulomb scaling if more than one .tpr file is tested */
1841 if (*rmin/rcoulomb < 0.75 || *rmax/rcoulomb > 1.25)
1843 fprintf(stderr, "WARNING: Applying extreme scaling factor. I hope you know what you are doing.\n");
1847 /* If a fixed number of PME nodes is set we do rcoulomb and PME gird tuning
1848 * only. We need to check whether the requested number of PME-only nodes
1850 if (npme_fixed > -1)
1852 /* No more than 50% of all nodes can be assigned as PME-only nodes. */
1853 if (2*npme_fixed > nnodes)
1855 gmx_fatal(FARGS, "Cannot have more than %d PME-only ranks for a total of %d ranks (you chose %d).\n",
1856 nnodes/2, nnodes, npme_fixed);
1858 if ((npme_fixed > 0) && (5*npme_fixed < nnodes))
1860 fprintf(stderr, "WARNING: Only %g percent of the ranks are assigned as PME-only ranks.\n",
1861 (100.0*npme_fixed)/nnodes);
1863 if (opt2parg_bSet("-min", npargs, pa) || opt2parg_bSet("-max", npargs, pa))
1865 fprintf(stderr, "NOTE: The -min, -max, and -npme options have no effect when a\n"
1866 " fixed number of PME-only ranks is requested with -fix.\n");
1872 /* Returns TRUE when "opt" is needed at launch time */
1873 static gmx_bool is_launch_file(char *opt, gmx_bool bSet)
1875 if (0 == std::strncmp(opt, "-swap", 5))
1880 /* Apart from the input .tpr and the output log files we need all options that
1881 * were set on the command line and that do not start with -b */
1882 if (0 == std::strncmp(opt, "-b", 2) || 0 == std::strncmp(opt, "-s", 2)
1883 || 0 == std::strncmp(opt, "-err", 4) || 0 == std::strncmp(opt, "-p", 2) )
1892 /* Returns TRUE when "opt" defines a file which is needed for the benchmarks runs */
1893 static gmx_bool is_bench_file(char *opt, gmx_bool bSet, gmx_bool bOptional, gmx_bool bIsOutput)
1895 /* Apart from the input .tpr, all files starting with "-b" are for
1896 * _b_enchmark files exclusively */
1897 if (0 == std::strncmp(opt, "-s", 2))
1902 if (0 == std::strncmp(opt, "-b", 2) || 0 == std::strncmp(opt, "-s", 2))
1904 if (!bOptional || bSet)
1921 if (bSet) /* These are additional input files like -cpi -ei */
1934 /* Adds 'buf' to 'str' */
1935 static void add_to_string(char **str, const char *buf)
1940 len = std::strlen(*str) + std::strlen(buf) + 1;
1942 std::strcat(*str, buf);
1946 /* Create the command line for the benchmark as well as for the real run */
1947 static void create_command_line_snippets(
1948 gmx_bool bAppendFiles,
1949 gmx_bool bKeepAndNumCPT,
1950 gmx_bool bResetHWay,
1954 char *cmd_args_bench[], /* command line arguments for benchmark runs */
1955 char *cmd_args_launch[], /* command line arguments for simulation run */
1956 char extra_args[]) /* Add this to the end of the command line */
1961 char strbuf[STRLEN];
1964 /* strlen needs at least '\0' as a string: */
1965 snew(*cmd_args_bench, 1);
1966 snew(*cmd_args_launch, 1);
1967 *cmd_args_launch[0] = '\0';
1968 *cmd_args_bench[0] = '\0';
1971 /*******************************************/
1972 /* 1. Process other command line arguments */
1973 /*******************************************/
1976 /* Add equilibration steps to benchmark options */
1977 sprintf(strbuf, "-resetstep %d ", presteps);
1978 add_to_string(cmd_args_bench, strbuf);
1980 /* These switches take effect only at launch time */
1981 if (FALSE == bAppendFiles)
1983 add_to_string(cmd_args_launch, "-noappend ");
1987 add_to_string(cmd_args_launch, "-cpnum ");
1991 add_to_string(cmd_args_launch, "-resethway ");
1994 /********************/
1995 /* 2. Process files */
1996 /********************/
1997 for (i = 0; i < nfile; i++)
1999 opt = (char *)fnm[i].opt;
2000 name = opt2fn(opt, nfile, fnm);
2002 /* Strbuf contains the options, now let's sort out where we need that */
2003 sprintf(strbuf, "%s %s ", opt, name);
2005 if (is_bench_file(opt, opt2bSet(opt, nfile, fnm), is_optional(&fnm[i]), is_output(&fnm[i])) )
2007 /* All options starting with -b* need the 'b' removed,
2008 * therefore overwrite strbuf */
2009 if (0 == std::strncmp(opt, "-b", 2))
2011 sprintf(strbuf, "-%s %s ", &opt[2], name);
2014 add_to_string(cmd_args_bench, strbuf);
2017 if (is_launch_file(opt, opt2bSet(opt, nfile, fnm)) )
2019 add_to_string(cmd_args_launch, strbuf);
2023 add_to_string(cmd_args_bench, extra_args);
2024 add_to_string(cmd_args_launch, extra_args);
2028 /* Set option opt */
2029 static void setopt(const char *opt, int nfile, t_filenm fnm[])
2033 for (i = 0; (i < nfile); i++)
2035 if (std::strcmp(opt, fnm[i].opt) == 0)
2037 fnm[i].flag |= ffSET;
2043 /* This routine inspects the tpr file and ...
