2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2009,2010,2011,2012,2013,2014, 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.
40 #ifdef HAVE_SYS_TIME_H
44 #include "gromacs/commandline/pargs.h"
46 #include "types/commrec.h"
47 #include "gromacs/utility/smalloc.h"
50 #include "gromacs/fileio/tpxio.h"
51 #include "gromacs/utility/cstringutil.h"
54 #include "checkpoint.h"
60 #include "gromacs/timing/walltime_accounting.h"
61 #include "gromacs/math/utilities.h"
63 #include "gmx_fatal.h"
65 /* Enum for situations that can occur during log file parsing, the
66 * corresponding string entries can be found in do_the_tests() in
67 * const char* ParseLog[] */
73 eParselogResetProblem,
77 eParselogLargePrimeFactor,
85 int nPMEnodes; /* number of PME-only nodes used in this test */
86 int nx, ny, nz; /* DD grid */
87 int guessPME; /* if nPMEnodes == -1, this is the guessed number of PME nodes */
88 double *Gcycles; /* This can contain more than one value if doing multiple tests */
92 float *PME_f_load; /* PME mesh/force load average*/
93 float PME_f_load_Av; /* Average average ;) ... */
94 char *mdrun_cmd_line; /* Mdrun command line used for this test */
100 int nr_inputfiles; /* The number of tpr and mdp input files */
101 gmx_int64_t orig_sim_steps; /* Number of steps to be done in the real simulation */
102 gmx_int64_t orig_init_step; /* Init step for the real simulation */
103 real *rcoulomb; /* The coulomb radii [0...nr_inputfiles] */
104 real *rvdw; /* The vdW radii */
105 real *rlist; /* Neighbourlist cutoff radius */
107 int *nkx, *nky, *nkz;
108 real *fsx, *fsy, *fsz; /* Fourierspacing in x,y,z dimension */
112 static void sep_line(FILE *fp)
114 fprintf(fp, "\n------------------------------------------------------------\n");
118 /* Wrapper for system calls */
119 static int gmx_system_call(char *command)
122 gmx_fatal(FARGS, "No calls to system(3) supported on this platform. Attempted to call:\n'%s'\n", command);
124 return ( system(command) );
129 /* Check if string starts with substring */
130 static gmx_bool str_starts(const char *string, const char *substring)
132 return ( strncmp(string, substring, strlen(substring)) == 0);
136 static void cleandata(t_perf *perfdata, int test_nr)
138 perfdata->Gcycles[test_nr] = 0.0;
139 perfdata->ns_per_day[test_nr] = 0.0;
140 perfdata->PME_f_load[test_nr] = 0.0;
146 static gmx_bool is_equal(real a, real b)
148 real diff, eps = 1.0e-7;
169 static void remove_if_exists(const char *fn)
173 fprintf(stdout, "Deleting %s\n", fn);
179 static void finalize(const char *fn_out)
185 fp = fopen(fn_out, "r");
186 fprintf(stdout, "\n\n");
188 while (fgets(buf, STRLEN-1, fp) != NULL)
190 fprintf(stdout, "%s", buf);
193 fprintf(stdout, "\n\n");
198 eFoundNothing, eFoundDDStr, eFoundAccountingStr, eFoundCycleStr
201 static int parse_logfile(const char *logfile, const char *errfile,
202 t_perf *perfdata, int test_nr, int presteps, gmx_int64_t cpt_steps,
206 char line[STRLEN], dumstring[STRLEN], dumstring2[STRLEN];
207 const char matchstrdd[] = "Domain decomposition grid";
208 const char matchstrcr[] = "resetting all time and cycle counters";
209 const char matchstrbal[] = "Average PME mesh/force load:";
210 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";
211 const char errSIG[] = "signal, stopping at the next";
213 float dum1, dum2, dum3, dum4;
216 gmx_int64_t resetsteps = -1;
217 gmx_bool bFoundResetStr = FALSE;
218 gmx_bool bResetChecked = FALSE;
221 if (!gmx_fexist(logfile))
223 fprintf(stderr, "WARNING: Could not find logfile %s.\n", logfile);
224 cleandata(perfdata, test_nr);
225 return eParselogNotFound;
228 fp = fopen(logfile, "r");
229 perfdata->PME_f_load[test_nr] = -1.0;
230 perfdata->guessPME = -1;
232 iFound = eFoundNothing;
235 iFound = eFoundDDStr; /* Skip some case statements */
238 while (fgets(line, STRLEN, fp) != NULL)
240 /* Remove leading spaces */
243 /* Check for TERM and INT signals from user: */
244 if (strstr(line, errSIG) != NULL)
247 cleandata(perfdata, test_nr);
248 return eParselogTerm;
251 /* Check whether cycle resetting worked */
252 if (presteps > 0 && !bFoundResetStr)
254 if (strstr(line, matchstrcr) != NULL)
256 sprintf(dumstring, "step %s", "%"GMX_SCNd64);
257 sscanf(line, dumstring, &resetsteps);
258 bFoundResetStr = TRUE;
259 if (resetsteps == presteps+cpt_steps)
261 bResetChecked = TRUE;
265 sprintf(dumstring, "%"GMX_PRId64, resetsteps);
266 sprintf(dumstring2, "%"GMX_PRId64, presteps+cpt_steps);
267 fprintf(stderr, "WARNING: Time step counters were reset at step %s,\n"
268 " though they were supposed to be reset at step %s!\n",
269 dumstring, dumstring2);
274 /* Look for strings that appear in a certain order in the log file: */
278 /* Look for domain decomp grid and separate PME nodes: */
279 if (str_starts(line, matchstrdd))
281 sscanf(line, "Domain decomposition grid %d x %d x %d, separate PME ranks %d",
282 &(perfdata->nx), &(perfdata->ny), &(perfdata->nz), &npme);
283 if (perfdata->nPMEnodes == -1)
285 perfdata->guessPME = npme;
287 else if (perfdata->nPMEnodes != npme)
289 gmx_fatal(FARGS, "PME ranks from command line and output file are not identical");
291 iFound = eFoundDDStr;
293 /* Catch a few errors that might have occured: */
294 else if (str_starts(line, "There is no domain decomposition for"))
297 return eParselogNoDDGrid;
299 else if (str_starts(line, "The number of ranks you selected"))
302 return eParselogLargePrimeFactor;
304 else if (str_starts(line, "reading tpx file"))
307 return eParselogTPXVersion;
309 else if (str_starts(line, "The -dd or -npme option request a parallel simulation"))
312 return eParselogNotParallel;
316 /* Look for PME mesh/force balance (not necessarily present, though) */
317 if (str_starts(line, matchstrbal))
319 sscanf(&line[strlen(matchstrbal)], "%f", &(perfdata->PME_f_load[test_nr]));
321 /* Look for matchstring */
322 if (str_starts(line, matchstring))
324 iFound = eFoundAccountingStr;
327 case eFoundAccountingStr:
328 /* Already found matchstring - look for cycle data */
329 if (str_starts(line, "Total "))
331 sscanf(line, "Total %*f %lf", &(perfdata->Gcycles[test_nr]));
332 iFound = eFoundCycleStr;
336 /* Already found cycle data - look for remaining performance info and return */
337 if (str_starts(line, "Performance:"))
339 ndum = sscanf(line, "%s %f %f %f %f", dumstring, &dum1, &dum2, &dum3, &dum4);
340 /* (ns/day) is the second last entry, depending on whether GMX_DETAILED_PERF_STATS was set in print_perf(), nrnb.c */
341 perfdata->ns_per_day[test_nr] = (ndum == 5) ? dum3 : dum1;
343 if (bResetChecked || presteps == 0)
349 return eParselogResetProblem;
356 /* Close the log file */
359 /* Check why there is no performance data in the log file.