2044 * 1. checks for output files that get triggered by a tpr option. These output
2045 * files are marked as 'set' to allow for a proper cleanup after each
2047 * 2. returns the PME:PP load ratio
2048 * 3. returns rcoulomb from the tpr */
2049 static float inspect_tpr(int nfile, t_filenm fnm[], real *rcoulomb)
2051 gmx_bool bTpi; /* Is test particle insertion requested? */
2052 gmx_bool bFree; /* Is a free energy simulation requested? */
2053 gmx_bool bNM; /* Is a normal mode analysis requested? */
2054 gmx_bool bSwap; /* Is water/ion position swapping requested? */
2060 /* Check tpr file for options that trigger extra output files */
2061 read_tpx_state(opt2fn("-s", nfile, fnm), &ir, &state, NULL, &mtop);
2062 bFree = (efepNO != ir.efep );
2063 bNM = (eiNM == ir.eI );
2064 bSwap = (eswapNO != ir.eSwapCoords);
2065 bTpi = EI_TPI(ir.eI);
2067 /* Set these output files on the tuning command-line */
2070 setopt("-pf", nfile, fnm);
2071 setopt("-px", nfile, fnm);
2075 setopt("-dhdl", nfile, fnm);
2079 setopt("-tpi", nfile, fnm);
2080 setopt("-tpid", nfile, fnm);
2084 setopt("-mtx", nfile, fnm);
2088 setopt("-swap", nfile, fnm);
2091 *rcoulomb = ir.rcoulomb;
2093 /* Return the estimate for the number of PME nodes */
2094 return pme_load_estimate(&mtop, &ir, state.box);
2098 static void couple_files_options(int nfile, t_filenm fnm[])
2101 gmx_bool bSet, bBench;
2106 for (i = 0; i < nfile; i++)
2108 opt = (char *)fnm[i].opt;
2109 bSet = ((fnm[i].flag & ffSET) != 0);
2110 bBench = (0 == std::strncmp(opt, "-b", 2));
2112 /* Check optional files */
2113 /* If e.g. -eo is set, then -beo also needs to be set */
2114 if (is_optional(&fnm[i]) && bSet && !bBench)
2116 sprintf(buf, "-b%s", &opt[1]);
2117 setopt(buf, nfile, fnm);
2119 /* If -beo is set, then -eo also needs to be! */
2120 if (is_optional(&fnm[i]) && bSet && bBench)
2122 sprintf(buf, "-%s", &opt[2]);
2123 setopt(buf, nfile, fnm);
2129 #define BENCHSTEPS (1000)
2131 int gmx_tune_pme(int argc, char *argv[])
2133 const char *desc[] = {
2134 "For a given number [TT]-np[tt] or [TT]-ntmpi[tt] of ranks, [THISMODULE] systematically",
2135 "times [gmx-mdrun] with various numbers of PME-only ranks and determines",
2136 "which setting is fastest. It will also test whether performance can",
2137 "be enhanced by shifting load from the reciprocal to the real space",
2138 "part of the Ewald sum. ",
2139 "Simply pass your [REF].tpr[ref] file to [THISMODULE] together with other options",
2140 "for [gmx-mdrun] as needed.[PAR]",
2141 "Which executables are used can be set in the environment variables",
2142 "MPIRUN and MDRUN. If these are not present, 'mpirun' and 'mdrun'",
2143 "will be used as defaults. Note that for certain MPI frameworks you",
2144 "need to provide a machine- or hostfile. This can also be passed",
2145 "via the MPIRUN variable, e.g.[PAR]",
2146 "[TT]export MPIRUN=\"/usr/local/mpirun -machinefile hosts\"[tt][PAR]",
2147 "Before doing the actual benchmark runs, [THISMODULE] will do a quick",
2148 "check whether mdrun works as expected with the provided parallel settings",
2149 "if the [TT]-check[tt] option is activated (the default).",
2150 "Please call [THISMODULE] with the normal options you would pass to",