360 * Did a fatal errors occur? */
361 if (gmx_fexist(errfile))
363 fp = fopen(errfile, "r");
364 while (fgets(line, STRLEN, fp) != NULL)
366 if (str_starts(line, "Fatal error:") )
368 if (fgets(line, STRLEN, fp) != NULL)
370 fprintf(stderr, "\nWARNING: An error occured during this benchmark:\n"
374 cleandata(perfdata, test_nr);
375 return eParselogFatal;
382 fprintf(stderr, "WARNING: Could not find stderr file %s.\n", errfile);
385 /* Giving up ... we could not find out why there is no performance data in
387 fprintf(stdout, "No performance data in log file.\n");
388 cleandata(perfdata, test_nr);
390 return eParselogNoPerfData;
394 static gmx_bool analyze_data(
403 int *index_tpr, /* OUT: Nr of mdp file with best settings */
404 int *npme_optimal) /* OUT: Optimal number of PME nodes */
407 int line = 0, line_win = -1;
408 int k_win = -1, i_win = -1, winPME;
409 double s = 0.0; /* standard deviation */
412 char str_PME_f_load[13];
413 gmx_bool bCanUseOrigTPR;
414 gmx_bool bRefinedCoul, bRefinedVdW, bRefinedGrid;
420 fprintf(fp, "Summary of successful runs:\n");
421 fprintf(fp, "Line tpr PME ranks Gcycles Av. Std.dev. ns/day PME/f");
424 fprintf(fp, " DD grid");
430 for (k = 0; k < ntprs; k++)
432 for (i = 0; i < ntests; i++)
434 /* Select the right dataset: */
435 pd = &(perfdata[k][i]);
437 pd->Gcycles_Av = 0.0;
438 pd->PME_f_load_Av = 0.0;
439 pd->ns_per_day_Av = 0.0;
441 if (pd->nPMEnodes == -1)
443 sprintf(strbuf, "(%3d)", pd->guessPME);
447 sprintf(strbuf, " ");
450 /* Get the average run time of a setting */
451 for (j = 0; j < nrepeats; j++)
453 pd->Gcycles_Av += pd->Gcycles[j];
454 pd->PME_f_load_Av += pd->PME_f_load[j];
456 pd->Gcycles_Av /= nrepeats;
457 pd->PME_f_load_Av /= nrepeats;
459 for (j = 0; j < nrepeats; j++)
461 if (pd->ns_per_day[j] > 0.0)
463 pd->ns_per_day_Av += pd->ns_per_day[j];
467 /* Somehow the performance number was not aquired for this run,
468 * therefor set the average to some negative value: */
469 pd->ns_per_day_Av = -1.0f*nrepeats;
473 pd->ns_per_day_Av /= nrepeats;
476 if (pd->PME_f_load_Av > 0.0)
478 sprintf(str_PME_f_load, "%12.3f", pd->PME_f_load_Av);
482 sprintf(str_PME_f_load, "%s", " - ");
486 /* We assume we had a successful run if both averages are positive */
487 if (pd->Gcycles_Av > 0.0 && pd->ns_per_day_Av > 0.0)
489 /* Output statistics if repeats were done */
492 /* Calculate the standard deviation */
494 for (j = 0; j < nrepeats; j++)
496 s += pow( pd->Gcycles[j] - pd->Gcycles_Av, 2 );
501 fprintf(fp, "%4d %3d %4d%s %12.3f %12.3f %12.3f %s",
502 line, k, pd->nPMEnodes, strbuf, pd->Gcycles_Av, s,
503 pd->ns_per_day_Av, str_PME_f_load);
506 fprintf(fp, " %3d %3d %3d", pd->nx, pd->ny, pd->nz);
510 /* Store the index of the best run found so far in 'winner': */
511 if ( (k_win == -1) || (pd->Gcycles_Av < perfdata[k_win][i_win].Gcycles_Av) )
524 gmx_fatal(FARGS, "None of the runs was successful! Check %s for problems.", fn);
529 winPME = perfdata[k_win][i_win].nPMEnodes;
533 /* We stuck to a fixed number of PME-only nodes */
534 sprintf(strbuf, "settings No. %d", k_win);
538 /* We have optimized the number of PME-only nodes */
541 sprintf(strbuf, "%s", "the automatic number of PME ranks");
545 sprintf(strbuf, "%d PME ranks", winPME);
548 fprintf(fp, "Best performance was achieved with %s", strbuf);
549 if ((nrepeats > 1) && (ntests > 1))
551 fprintf(fp, " (see line %d)", line_win);
555 /* Only mention settings if they were modified: */
556 bRefinedCoul = !is_equal(info->rcoulomb[k_win], info->rcoulomb[0]);
557 bRefinedVdW = !is_equal(info->rvdw[k_win], info->rvdw[0] );
558 bRefinedGrid = !(info->nkx[k_win] == info->nkx[0] &&
559 info->nky[k_win] == info->nky[0] &&
560 info->nkz[k_win] == info->nkz[0]);
562 if (bRefinedCoul || bRefinedVdW || bRefinedGrid)
564 fprintf(fp, "Optimized PME settings:\n");
565 bCanUseOrigTPR = FALSE;
569 bCanUseOrigTPR = TRUE;
574 fprintf(fp, " New Coulomb radius: %f nm (was %f nm)\n", info->rcoulomb[k_win], info->rcoulomb[0]);
579 fprintf(fp, " New Van der Waals radius: %f nm (was %f nm)\n", info->rvdw[k_win], info->rvdw[0]);
584 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],
585 info->nkx[0], info->nky[0], info->nkz[0]);
588 if (bCanUseOrigTPR && ntprs > 1)
590 fprintf(fp, "and original PME settings.\n");
595 /* Return the index of the mdp file that showed the highest performance
596 * and the optimal number of PME nodes */
598 *npme_optimal = winPME;
600 return bCanUseOrigTPR;
604 /* Get the commands we need to set up the runs from environment variables */
605 static void get_program_paths(gmx_bool bThreads, char *cmd_mpirun[], char *cmd_mdrun[])
609 const char def_mpirun[] = "mpirun";
610 const char def_mdrun[] = "mdrun";
612 const char empty_mpirun[] = "";
614 /* Get the commands we need to set up the runs from environment variables */
617 if ( (cp = getenv("MPIRUN")) != NULL)
619 *cmd_mpirun = strdup(cp);
623 *cmd_mpirun = strdup(def_mpirun);
628 *cmd_mpirun = strdup(empty_mpirun);
631 if ( (cp = getenv("MDRUN" )) != NULL)
633 *cmd_mdrun = strdup(cp);
637 *cmd_mdrun = strdup(def_mdrun);
641 /* Check that the commands will run mdrun (perhaps via mpirun) by
642 * running a very quick test simulation. Requires MPI environment to
643 * be available if applicable. */
644 static void check_mdrun_works(gmx_bool bThreads,
645 const char *cmd_mpirun,
647 const char *cmd_mdrun)
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 gmx_bool bMdrun = FALSE;
660 gmx_bool bMPI = FALSE;
662 /* Run a small test to see whether mpirun + mdrun work */
663 fprintf(stdout, "Making sure that mdrun can be executed. ");
666 snew(command, strlen(cmd_mdrun) + strlen(cmd_np) + strlen(filename) + 50);
667 sprintf(command, "%s%s-version -maxh 0.001 1> %s 2>&1", cmd_mdrun, cmd_np, filename);
671 snew(command, strlen(cmd_mpirun) + strlen(cmd_np) + strlen(cmd_mdrun) + strlen(filename) + 50);
672 sprintf(command, "%s%s%s -version -maxh 0.001 1> %s 2>&1", cmd_mpirun, cmd_np, cmd_mdrun, filename);
674 fprintf(stdout, "Trying '%s' ... ", command);
675 make_backup(filename);
676 gmx_system_call(command);
678 /* Check if we find the characteristic string in the output: */
679 if (!gmx_fexist(filename))
681 gmx_fatal(FARGS, "Output from test run could not be found.");
684 fp = fopen(filename, "r");