2151 "[gmx-mdrun] and add [TT]-np[tt] for the number of ranks to perform the",
2152 "tests on, or [TT]-ntmpi[tt] for the number of threads. You can also add [TT]-r[tt]",
2153 "to repeat each test several times to get better statistics. [PAR]",
2154 "[THISMODULE] can test various real space / reciprocal space workloads",
2155 "for you. With [TT]-ntpr[tt] you control how many extra [REF].tpr[ref] files will be",
2156 "written with enlarged cutoffs and smaller Fourier grids respectively.",
2157 "Typically, the first test (number 0) will be with the settings from the input",
2158 "[REF].tpr[ref] file; the last test (number [TT]ntpr[tt]) will have the Coulomb cutoff",
2159 "specified by [TT]-rmax[tt] with a somewhat smaller PME grid at the same time. ",
2160 "In this last test, the Fourier spacing is multiplied with [TT]rmax[tt]/rcoulomb. ",
2161 "The remaining [REF].tpr[ref] files will have equally-spaced Coulomb radii (and Fourier "
2162 "spacings) between these extremes. [BB]Note[bb] that you can set [TT]-ntpr[tt] to 1",
2163 "if you just seek the optimal number of PME-only ranks; in that case",
2164 "your input [REF].tpr[ref] file will remain unchanged.[PAR]",
2165 "For the benchmark runs, the default of 1000 time steps should suffice for most",
2166 "MD systems. The dynamic load balancing needs about 100 time steps",
2167 "to adapt to local load imbalances, therefore the time step counters",
2168 "are by default reset after 100 steps. For large systems (>1M atoms), as well as ",
2169 "for a higher accuracy of the measurements, you should set [TT]-resetstep[tt] to a higher value.",
2170 "From the 'DD' load imbalance entries in the md.log output file you",
2171 "can tell after how many steps the load is sufficiently balanced. Example call:[PAR]"
2172 "[TT]gmx tune_pme -np 64 -s protein.tpr -launch[tt][PAR]",
2173 "After calling [gmx-mdrun] several times, detailed performance information",
2174 "is available in the output file [TT]perf.out[tt].",
2175 "[BB]Note[bb] that during the benchmarks, a couple of temporary files are written",
2176 "(options [TT]-b*[tt]), these will be automatically deleted after each test.[PAR]",
2177 "If you want the simulation to be started automatically with the",
2178 "optimized parameters, use the command line option [TT]-launch[tt].[PAR]",
2179 "Basic support for GPU-enabled [TT]mdrun[tt] exists. Give a string containing the IDs",
2180 "of the GPUs that you wish to use in the optimization in the [TT]-gpu_id[tt]",
2181 "command-line argument. Unlike [TT]mdrun -gpu_id[tt], this does not imply a mapping",
2182 "but merely the eligible set. [TT]g_tune_pme[tt] will construct calls to",
2183 "mdrun that use this set appropriately, assuming that PP ranks with low indices",
2184 "should map to GPUs with low indices, and increasing both monotonically",
2185 "over the respective sets.[PAR]",
2190 int pmeentries = 0; /* How many values for -npme do we actually test for each tpr file */
2191 real maxPMEfraction = 0.50;
2192 real minPMEfraction = 0.25;
2193 int maxPMEnodes, minPMEnodes;
2194 float guessPMEratio; /* guessed PME:PP ratio based on the tpr file */
2195 float guessPMEnodes;