685 /* We need to scan the whole output file, since sometimes the queuing system
686 * also writes stuff to stdout/err */
689 cp = fgets(line, STRLEN, fp);
692 if (str_starts(line, match_mdrun) )
696 if (str_starts(line, match_mpi) )
708 gmx_fatal(FARGS, "Need a threaded version of mdrun. This one\n"
710 "seems to have been compiled with MPI instead.",
718 gmx_fatal(FARGS, "Need an MPI-enabled version of mdrun. This one\n"
720 "seems to have been compiled without MPI support.",
727 gmx_fatal(FARGS, "Cannot execute mdrun. Please check %s for problems!",
731 fprintf(stdout, "passed.\n");
739 static void launch_simulation(
740 gmx_bool bLaunch, /* Should the simulation be launched? */
741 FILE *fp, /* General log file */
742 gmx_bool bThreads, /* whether to use threads */
743 char *cmd_mpirun, /* Command for mpirun */
744 char *cmd_np, /* Switch for -np or -ntmpi or empty */
745 char *cmd_mdrun, /* Command for mdrun */
746 char *args_for_mdrun, /* Arguments for mdrun */
747 const char *simulation_tpr, /* This tpr will be simulated */
748 int nPMEnodes) /* Number of PME nodes to use */
753 /* Make enough space for the system call command,
754 * (100 extra chars for -npme ... etc. options should suffice): */
755 snew(command, strlen(cmd_mpirun)+strlen(cmd_mdrun)+strlen(cmd_np)+strlen(args_for_mdrun)+strlen(simulation_tpr)+100);
757 /* Note that the -passall options requires args_for_mdrun to be at the end
758 * of the command line string */
761 sprintf(command, "%s%s-npme %d -s %s %s",
762 cmd_mdrun, cmd_np, nPMEnodes, simulation_tpr, args_for_mdrun);
766 sprintf(command, "%s%s%s -npme %d -s %s %s",
767 cmd_mpirun, cmd_np, cmd_mdrun, nPMEnodes, simulation_tpr, args_for_mdrun);
770 fprintf(fp, "%s this command line to launch the simulation:\n\n%s", bLaunch ? "Using" : "Please use", command);
774 /* Now the real thing! */
777 fprintf(stdout, "\nLaunching simulation with best parameters now.\nExecuting '%s'", command);
780 gmx_system_call(command);
785 static void modify_PMEsettings(
786 gmx_int64_t simsteps, /* Set this value as number of time steps */
787 gmx_int64_t init_step, /* Set this value as init_step */
788 const char *fn_best_tpr, /* tpr file with the best performance */
789 const char *fn_sim_tpr) /* name of tpr file to be launched */
797 read_tpx_state(fn_best_tpr, ir, &state, NULL, &mtop);
799 /* Reset nsteps and init_step to the value of the input .tpr file */
800 ir->nsteps = simsteps;
801 ir->init_step = init_step;
803 /* Write the tpr file which will be launched */
804 sprintf(buf, "Writing optimized simulation file %s with nsteps=%s.\n", fn_sim_tpr, "%"GMX_PRId64);
805 fprintf(stdout, buf, ir->nsteps);
807 write_tpx_state(fn_sim_tpr, ir, &state, &mtop);
812 static gmx_bool can_scale_rvdw(int vdwtype)
814 return (evdwCUT == vdwtype ||
818 #define EPME_SWITCHED(e) ((e) == eelPMESWITCH || (e) == eelPMEUSERSWITCH)
820 /* Make additional TPR files with more computational load for the
821 * direct space processors: */
822 static void make_benchmark_tprs(
823 const char *fn_sim_tpr, /* READ : User-provided tpr file */
824 char *fn_bench_tprs[], /* WRITE: Names of benchmark tpr files */
825 gmx_int64_t benchsteps, /* Number of time steps for benchmark runs */
826 gmx_int64_t statesteps, /* Step counter in checkpoint file */
827 real rmin, /* Minimal Coulomb radius */
828 real rmax, /* Maximal Coulomb radius */
829 real bScaleRvdw, /* Scale rvdw along with rcoulomb */
830 int *ntprs, /* No. of TPRs to write, each with a different
831 rcoulomb and fourierspacing */
832 t_inputinfo *info, /* Contains information about mdp file options */
833 FILE *fp) /* Write the output here */
839 real nlist_buffer; /* Thickness of the buffer regions for PME-switch potentials */
842 gmx_bool bNote = FALSE;
843 real add; /* Add this to rcoul for the next test */
844 real fac = 1.0; /* Scaling factor for Coulomb radius */
845 real fourierspacing; /* Basic fourierspacing from tpr */
848 sprintf(buf, "Making benchmark tpr file%s with %s time step%s",
849 *ntprs > 1 ? "s" : "", "%"GMX_PRId64, benchsteps > 1 ? "s" : "");
850 fprintf(stdout, buf, benchsteps);
853 sprintf(buf, " (adding %s steps from checkpoint file)", "%"GMX_PRId64);
854 fprintf(stdout, buf, statesteps);
855 benchsteps += statesteps;
857 fprintf(stdout, ".\n");
861 read_tpx_state(fn_sim_tpr, ir, &state, NULL, &mtop);
863 /* Check if some kind of PME was chosen */
864 if (EEL_PME(ir->coulombtype) == FALSE)
866 gmx_fatal(FARGS, "Can only do optimizations for simulations with %s electrostatics.",
870 /* Check if rcoulomb == rlist, which is necessary for plain PME. */
871 if ( (ir->cutoff_scheme != ecutsVERLET) &&
872 (eelPME == ir->coulombtype) && !(ir->rcoulomb == ir->rlist))
874 gmx_fatal(FARGS, "%s requires rcoulomb (%f) to be equal to rlist (%f).",
875 EELTYPE(eelPME), ir->rcoulomb, ir->rlist);
877 /* For other PME types, rcoulomb is allowed to be smaller than rlist */
878 else if (ir->rcoulomb > ir->rlist)
880 gmx_fatal(FARGS, "%s requires rcoulomb (%f) to be equal to or smaller than rlist (%f)",
881 EELTYPE(ir->coulombtype), ir->rcoulomb, ir->rlist);
884 if (bScaleRvdw && ir->rvdw != ir->rcoulomb)
886 fprintf(stdout, "NOTE: input rvdw != rcoulomb, will not scale rvdw\n");
890 /* Reduce the number of steps for the benchmarks */
891 info->orig_sim_steps = ir->nsteps;
892 ir->nsteps = benchsteps;
893 /* We must not use init_step from the input tpr file for the benchmarks */
894 info->orig_init_step = ir->init_step;
897 /* For PME-switch potentials, keep the radial distance of the buffer region */
898 nlist_buffer = ir->rlist - ir->rcoulomb;
900 /* Determine length of triclinic box vectors */
901 for (d = 0; d < DIM; d++)
904 for (i = 0; i < DIM; i++)
906 box_size[d] += state.box[d][i]*state.box[d][i];
908 box_size[d] = sqrt(box_size[d]);
911 if (ir->fourier_spacing > 0)
913 info->fsx[0] = ir->fourier_spacing;
914 info->fsy[0] = ir->fourier_spacing;
915 info->fsz[0] = ir->fourier_spacing;
919 /* Reconstruct fourierspacing per dimension from the number of grid points and box size */
920 info->fsx[0] = box_size[XX]/ir->nkx;
921 info->fsy[0] = box_size[YY]/ir->nky;
922 info->fsz[0] = box_size[ZZ]/ir->nkz;
925 /* If no value for the fourierspacing was provided on the command line, we
926 * use the reconstruction from the tpr file */