2196 int npme_fixed = -2; /* If >= -1, use only this number
2197 * of PME-only nodes */
2199 real rmin = 0.0, rmax = 0.0; /* min and max value for rcoulomb if scaling is requested */
2200 real rcoulomb = -1.0; /* Coulomb radius as set in .tpr file */
2201 gmx_bool bScaleRvdw = TRUE;
2202 gmx_int64_t bench_nsteps = BENCHSTEPS;
2203 gmx_int64_t new_sim_nsteps = -1; /* -1 indicates: not set by the user */
2204 gmx_int64_t cpt_steps = 0; /* Step counter in .cpt input file */
2205 int presteps = 100; /* Do a full cycle reset after presteps steps */
2206 gmx_bool bOverwrite = FALSE, bKeepTPR;
2207 gmx_bool bLaunch = FALSE;
2208 char *ExtraArgs = NULL;
2209 char **tpr_names = NULL;
2210 const char *simulation_tpr = NULL;
2211 int best_npme, best_tpr;
2212 int sim_part = 1; /* For benchmarks with checkpoint files */
2216 /* Default program names if nothing else is found */
2217 char *cmd_mpirun = NULL, *cmd_mdrun = NULL;
2218 char *cmd_args_bench, *cmd_args_launch;
2219 char *cmd_np = NULL;
2221 /* IDs of GPUs that are eligible for computation */
2222 char *eligible_gpu_ids = NULL;
2223 t_eligible_gpu_ids *gpu_ids = NULL;
2225 t_perf **perfdata = NULL;
2230 /* Print out how long the tuning took */
2233 static t_filenm fnm[] = {
2235 { efOUT, "-p", "perf", ffWRITE },
2236 { efLOG, "-err", "bencherr", ffWRITE },
2237 { efTPR, "-so", "tuned", ffWRITE },
2239 { efTPR, NULL, NULL, ffREAD },
2240 { efTRN, "-o", NULL, ffWRITE },
2241 { efCOMPRESSED, "-x", NULL, ffOPTWR },
2242 { efCPT, "-cpi", NULL, ffOPTRD },
2243 { efCPT, "-cpo", NULL, ffOPTWR },
2244 { efSTO, "-c", "confout", ffWRITE },
2245 { efEDR, "-e", "ener", ffWRITE },
2246 { efLOG, "-g", "md", ffWRITE },
2247 { efXVG, "-dhdl", "dhdl", ffOPTWR },
2248 { efXVG, "-field", "field", ffOPTWR },
2249 { efXVG, "-table", "table", ffOPTRD },
2250 { efXVG, "-tabletf", "tabletf", ffOPTRD },
2251 { efXVG, "-tablep", "tablep", ffOPTRD },
2252 { efXVG, "-tableb", "table", ffOPTRD },
2253 { efTRX, "-rerun", "rerun", ffOPTRD },
2254 { efXVG, "-tpi", "tpi", ffOPTWR },
2255 { efXVG, "-tpid", "tpidist", ffOPTWR },
2256 { efEDI, "-ei", "sam", ffOPTRD },
2257 { efXVG, "-eo", "edsam", ffOPTWR },
2258 { efXVG, "-devout", "deviatie", ffOPTWR },
2259 { efXVG, "-runav", "runaver", ffOPTWR },
2260 { efXVG, "-px", "pullx", ffOPTWR },
2261 { efXVG, "-pf", "pullf", ffOPTWR },
2262 { efXVG, "-ro", "rotation", ffOPTWR },
2263 { efLOG, "-ra", "rotangles", ffOPTWR },
2264 { efLOG, "-rs", "rotslabs", ffOPTWR },
2265 { efLOG, "-rt", "rottorque", ffOPTWR },
2266 { efMTX, "-mtx", "nm", ffOPTWR },
2267 { efNDX, "-dn", "dipole", ffOPTWR },
2268 { efXVG, "-swap", "swapions", ffOPTWR },
2269 /* Output files that are deleted after each benchmark run */
2270 { efTRN, "-bo", "bench", ffWRITE },
2271 { efXTC, "-bx", "bench", ffWRITE },
2272 { efCPT, "-bcpo", "bench", ffWRITE },
2273 { efSTO, "-bc", "bench", ffWRITE },
2274 { efEDR, "-be", "bench", ffWRITE },
2275 { efLOG, "-bg", "bench", ffWRITE },
2276 { efXVG, "-beo", "benchedo", ffOPTWR },
2277 { efXVG, "-bdhdl", "benchdhdl", ffOPTWR },
2278 { efXVG, "-bfield", "benchfld", ffOPTWR },
2279 { efXVG, "-btpi", "benchtpi", ffOPTWR },
2280 { efXVG, "-btpid", "benchtpid", ffOPTWR },
2281 { efXVG, "-bdevout", "benchdev", ffOPTWR },