927 if (ir->fourier_spacing > 0)
929 /* Use the spacing from the tpr */
930 fourierspacing = ir->fourier_spacing;
934 /* Use the maximum observed spacing */
935 fourierspacing = max(max(info->fsx[0], info->fsy[0]), info->fsz[0]);
938 fprintf(stdout, "Calculating PME grid points on the basis of a fourierspacing of %f nm\n", fourierspacing);
940 /* For performance comparisons the number of particles is useful to have */
941 fprintf(fp, " Number of particles : %d\n", mtop.natoms);
943 /* Print information about settings of which some are potentially modified: */
944 fprintf(fp, " Coulomb type : %s\n", EELTYPE(ir->coulombtype));
945 fprintf(fp, " Grid spacing x y z : %f %f %f\n",
946 box_size[XX]/ir->nkx, box_size[YY]/ir->nky, box_size[ZZ]/ir->nkz);
947 fprintf(fp, " Van der Waals type : %s\n", EVDWTYPE(ir->vdwtype));
948 if (ir_vdw_switched(ir))
950 fprintf(fp, " rvdw_switch : %f nm\n", ir->rvdw_switch);
952 if (EPME_SWITCHED(ir->coulombtype))
954 fprintf(fp, " rlist : %f nm\n", ir->rlist);
956 if (ir->rlistlong != max_cutoff(ir->rvdw, ir->rcoulomb))
958 fprintf(fp, " rlistlong : %f nm\n", ir->rlistlong);
961 /* Print a descriptive line about the tpr settings tested */
962 fprintf(fp, "\nWill try these real/reciprocal workload settings:\n");
963 fprintf(fp, " No. scaling rcoulomb");
964 fprintf(fp, " nkx nky nkz");
965 fprintf(fp, " spacing");
966 if (can_scale_rvdw(ir->vdwtype))
968 fprintf(fp, " rvdw");
970 if (EPME_SWITCHED(ir->coulombtype))
972 fprintf(fp, " rlist");
974 if (ir->rlistlong != max_cutoff(ir->rlist, max_cutoff(ir->rvdw, ir->rcoulomb)) )
976 fprintf(fp, " rlistlong");
978 fprintf(fp, " tpr file\n");
980 /* Loop to create the requested number of tpr input files */
981 for (j = 0; j < *ntprs; j++)
983 /* The first .tpr is the provided one, just need to modify nsteps,
984 * so skip the following block */
987 /* Determine which Coulomb radii rc to use in the benchmarks */
988 add = (rmax-rmin)/(*ntprs-1);
989 if (is_equal(rmin, info->rcoulomb[0]))
991 ir->rcoulomb = rmin + j*add;
993 else if (is_equal(rmax, info->rcoulomb[0]))
995 ir->rcoulomb = rmin + (j-1)*add;
999 /* rmin != rcoul != rmax, ergo test between rmin and rmax */
1000 add = (rmax-rmin)/(*ntprs-2);
1001 ir->rcoulomb = rmin + (j-1)*add;
1004 /* Determine the scaling factor fac */
1005 fac = ir->rcoulomb/info->rcoulomb[0];
1007 /* Scale the Fourier grid spacing */
1008 ir->nkx = ir->nky = ir->nkz = 0;
1009 calc_grid(NULL, state.box, fourierspacing*fac, &ir->nkx, &ir->nky, &ir->nkz);
1011 /* Adjust other radii since various conditions need to be fulfilled */
1012 if (eelPME == ir->coulombtype)
1014 /* plain PME, rcoulomb must be equal to rlist */
1015 ir->rlist = ir->rcoulomb;
1019 /* rlist must be >= rcoulomb, we keep the size of the buffer region */
1020 ir->rlist = ir->rcoulomb + nlist_buffer;
1023 if (bScaleRvdw && can_scale_rvdw(ir->vdwtype))
1025 if (ecutsVERLET == ir->cutoff_scheme ||
1026 evdwPME == ir->vdwtype)
1028 /* With either the Verlet cutoff-scheme or LJ-PME,
1029 the van der Waals radius must always equal the
1031 ir->rvdw = ir->rcoulomb;
1035 /* For vdw cutoff, rvdw >= rlist */
1036 ir->rvdw = max(info->rvdw[0], ir->rlist);
1040 ir->rlistlong = max_cutoff(ir->rlist, max_cutoff(ir->rvdw, ir->rcoulomb));
1042 } /* end of "if (j != 0)" */
1044 /* for j==0: Save the original settings
1045 * for j >0: Save modified radii and Fourier grids */
1046 info->rcoulomb[j] = ir->rcoulomb;
1047 info->rvdw[j] = ir->rvdw;
1048 info->nkx[j] = ir->nkx;
1049 info->nky[j] = ir->nky;
1050 info->nkz[j] = ir->nkz;
1051 info->rlist[j] = ir->rlist;
1052 info->rlistlong[j] = ir->rlistlong;
1053 info->fsx[j] = fac*fourierspacing;
1054 info->fsy[j] = fac*fourierspacing;
1055 info->fsz[j] = fac*fourierspacing;
1057 /* Write the benchmark tpr file */
1058 strncpy(fn_bench_tprs[j], fn_sim_tpr, strlen(fn_sim_tpr)-strlen(".tpr"));
1059 sprintf(buf, "_bench%.2d.tpr", j);
1060 strcat(fn_bench_tprs[j], buf);
1061 fprintf(stdout, "Writing benchmark tpr %s with nsteps=", fn_bench_tprs[j]);
1062 fprintf(stdout, "%"GMX_PRId64, ir->nsteps);
1065 fprintf(stdout, ", scaling factor %f\n", fac);
1069 fprintf(stdout, ", unmodified settings\n");
1072 write_tpx_state(fn_bench_tprs[j], ir, &state, &mtop);
1074 /* Write information about modified tpr settings to log file */
1075 fprintf(fp, "%4d%10f%10f", j, fac, ir->rcoulomb);
1076 fprintf(fp, "%5d%5d%5d", ir->nkx, ir->nky, ir->nkz);
1077 fprintf(fp, " %9f ", info->fsx[j]);
1078 if (can_scale_rvdw(ir->vdwtype))
1080 fprintf(fp, "%10f", ir->rvdw);
1082 if (EPME_SWITCHED(ir->coulombtype))
1084 fprintf(fp, "%10f", ir->rlist);
1086 if (info->rlistlong[0] != max_cutoff(info->rlist[0], max_cutoff(info->rvdw[0], info->rcoulomb[0])) )
1088 fprintf(fp, "%10f", ir->rlistlong);
1090 fprintf(fp, " %-14s\n", fn_bench_tprs[j]);
1092 /* Make it clear to the user that some additional settings were modified */
1093 if (!is_equal(ir->rvdw, info->rvdw[0])
1094 || !is_equal(ir->rlistlong, info->rlistlong[0]) )
1101 fprintf(fp, "\nNote that in addition to the Coulomb radius and the Fourier grid\n"
1102 "other input settings were also changed (see table above).\n"
1103 "Please check if the modified settings are appropriate.\n");
1111 /* Rename the files we want to keep to some meaningful filename and
1112 * delete the rest */
1113 static void cleanup(const t_filenm *fnm, int nfile, int k, int nnodes,
1114 int nPMEnodes, int nr, gmx_bool bKeepStderr)
1116 char numstring[STRLEN];
1117 char newfilename[STRLEN];
1118 const char *fn = NULL;
1123 fprintf(stdout, "Cleaning up, deleting benchmark temp files ...\n");
1125 for (i = 0; i < nfile; i++)
1127 opt = (char *)fnm[i].opt;
1128 if (strcmp(opt, "-p") == 0)
1130 /* do nothing; keep this file */
1133 else if (strcmp(opt, "-bg") == 0)
1135 /* Give the log file a nice name so one can later see which parameters were used */
1136 numstring[0] = '\0';
1139 sprintf(numstring, "_%d", nr);
1141 sprintf(newfilename, "%s_no%d_np%d_npme%d%s", opt2fn("-bg", nfile, fnm), k, nnodes, nPMEnodes, numstring);
1142 if (gmx_fexist(opt2fn("-bg", nfile, fnm)))
1144 fprintf(stdout, "renaming log file to %s\n", newfilename);
1145 make_backup(newfilename);
1146 rename(opt2fn("-bg", nfile, fnm), newfilename);
1149 else if (strcmp(opt, "-err") == 0)
1151 /* This file contains the output of stderr. We want to keep it in
1152 * cases where there have been problems. */
1153 fn = opt2fn(opt, nfile, fnm);
1154 numstring[0] = '\0';
1157 sprintf(numstring, "_%d", nr);
1159 sprintf(newfilename, "%s_no%d_np%d_npme%d%s", fn, k, nnodes, nPMEnodes, numstring);
1164 fprintf(stdout, "Saving stderr output in %s\n", newfilename);
1165 make_backup(newfilename);
1166 rename(fn, newfilename);
1170 fprintf(stdout, "Deleting %s\n", fn);
1175 /* Delete the files which are created for each benchmark run: (options -b*) */
1176 else if ( (0 == strncmp(opt, "-b", 2)) && (opt2bSet(opt, nfile, fnm) || !is_optional(&fnm[i])) )
1178 remove_if_exists(opt2fn(opt, nfile, fnm));
1185 eNpmeAuto, eNpmeAll, eNpmeReduced, eNpmeSubset, eNpmeNr
1188 /* Create a list of numbers of PME nodes to test */
1189 static void make_npme_list(
1190 const char *npmevalues_opt, /* Make a complete list with all
1191 * possibilities or a short list that keeps only
1192 * reasonable numbers of PME nodes */
1193 int *nentries, /* Number of entries we put in the nPMEnodes list */
1194 int *nPMEnodes[], /* Each entry contains the value for -npme */
1195 int nnodes, /* Total number of nodes to do the tests on */
1196 int minPMEnodes, /* Minimum number of PME nodes */
1197 int maxPMEnodes) /* Maximum number of PME nodes */
1200 int min_factor = 1; /* We request that npp and npme have this minimal
1201 * largest common factor (depends on npp) */
1202 int nlistmax; /* Max. list size */
1203 int nlist; /* Actual number of entries in list */
1207 /* Do we need to check all possible values for -npme or is a reduced list enough? */
1208 if (0 == strcmp(npmevalues_opt, "all") )
1212 else if (0 == strcmp(npmevalues_opt, "subset") )
1214 eNPME = eNpmeSubset;
1216 else /* "auto" or "range" */
1222 else if (nnodes < 128)
1224 eNPME = eNpmeReduced;
1228 eNPME = eNpmeSubset;
1232 /* Calculate how many entries we could possibly have (in case of -npme all) */
1235 nlistmax = maxPMEnodes - minPMEnodes + 3;
1236 if (0 == minPMEnodes)
1246 /* Now make the actual list which is at most of size nlist */
1247 snew(*nPMEnodes, nlistmax);
1248 nlist = 0; /* start counting again, now the real entries in the list */
1249 for (i = 0; i < nlistmax - 2; i++)
1251 npme = maxPMEnodes - i;
1262 /* For 2d PME we want a common largest factor of at least the cube
1263 * root of the number of PP nodes */
1264 min_factor = (int) pow(npp, 1.0/3.0);
1267 gmx_fatal(FARGS, "Unknown option for eNPME in make_npme_list");
1270 if (gmx_greatest_common_divisor(npp, npme) >= min_factor)
1272 (*nPMEnodes)[nlist] = npme;
1276 /* We always test 0 PME nodes and the automatic number */
1277 *nentries = nlist + 2;
1278 (*nPMEnodes)[nlist ] = 0;
1279 (*nPMEnodes)[nlist+1] = -1;
1281 fprintf(stderr, "Will try the following %d different values for -npme:\n", *nentries);
1282 for (i = 0; i < *nentries-1; i++)
1284 fprintf(stderr, "%d, ", (*nPMEnodes)[i]);
1286 fprintf(stderr, "and %d (auto).\n", (*nPMEnodes)[*nentries-1]);
1290 /* Allocate memory to store the performance data */
1291 static void init_perfdata(t_perf *perfdata[], int ntprs, int datasets, int repeats)
1296 for (k = 0; k < ntprs; k++)
1298 snew(perfdata[k], datasets);
1299 for (i = 0; i < datasets; i++)
1301 for (j = 0; j < repeats; j++)
1303 snew(perfdata[k][i].Gcycles, repeats);
1304 snew(perfdata[k][i].ns_per_day, repeats);
1305 snew(perfdata[k][i].PME_f_load, repeats);
1312 /* Check for errors on mdrun -h */
1313 static void make_sure_it_runs(char *mdrun_cmd_line, int length, FILE *fp,
1314 const t_filenm *fnm, int nfile)
1316 const char *fn = NULL;
1317 char *command, *msg;
1321 snew(command, length + 15);
1322 snew(msg, length + 500);
1324 fprintf(stdout, "Making sure the benchmarks can be executed by running just 1 step...\n");
1325 sprintf(command, "%s -nsteps 1 -quiet", mdrun_cmd_line);
1326 fprintf(stdout, "Executing '%s' ...\n", command);
1327 ret = gmx_system_call(command);
1331 /* To prevent confusion, do not again issue a gmx_fatal here since we already
1332 * get the error message from mdrun itself */
1333 sprintf(msg, "Cannot run the benchmark simulations! Please check the error message of\n"
1334 "mdrun for the source of the problem. Did you provide a command line\n"
1335 "argument that neither g_tune_pme nor mdrun understands? Offending command:\n"
1336 "\n%s\n\n", command);
1338 fprintf(stderr, "%s", msg);
1340 fprintf(fp, "%s", msg);
1344 fprintf(stdout, "Benchmarks can be executed!\n");
1346 /* Clean up the benchmark output files we just created */
1347 fprintf(stdout, "Cleaning up ...\n");
1348 remove_if_exists(opt2fn("-bc", nfile, fnm));
1349 remove_if_exists(opt2fn("-be", nfile, fnm));
1350 remove_if_exists(opt2fn("-bcpo", nfile, fnm));
1351 remove_if_exists(opt2fn("-bg", nfile, fnm));
1358 static void do_the_tests(
1359 FILE *fp, /* General g_tune_pme output file */
1360 char **tpr_names, /* Filenames of the input files to test */
1361 int maxPMEnodes, /* Max fraction of nodes to use for PME */
1362 int minPMEnodes, /* Min fraction of nodes to use for PME */
1363 int npme_fixed, /* If >= -1, test fixed number of PME
1365 const char *npmevalues_opt, /* Which -npme values should be tested */
1366 t_perf **perfdata, /* Here the performace data is stored */
1367 int *pmeentries, /* Entries in the nPMEnodes list */
1368 int repeats, /* Repeat each test this often */
1369 int nnodes, /* Total number of nodes = nPP + nPME */
1370 int nr_tprs, /* Total number of tpr files to test */
1371 gmx_bool bThreads, /* Threads or MPI? */
1372 char *cmd_mpirun, /* mpirun command string */
1373 char *cmd_np, /* "-np", "-n", whatever mpirun needs */
1374 char *cmd_mdrun, /* mdrun command string */
1375 char *cmd_args_bench, /* arguments for mdrun in a string */
1376 const t_filenm *fnm, /* List of filenames from command line */
1377 int nfile, /* Number of files specified on the cmdl. */
1378 int presteps, /* DLB equilibration steps, is checked */
1379 gmx_int64_t cpt_steps) /* Time step counter in the checkpoint */
1381 int i, nr, k, ret, count = 0, totaltests;
1382 int *nPMEnodes = NULL;
1385 char *command, *cmd_stub;
1387 gmx_bool bResetProblem = FALSE;
1388 gmx_bool bFirst = TRUE;
1391 /* This string array corresponds to the eParselog enum type at the start