2282 { efXVG, "-brunav", "benchrnav", ffOPTWR },
2283 { efXVG, "-bpx", "benchpx", ffOPTWR },
2284 { efXVG, "-bpf", "benchpf", ffOPTWR },
2285 { efXVG, "-bro", "benchrot", ffOPTWR },
2286 { efLOG, "-bra", "benchrota", ffOPTWR },
2287 { efLOG, "-brs", "benchrots", ffOPTWR },
2288 { efLOG, "-brt", "benchrott", ffOPTWR },
2289 { efMTX, "-bmtx", "benchn", ffOPTWR },
2290 { efNDX, "-bdn", "bench", ffOPTWR },
2291 { efXVG, "-bswap", "benchswp", ffOPTWR }
2294 gmx_bool bThreads = FALSE;
2298 const char *procstring[] =
2299 { NULL, "-np", "-n", "none", NULL };
2300 const char *npmevalues_opt[] =
2301 { NULL, "auto", "all", "subset", NULL };
2303 gmx_bool bAppendFiles = TRUE;
2304 gmx_bool bKeepAndNumCPT = FALSE;
2305 gmx_bool bResetCountersHalfWay = FALSE;
2306 gmx_bool bBenchmark = TRUE;
2307 gmx_bool bCheck = TRUE;
2309 output_env_t oenv = NULL;
2312 /***********************/
2313 /* g_tune_pme options: */
2314 /***********************/
2315 { "-np", FALSE, etINT, {&nnodes},
2316 "Number of ranks to run the tests on (must be > 2 for separate PME ranks)" },
2317 { "-npstring", FALSE, etENUM, {procstring},
2318 "Specify the number of ranks to [TT]$MPIRUN[tt] using this string"},
2319 { "-ntmpi", FALSE, etINT, {&nthreads},
2320 "Number of MPI-threads to run the tests on (turns MPI & mpirun off)"},
2321 { "-r", FALSE, etINT, {&repeats},
2322 "Repeat each test this often" },
2323 { "-max", FALSE, etREAL, {&maxPMEfraction},
2324 "Max fraction of PME ranks to test with" },
2325 { "-min", FALSE, etREAL, {&minPMEfraction},
2326 "Min fraction of PME ranks to test with" },
2327 { "-npme", FALSE, etENUM, {npmevalues_opt},
2328 "Within -min and -max, benchmark all possible values for [TT]-npme[tt], or just a reasonable subset. "
2329 "Auto neglects -min and -max and chooses reasonable values around a guess for npme derived from the .tpr"},
2330 { "-fix", FALSE, etINT, {&npme_fixed},
2331 "If >= -1, do not vary the number of PME-only ranks, instead use this fixed value and only vary rcoulomb and the PME grid spacing."},
2332 { "-rmax", FALSE, etREAL, {&rmax},
2333 "If >0, maximal rcoulomb for -ntpr>1 (rcoulomb upscaling results in fourier grid downscaling)" },
2334 { "-rmin", FALSE, etREAL, {&rmin},
2335 "If >0, minimal rcoulomb for -ntpr>1" },
2336 { "-scalevdw", FALSE, etBOOL, {&bScaleRvdw},
2337 "Scale rvdw along with rcoulomb"},
2338 { "-ntpr", FALSE, etINT, {&ntprs},
2339 "Number of [REF].tpr[ref] files to benchmark. Create this many files with different rcoulomb scaling factors depending on -rmin and -rmax. "
2340 "If < 1, automatically choose the number of [REF].tpr[ref] files to test" },
2341 { "-steps", FALSE, etINT64, {&bench_nsteps},
2342 "Take timings for this many steps in the benchmark runs" },
2343 { "-resetstep", FALSE, etINT, {&presteps},
2344 "Let dlb equilibrate this many steps before timings are taken (reset cycle counters after this many steps)" },
2345 { "-nsteps", FALSE, etINT64, {&new_sim_nsteps},
2346 "If non-negative, perform this many steps in the real run (overwrites nsteps from [REF].tpr[ref], add [REF].cpt[ref] steps)" },
2347 { "-launch", FALSE, etBOOL, {&bLaunch},
2348 "Launch the real simulation after optimization" },