1393 const char* ParseLog[] = {
1395 "Logfile not found!",
1396 "No timings, logfile truncated?",
1397 "Run was terminated.",
1398 "Counters were not reset properly.",
1399 "No DD grid found for these settings.",
1400 "TPX version conflict!",
1401 "mdrun was not started in parallel!",
1402 "Number of PP ranks has a prime factor that is too large.",
1405 char str_PME_f_load[13];
1408 /* Allocate space for the mdrun command line. 100 extra characters should
1409 be more than enough for the -npme etcetera arguments */
1410 cmdline_length = strlen(cmd_mpirun)
1413 + strlen(cmd_args_bench)
1414 + strlen(tpr_names[0]) + 100;
1415 snew(command, cmdline_length);
1416 snew(cmd_stub, cmdline_length);
1418 /* Construct the part of the command line that stays the same for all tests: */
1421 sprintf(cmd_stub, "%s%s", cmd_mdrun, cmd_np);
1425 sprintf(cmd_stub, "%s%s%s ", cmd_mpirun, cmd_np, cmd_mdrun);
1428 /* Create a list of numbers of PME nodes to test */
1429 if (npme_fixed < -1)
1431 make_npme_list(npmevalues_opt, pmeentries, &nPMEnodes,
1432 nnodes, minPMEnodes, maxPMEnodes);
1438 nPMEnodes[0] = npme_fixed;
1439 fprintf(stderr, "Will use a fixed number of %d PME-only ranks.\n", nPMEnodes[0]);
1444 fprintf(fp, "\nNo benchmarks done since number of repeats (-r) is 0.\n");
1446 finalize(opt2fn("-p", nfile, fnm));
1450 /* Allocate one dataset for each tpr input file: */
1451 init_perfdata(perfdata, nr_tprs, *pmeentries, repeats);
1453 /*****************************************/
1454 /* Main loop over all tpr files to test: */
1455 /*****************************************/
1456 totaltests = nr_tprs*(*pmeentries)*repeats;
1457 for (k = 0; k < nr_tprs; k++)
1459 fprintf(fp, "\nIndividual timings for input file %d (%s):\n", k, tpr_names[k]);
1460 fprintf(fp, "PME ranks Gcycles ns/day PME/f Remark\n");
1461 /* Loop over various numbers of PME nodes: */
1462 for (i = 0; i < *pmeentries; i++)
1464 pd = &perfdata[k][i];
1466 /* Loop over the repeats for each scenario: */
1467 for (nr = 0; nr < repeats; nr++)
1469 pd->nPMEnodes = nPMEnodes[i];
1471 /* Add -npme and -s to the command line and save it. Note that
1472 * the -passall (if set) options requires cmd_args_bench to be
1473 * at the end of the command line string */
1474 snew(pd->mdrun_cmd_line, cmdline_length);
1475 sprintf(pd->mdrun_cmd_line, "%s-npme %d -s %s %s",
1476 cmd_stub, pd->nPMEnodes, tpr_names[k], cmd_args_bench);
1478 /* To prevent that all benchmarks fail due to a show-stopper argument
1479 * on the mdrun command line, we make a quick check first */
1482 make_sure_it_runs(pd->mdrun_cmd_line, cmdline_length, fp, fnm, nfile);
1486 /* Do a benchmark simulation: */
1489 sprintf(buf, ", pass %d/%d", nr+1, repeats);
1495 fprintf(stdout, "\n=== Progress %2.0f%%, tpr %d/%d, run %d/%d%s:\n",
1496 (100.0*count)/totaltests,
1497 k+1, nr_tprs, i+1, *pmeentries, buf);
1498 make_backup(opt2fn("-err", nfile, fnm));
1499 sprintf(command, "%s 1> /dev/null 2>%s", pd->mdrun_cmd_line, opt2fn("-err", nfile, fnm));
1500 fprintf(stdout, "%s\n", pd->mdrun_cmd_line);
1501 gmx_system_call(command);
1503 /* Collect the performance data from the log file; also check stderr
1504 * for fatal errors */
1505 ret = parse_logfile(opt2fn("-bg", nfile, fnm), opt2fn("-err", nfile, fnm),
1506 pd, nr, presteps, cpt_steps, nnodes);
1507 if ((presteps > 0) && (ret == eParselogResetProblem))
1509 bResetProblem = TRUE;
1512 if (-1 == pd->nPMEnodes)
1514 sprintf(buf, "(%3d)", pd->guessPME);
1522 if (pd->PME_f_load[nr] > 0.0)
1524 sprintf(str_PME_f_load, "%12.3f", pd->PME_f_load[nr]);
1528 sprintf(str_PME_f_load, "%s", " - ");
1531 /* Write the data we got to disk */
1532 fprintf(fp, "%4d%s %12.3f %12.3f %s %s", pd->nPMEnodes,
1533 buf, pd->Gcycles[nr], pd->ns_per_day[nr], str_PME_f_load, ParseLog[ret]);
1534 if (!(ret == eParselogOK || ret == eParselogNoDDGrid || ret == eParselogNotFound) )
1536 fprintf(fp, " Check %s file for problems.", ret == eParselogFatal ? "err" : "log");
1542 /* Do some cleaning up and delete the files we do not need any more */
1543 cleanup(fnm, nfile, k, nnodes, pd->nPMEnodes, nr, ret == eParselogFatal);
1545 /* If the first run with this number of processors already failed, do not try again: */
1546 if (pd->Gcycles[0] <= 0.0 && repeats > 1)
1548 fprintf(stdout, "Skipping remaining passes of unsuccessful setting, see log file for details.\n");
1549 count += repeats-(nr+1);
1552 } /* end of repeats loop */
1553 } /* end of -npme loop */
1554 } /* end of tpr file loop */
1559 fprintf(fp, "WARNING: The cycle and time step counters could not be reset properly. ");
1567 static void check_input(
1574 real maxPMEfraction,
1575 real minPMEfraction,
1577 gmx_int64_t bench_nsteps,
1578 const t_filenm *fnm,
1588 /* Make sure the input file exists */
1589 if (!gmx_fexist(opt2fn("-s", nfile, fnm)))
1591 gmx_fatal(FARGS, "File %s not found.", opt2fn("-s", nfile, fnm));
1594 /* Make sure that the checkpoint file is not overwritten during benchmarking */
1595 if ( (0 == strcmp(opt2fn("-cpi", nfile, fnm), opt2fn("-bcpo", nfile, fnm)) ) && (sim_part > 1) )
1597 gmx_fatal(FARGS, "Checkpoint input (-cpi) and benchmark checkpoint output (-bcpo) files must not be identical.\n"
1598 "The checkpoint input file must not be overwritten during the benchmarks.\n");
1601 /* Make sure that repeats is >= 0 (if == 0, only write tpr files) */
1604 gmx_fatal(FARGS, "Number of repeats < 0!");
1607 /* Check number of nodes */
1610 gmx_fatal(FARGS, "Number of ranks/threads must be a positive integer.");
1613 /* Automatically choose -ntpr if not set */
1623 /* Set a reasonable scaling factor for rcoulomb */
1626 *rmax = rcoulomb * 1.2;
1629 fprintf(stderr, "Will test %d tpr file%s.\n", *ntprs, *ntprs == 1 ? "" : "s");
1635 fprintf(stderr, "Note: Choose ntpr>1 to shift PME load between real and reciprocal space.\n");
1639 /* Make shure that rmin <= rcoulomb <= rmax */
1648 if (!(*rmin <= *rmax) )
1650 gmx_fatal(FARGS, "Please choose the Coulomb radii such that rmin <= rmax.\n"
1651 "rmin = %g, rmax = %g, actual rcoul from .tpr file = %g\n", *rmin, *rmax, rcoulomb);
1653 /* Add test scenarios if rmin or rmax were set */
1656 if (!is_equal(*rmin, rcoulomb) && (*ntprs == 1) )
1659 fprintf(stderr, "NOTE: Setting -rmin to %g changed -ntpr to %d\n",
1662 if (!is_equal(*rmax, rcoulomb) && (*ntprs == 1) )
1665 fprintf(stderr, "NOTE: Setting -rmax to %g changed -ntpr to %d\n",
1670 /* If one of rmin, rmax is set, we need 2 tpr files at minimum */
1671 if (!is_equal(*rmax, rcoulomb) || !is_equal(*rmin, rcoulomb) )
1673 *ntprs = max(*ntprs, 2);
1676 /* If both rmin, rmax are set, we need 3 tpr files at minimum */
1677 if (!is_equal(*rmax, rcoulomb) && !is_equal(*rmin, rcoulomb) )
1679 *ntprs = max(*ntprs, 3);
1684 fprintf(stderr, "NOTE: Your rmin, rmax setting changed -ntpr to %d\n", *ntprs);
1689 if (is_equal(*rmin, rcoulomb) && is_equal(rcoulomb, *rmax)) /* We have just a single rc */
1691 fprintf(stderr, "WARNING: Resetting -ntpr to 1 since no Coulomb radius scaling is requested.\n"
1692 "Please set rmin < rmax to test Coulomb radii in the [rmin, rmax] interval\n"
1693 "with correspondingly adjusted PME grid settings\n");
1698 /* Check whether max and min fraction are within required values */
1699 if (maxPMEfraction > 0.5 || maxPMEfraction < 0)
1701 gmx_fatal(FARGS, "-max must be between 0 and 0.5");
1703 if (minPMEfraction > 0.5 || minPMEfraction < 0)
1705 gmx_fatal(FARGS, "-min must be between 0 and 0.5");
1707 if (maxPMEfraction < minPMEfraction)
1709 gmx_fatal(FARGS, "-max must be larger or equal to -min");
1712 /* Check whether the number of steps - if it was set - has a reasonable value */
1713 if (bench_nsteps < 0)
1715 gmx_fatal(FARGS, "Number of steps must be positive.");
1718 if (bench_nsteps > 10000 || bench_nsteps < 100)
1720 fprintf(stderr, "WARNING: steps=");
1721 fprintf(stderr, "%"GMX_PRId64, bench_nsteps);
1722 fprintf(stderr, ". Are you sure you want to perform so %s steps for each benchmark?\n", (bench_nsteps < 100) ? "few" : "many");
1727 gmx_fatal(FARGS, "Cannot have a negative number of presteps.\n");
1730 /* Check for rcoulomb scaling if more than one .tpr file is tested */
1733 if (*rmin/rcoulomb < 0.75 || *rmax/rcoulomb > 1.25)
1735 fprintf(stderr, "WARNING: Applying extreme scaling factor. I hope you know what you are doing.\n");
1739 /* If a fixed number of PME nodes is set we do rcoulomb and PME gird tuning
1740 * only. We need to check whether the requested number of PME-only nodes
1742 if (npme_fixed > -1)
1744 /* No more than 50% of all nodes can be assigned as PME-only nodes. */
1745 if (2*npme_fixed > nnodes)
1747 gmx_fatal(FARGS, "Cannot have more than %d PME-only ranks for a total of %d ranks (you chose %d).\n",
1748 nnodes/2, nnodes, npme_fixed);
1750 if ((npme_fixed > 0) && (5*npme_fixed < nnodes))
1752 fprintf(stderr, "WARNING: Only %g percent of the ranks are assigned as PME-only ranks.\n",
1753 100.0*((real)npme_fixed / (real)nnodes));
1755 if (opt2parg_bSet("-min", npargs, pa) || opt2parg_bSet("-max", npargs, pa))
1757 fprintf(stderr, "NOTE: The -min, -max, and -npme options have no effect when a\n"
1758 " fixed number of PME-only ranks is requested with -fix.\n");
1764 /* Returns TRUE when "opt" is needed at launch time */
1765 static gmx_bool is_launch_file(char *opt, gmx_bool bSet)
1767 if (0 == strncmp(opt, "-swap", 5))
1772 /* Apart from the input .tpr and the output log files we need all options that
1773 * were set on the command line and that do not start with -b */
1774 if (0 == strncmp(opt, "-b", 2) || 0 == strncmp(opt, "-s", 2)
1775 || 0 == strncmp(opt, "-err", 4) || 0 == strncmp(opt, "-p", 2) )
1784 /* Returns TRUE when "opt" defines a file which is needed for the benchmarks runs */
1785 static gmx_bool is_bench_file(char *opt, gmx_bool bSet, gmx_bool bOptional, gmx_bool bIsOutput)
1787 /* Apart from the input .tpr, all files starting with "-b" are for
1788 * _b_enchmark files exclusively */
1789 if (0 == strncmp(opt, "-s", 2))
1794 if (0 == strncmp(opt, "-b", 2) || 0 == strncmp(opt, "-s", 2))
1796 if (!bOptional || bSet)
1813 if (bSet) /* These are additional input files like -cpi -ei */
1826 /* Adds 'buf' to 'str' */
1827 static void add_to_string(char **str, char *buf)
1832 len = strlen(*str) + strlen(buf) + 1;
1838 /* Create the command line for the benchmark as well as for the real run */
1839 static void create_command_line_snippets(
1840 gmx_bool bAppendFiles,
1841 gmx_bool bKeepAndNumCPT,
1842 gmx_bool bResetHWay,
1846 char *cmd_args_bench[], /* command line arguments for benchmark runs */
1847 char *cmd_args_launch[], /* command line arguments for simulation run */
1848 char extra_args[]) /* Add this to the end of the command line */
1853 char strbuf[STRLEN];
1856 /* strlen needs at least '\0' as a string: */
1857 snew(*cmd_args_bench, 1);
1858 snew(*cmd_args_launch, 1);
1859 *cmd_args_launch[0] = '\0';
1860 *cmd_args_bench[0] = '\0';
1863 /*******************************************/
1864 /* 1. Process other command line arguments */
1865 /*******************************************/
1868 /* Add equilibration steps to benchmark options */
1869 sprintf(strbuf, "-resetstep %d ", presteps);
1870 add_to_string(cmd_args_bench, strbuf);
1872 /* These switches take effect only at launch time */
1873 if (FALSE == bAppendFiles)
1875 add_to_string(cmd_args_launch, "-noappend ");
1879 add_to_string(cmd_args_launch, "-cpnum ");
1883 add_to_string(cmd_args_launch, "-resethway ");
1886 /********************/
1887 /* 2. Process files */
1888 /********************/
1889 for (i = 0; i < nfile; i++)
1891 opt = (char *)fnm[i].opt;
1892 name = opt2fn(opt, nfile, fnm);
1894 /* Strbuf contains the options, now let's sort out where we need that */
1895 sprintf(strbuf, "%s %s ", opt, name);
1897 if (is_bench_file(opt, opt2bSet(opt, nfile, fnm), is_optional(&fnm[i]), is_output(&fnm[i])) )
1899 /* All options starting with -b* need the 'b' removed,
1900 * therefore overwrite strbuf */
1901 if (0 == strncmp(opt, "-b", 2))
1903 sprintf(strbuf, "-%s %s ", &opt[2], name);
1906 add_to_string(cmd_args_bench, strbuf);
1909 if (is_launch_file(opt, opt2bSet(opt, nfile, fnm)) )
1911 add_to_string(cmd_args_launch, strbuf);
1915 add_to_string(cmd_args_bench, extra_args);
1916 add_to_string(cmd_args_launch, extra_args);
1920 /* Set option opt */
1921 static void setopt(const char *opt, int nfile, t_filenm fnm[])
1925 for (i = 0; (i < nfile); i++)
1927 if (strcmp(opt, fnm[i].opt) == 0)
1929 fnm[i].flag |= ffSET;
1935 /* This routine inspects the tpr file and ...