2349 { "-bench", FALSE, etBOOL, {&bBenchmark},
2350 "Run the benchmarks or just create the input [REF].tpr[ref] files?" },
2351 { "-check", FALSE, etBOOL, {&bCheck},
2352 "Before the benchmark runs, check whether mdrun works in parallel" },
2353 { "-gpu_id", FALSE, etSTR, {&eligible_gpu_ids},
2354 "List of GPU device id-s that are eligible for use (unlike mdrun, does not imply any mapping)" },
2355 /******************/
2356 /* mdrun options: */
2357 /******************/
2358 /* We let g_tune_pme parse and understand these options, because we need to
2359 * prevent that they appear on the mdrun command line for the benchmarks */
2360 { "-append", FALSE, etBOOL, {&bAppendFiles},
2361 "Append to previous output files when continuing from checkpoint instead of adding the simulation part number to all file names (for launch only)" },
2362 { "-cpnum", FALSE, etBOOL, {&bKeepAndNumCPT},
2363 "Keep and number checkpoint files (launch only)" },
2364 { "-resethway", FALSE, etBOOL, {&bResetCountersHalfWay},
2365 "HIDDENReset the cycle counters after half the number of steps or halfway [TT]-maxh[tt] (launch only)" }
2368 #define NFILE asize(fnm)
2370 seconds = gmx_gettime();
2372 if (!parse_common_args(&argc, argv, PCA_NOEXIT_ON_ARGS,
2373 NFILE, fnm, asize(pa), pa, asize(desc), desc,
2379 // procstring[0]Â is used inside two different conditionals further down
2380 GMX_RELEASE_ASSERT(procstring[0] != NULL, "Options inconsistency; procstring[0]Â is NULL");
2382 /* Store the remaining unparsed command line entries in a string which
2383 * is then attached to the mdrun command line */
2385 ExtraArgs[0] = '\0';
2386 for (i = 1; i < argc; i++) /* argc will now be 1 if everything was understood */
2388 add_to_string(&ExtraArgs, argv[i]);
2389 add_to_string(&ExtraArgs, " ");
2392 if (opt2parg_bSet("-ntmpi", asize(pa), pa))
2395 if (opt2parg_bSet("-npstring", asize(pa), pa))
2397 fprintf(stderr, "WARNING: -npstring has no effect when using threads.\n");
2402 gmx_fatal(FARGS, "Can't run multi-threaded MPI simulation yet!");
2404 /* and now we just set this; a bit of an ugly hack*/
2407 /* Check for PME:PP ratio and whether tpr triggers additional output files */
2408 guessPMEratio = inspect_tpr(NFILE, fnm, &rcoulomb);
2410 /* Automatically set -beo options if -eo is set etc. */
2411 couple_files_options(NFILE, fnm);
2413 /* Construct the command line arguments for benchmark runs
2414 * as well as for the simulation run */
2417 sprintf(bbuf, " -ntmpi %d ", nthreads);
2421 /* This string will be used for MPI runs and will appear after the
2422 * mpirun command. */
2423 if (std::strcmp(procstring[0], "none") != 0)
2425 sprintf(bbuf, " %s %d ", procstring[0], nnodes);
2435 create_command_line_snippets(bAppendFiles, bKeepAndNumCPT, bResetCountersHalfWay, presteps,
2436 NFILE, fnm, &cmd_args_bench, &cmd_args_launch, ExtraArgs);
2438 /* Prepare to use checkpoint file if requested */
2440 if (opt2bSet("-cpi", NFILE, fnm))
2442 const char *filename = opt2fn("-cpi", NFILE, fnm);
2444 read_checkpoint_part_and_step(filename,
2445 &cpt_sim_part, &cpt_steps);
2446 if (cpt_sim_part == 0)
2448 gmx_fatal(FARGS, "Checkpoint file %s could not be read!", filename);
2450 /* tune_pme will run the next part of the simulation */
2451 sim_part = cpt_sim_part + 1;
2454 /* Open performance output file and write header info */
2455 fp = gmx_ffopen(opt2fn("-p", NFILE, fnm), "w");
2457 /* Make a quick consistency check of command line parameters */
2458 check_input(nnodes, repeats, &ntprs, &rmin, rcoulomb, &rmax,
2459 maxPMEfraction, minPMEfraction, npme_fixed,
2460 bench_nsteps, fnm, NFILE, sim_part, presteps,
2462 /* Check any GPU IDs passed make sense, and fill the data structure for them */
2464 parse_digits_from_plain_string(eligible_gpu_ids, &gpu_ids->n, &gpu_ids->ids);
2466 /* Determine the maximum and minimum number of PME nodes to test,
2467 * the actual list of settings is build in do_the_tests(). */
2468 if ((nnodes > 2) && (npme_fixed < -1))
2470 if (0 == std::strcmp(npmevalues_opt[0], "auto"))
2472 /* Determine the npme range automatically based on the PME:PP load guess */
2473 if (guessPMEratio > 1.0)
2475 /* More PME than PP work, probably we do not need separate PME nodes at all! */
2476 maxPMEnodes = nnodes/2;
2477 minPMEnodes = nnodes/2;
2481 /* PME : PP load is in the range 0..1, let's test around the guess */
2482 guessPMEnodes = static_cast<int>(nnodes/(1.0 + 1.0/guessPMEratio));
2483 minPMEnodes = static_cast<int>(std::floor(0.7*guessPMEnodes));
2484 maxPMEnodes = static_cast<int>(std::ceil(1.6*guessPMEnodes));
2485 maxPMEnodes = std::min(maxPMEnodes, nnodes/2);
2490 /* Determine the npme range based on user input */
2491 maxPMEnodes = static_cast<int>(std::floor(maxPMEfraction*nnodes));
2492 minPMEnodes = std::max(static_cast<int>(std::floor(minPMEfraction*nnodes)), 0);
2493 fprintf(stdout, "Will try runs with %d ", minPMEnodes);
2494 if (maxPMEnodes != minPMEnodes)
2496 fprintf(stdout, "- %d ", maxPMEnodes);
2498 fprintf(stdout, "PME-only ranks.\n Note that the automatic number of PME-only ranks and no separate PME ranks are always tested.\n");
2507 /* Get the commands we need to set up the runs from environment variables */
2508 get_program_paths(bThreads, &cmd_mpirun, &cmd_mdrun);
2509 if (bBenchmark && repeats > 0)
2511 check_mdrun_works(bThreads, cmd_mpirun, cmd_np, cmd_mdrun, NULL != eligible_gpu_ids);
2514 /* Print some header info to file */
2516 fprintf(fp, "\n P E R F O R M A N C E R E S U L T S\n");
2518 fprintf(fp, "%s for GROMACS %s\n", output_env_get_program_display_name(oenv),
2522 fprintf(fp, "Number of ranks : %d\n", nnodes);
2523 fprintf(fp, "The mpirun command is : %s\n", cmd_mpirun);
2524 if (std::strcmp(procstring[0], "none") != 0)
2526 fprintf(fp, "Passing # of ranks via : %s\n", procstring[0]);
2530 fprintf(fp, "Not setting number of ranks in system call\n");
2535 fprintf(fp, "Number of threads : %d\n", nnodes);
2538 fprintf(fp, "The mdrun command is : %s\n", cmd_mdrun);
2539 fprintf(fp, "mdrun args benchmarks : %s\n", cmd_args_bench);
2540 fprintf(fp, "Benchmark steps : ");
2541 fprintf(fp, "%" GMX_PRId64, bench_nsteps);
2543 fprintf(fp, "dlb equilibration steps : %d\n", presteps);
2546 fprintf(fp, "Checkpoint time step : ");
2547 fprintf(fp, "%" GMX_PRId64, cpt_steps);
2550 fprintf(fp, "mdrun args at launchtime: %s\n", cmd_args_launch);
2552 if (new_sim_nsteps >= 0)
2555 fprintf(stderr, "Note: Simulation input file %s will have ", opt2fn("-so", NFILE, fnm));
2556 fprintf(stderr, "%" GMX_PRId64, new_sim_nsteps+cpt_steps);
2557 fprintf(stderr, " steps.\n");
2558 fprintf(fp, "Simulation steps : ");
2559 fprintf(fp, "%" GMX_PRId64, new_sim_nsteps);
2564 fprintf(fp, "Repeats for each test : %d\n", repeats);
2567 if (npme_fixed >= -1)
2569 fprintf(fp, "Fixing -npme at : %d\n", npme_fixed);
2572 fprintf(fp, "Input file : %s\n", opt2fn("-s", NFILE, fnm));
2573 fprintf(fp, " PME/PP load estimate : %g\n", guessPMEratio);
2575 /* Allocate memory for the inputinfo struct: */
2577 info->nr_inputfiles = ntprs;
2578 for (i = 0; i < ntprs; i++)
2580 snew(info->rcoulomb, ntprs);
2581 snew(info->rvdw, ntprs);
2582 snew(info->rlist, ntprs);
2583 snew(info->rlistlong, ntprs);
2584 snew(info->nkx, ntprs);
2585 snew(info->nky, ntprs);
2586 snew(info->nkz, ntprs);
2587 snew(info->fsx, ntprs);
2588 snew(info->fsy, ntprs);
2589 snew(info->fsz, ntprs);
2591 /* Make alternative tpr files to test: */
2592 snew(tpr_names, ntprs);
2593 for (i = 0; i < ntprs; i++)
2595 snew(tpr_names[i], STRLEN);
2598 /* It can be that ntprs is reduced by make_benchmark_tprs if not enough
2599 * different grids could be found. */
2600 make_benchmark_tprs(opt2fn("-s", NFILE, fnm), tpr_names, bench_nsteps+presteps,
2601 cpt_steps, rmin, rmax, bScaleRvdw, &ntprs, info, fp);
2603 /********************************************************************************/
2604 /* Main loop over all scenarios we need to test: tpr files, PME nodes, repeats */
2605 /********************************************************************************/
2606 snew(perfdata, ntprs);
2609 GMX_RELEASE_ASSERT(npmevalues_opt[0] != NULL, "Options inconsistency; npmevalues_opt[0] is NULL");
2610 do_the_tests(fp, tpr_names, maxPMEnodes, minPMEnodes, npme_fixed, npmevalues_opt[0], perfdata, &pmeentries,
2611 repeats, nnodes, ntprs, bThreads, cmd_mpirun, cmd_np, cmd_mdrun,
2612 cmd_args_bench, fnm, NFILE, presteps, cpt_steps, bCheck, gpu_ids);
2614 fprintf(fp, "\nTuning took%8.1f minutes.\n", (gmx_gettime()-seconds)/60.0);
2616 /* Analyse the results and give a suggestion for optimal settings: */
2617 bKeepTPR = analyze_data(fp, opt2fn("-p", NFILE, fnm), perfdata, nnodes, ntprs, pmeentries,
2618 repeats, info, &best_tpr, &best_npme);
2620 /* Take the best-performing tpr file and enlarge nsteps to original value */
2621 if (bKeepTPR && !bOverwrite)
2623 simulation_tpr = opt2fn("-s", NFILE, fnm);
2627 simulation_tpr = opt2fn("-so", NFILE, fnm);
2628 modify_PMEsettings(bOverwrite ? (new_sim_nsteps+cpt_steps) : info->orig_sim_steps,
2629 info->orig_init_step, tpr_names[best_tpr], simulation_tpr);
2632 /* Let's get rid of the temporary benchmark input files */
2633 for (i = 0; i < ntprs; i++)
2635 fprintf(stdout, "Deleting temporary benchmark input file %s\n", tpr_names[i]);
2636 remove(tpr_names[i]);
2639 /* Now start the real simulation if the user requested it ... */
2640 launch_simulation(bLaunch, fp, bThreads, cmd_mpirun, cmd_np, cmd_mdrun,
2641 cmd_args_launch, simulation_tpr, nnodes, best_npme, gpu_ids);
2645 /* ... or simply print the performance results to screen: */
2648 finalize(opt2fn("-p", NFILE, fnm));