1936 * 1. checks for output files that get triggered by a tpr option. These output
1937 * files are marked as 'set' to allow for a proper cleanup after each
1939 * 2. returns the PME:PP load ratio
1940 * 3. returns rcoulomb from the tpr */
1941 static float inspect_tpr(int nfile, t_filenm fnm[], real *rcoulomb)
1943 gmx_bool bPull; /* Is pulling requested in .tpr file? */
1944 gmx_bool bTpi; /* Is test particle insertion requested? */
1945 gmx_bool bFree; /* Is a free energy simulation requested? */
1946 gmx_bool bNM; /* Is a normal mode analysis requested? */
1947 gmx_bool bSwap; /* Is water/ion position swapping requested? */
1953 /* Check tpr file for options that trigger extra output files */
1954 read_tpx_state(opt2fn("-s", nfile, fnm), &ir, &state, NULL, &mtop);
1955 bPull = (epullNO != ir.ePull);
1956 bFree = (efepNO != ir.efep );
1957 bNM = (eiNM == ir.eI );
1958 bSwap = (eswapNO != ir.eSwapCoords);
1959 bTpi = EI_TPI(ir.eI);
1961 /* Set these output files on the tuning command-line */
1964 setopt("-pf", nfile, fnm);
1965 setopt("-px", nfile, fnm);
1969 setopt("-dhdl", nfile, fnm);
1973 setopt("-tpi", nfile, fnm);
1974 setopt("-tpid", nfile, fnm);
1978 setopt("-mtx", nfile, fnm);
1982 setopt("-swap", nfile, fnm);
1985 *rcoulomb = ir.rcoulomb;
1987 /* Return the estimate for the number of PME nodes */
1988 return pme_load_estimate(&mtop, &ir, state.box);
1992 static void couple_files_options(int nfile, t_filenm fnm[])
1995 gmx_bool bSet, bBench;
2000 for (i = 0; i < nfile; i++)
2002 opt = (char *)fnm[i].opt;
2003 bSet = ((fnm[i].flag & ffSET) != 0);
2004 bBench = (0 == strncmp(opt, "-b", 2));
2006 /* Check optional files */
2007 /* If e.g. -eo is set, then -beo also needs to be set */
2008 if (is_optional(&fnm[i]) && bSet && !bBench)
2010 sprintf(buf, "-b%s", &opt[1]);
2011 setopt(buf, nfile, fnm);
2013 /* If -beo is set, then -eo also needs to be! */
2014 if (is_optional(&fnm[i]) && bSet && bBench)
2016 sprintf(buf, "-%s", &opt[2]);
2017 setopt(buf, nfile, fnm);
2023 #define BENCHSTEPS (1000)
2025 int gmx_tune_pme(int argc, char *argv[])
2027 const char *desc[] = {
2028 "For a given number [TT]-np[tt] or [TT]-ntmpi[tt] of ranks, [THISMODULE] systematically",
2029 "times [gmx-mdrun] with various numbers of PME-only ranks and determines",
2030 "which setting is fastest. It will also test whether performance can",
2031 "be enhanced by shifting load from the reciprocal to the real space",
2032 "part of the Ewald sum. ",
2033 "Simply pass your [TT].tpr[tt] file to [THISMODULE] together with other options",
2034 "for [gmx-mdrun] as needed.[PAR]",
2035 "Which executables are used can be set in the environment variables",
2036 "MPIRUN and MDRUN. If these are not present, 'mpirun' and 'mdrun'",
2037 "will be used as defaults. Note that for certain MPI frameworks you",
2038 "need to provide a machine- or hostfile. This can also be passed",
2039 "via the MPIRUN variable, e.g.[PAR]",
2040 "[TT]export MPIRUN=\"/usr/local/mpirun -machinefile hosts\"[tt][PAR]",
2041 "Please call [THISMODULE] with the normal options you would pass to",
2042 "[gmx-mdrun] and add [TT]-np[tt] for the number of ranks to perform the",
2043 "tests on, or [TT]-ntmpi[tt] for the number of threads. You can also add [TT]-r[tt]",
2044 "to repeat each test several times to get better statistics. [PAR]",
2045 "[THISMODULE] can test various real space / reciprocal space workloads",
2046 "for you. With [TT]-ntpr[tt] you control how many extra [TT].tpr[tt] files will be",
2047 "written with enlarged cutoffs and smaller Fourier grids respectively.",
2048 "Typically, the first test (number 0) will be with the settings from the input",
2049 "[TT].tpr[tt] file; the last test (number [TT]ntpr[tt]) will have the Coulomb cutoff",
2050 "specified by [TT]-rmax[tt] with a somwhat smaller PME grid at the same time. ",
2051 "In this last test, the Fourier spacing is multiplied with [TT]rmax[tt]/rcoulomb. ",
2052 "The remaining [TT].tpr[tt] files will have equally-spaced Coulomb radii (and Fourier "
2053 "spacings) between these extremes. [BB]Note[bb] that you can set [TT]-ntpr[tt] to 1",
2054 "if you just seek the optimal number of PME-only ranks; in that case",
2055 "your input [TT].tpr[tt] file will remain unchanged.[PAR]",
2056 "For the benchmark runs, the default of 1000 time steps should suffice for most",
2057 "MD systems. The dynamic load balancing needs about 100 time steps",
2058 "to adapt to local load imbalances, therefore the time step counters",
2059 "are by default reset after 100 steps. For large systems (>1M atoms), as well as ",
2060 "for a higher accuarcy of the measurements, you should set [TT]-resetstep[tt] to a higher value.",
2061 "From the 'DD' load imbalance entries in the md.log output file you",
2062 "can tell after how many steps the load is sufficiently balanced. Example call:[PAR]"
2063 "[TT]gmx tune_pme -np 64 -s protein.tpr -launch[tt][PAR]",
2064 "After calling [gmx-mdrun] several times, detailed performance information",
2065 "is available in the output file [TT]perf.out[tt].",
2066 "[BB]Note[bb] that during the benchmarks, a couple of temporary files are written",
2067 "(options [TT]-b[tt]*), these will be automatically deleted after each test.[PAR]",
2068 "If you want the simulation to be started automatically with the",
2069 "optimized parameters, use the command line option [TT]-launch[tt].[PAR]",
2074 int pmeentries = 0; /* How many values for -npme do we actually test for each tpr file */
2075 real maxPMEfraction = 0.50;
2076 real minPMEfraction = 0.25;
2077 int maxPMEnodes, minPMEnodes;
2078 float guessPMEratio; /* guessed PME:PP ratio based on the tpr file */
2079 float guessPMEnodes;
2080 int npme_fixed = -2; /* If >= -1, use only this number
2081 * of PME-only nodes */
2083 real rmin = 0.0, rmax = 0.0; /* min and max value for rcoulomb if scaling is requested */
2084 real rcoulomb = -1.0; /* Coulomb radius as set in .tpr file */
2085 gmx_bool bScaleRvdw = TRUE;
2086 gmx_int64_t bench_nsteps = BENCHSTEPS;
2087 gmx_int64_t new_sim_nsteps = -1; /* -1 indicates: not set by the user */
2088 gmx_int64_t cpt_steps = 0; /* Step counter in .cpt input file */
2089 int presteps = 100; /* Do a full cycle reset after presteps steps */
2090 gmx_bool bOverwrite = FALSE, bKeepTPR;
2091 gmx_bool bLaunch = FALSE;
2092 char *ExtraArgs = NULL;
2093 char **tpr_names = NULL;
2094 const char *simulation_tpr = NULL;
2095 int best_npme, best_tpr;
2096 int sim_part = 1; /* For benchmarks with checkpoint files */
2099 /* Default program names if nothing else is found */
2100 char *cmd_mpirun = NULL, *cmd_mdrun = NULL;
2101 char *cmd_args_bench, *cmd_args_launch;
2102 char *cmd_np = NULL;
2104 t_perf **perfdata = NULL;
2110 /* Print out how long the tuning took */
2113 static t_filenm fnm[] = {
2115 { efOUT, "-p", "perf", ffWRITE },
2116 { efLOG, "-err", "bencherr", ffWRITE },
2117 { efTPX, "-so", "tuned", ffWRITE },
2119 { efTPX, NULL, NULL, ffREAD },
2120 { efTRN, "-o", NULL, ffWRITE },
2121 { efCOMPRESSED, "-x", NULL, ffOPTWR },
2122 { efCPT, "-cpi", NULL, ffOPTRD },
2123 { efCPT, "-cpo", NULL, ffOPTWR },
2124 { efSTO, "-c", "confout", ffWRITE },
2125 { efEDR, "-e", "ener", ffWRITE },
2126 { efLOG, "-g", "md", ffWRITE },
2127 { efXVG, "-dhdl", "dhdl", ffOPTWR },
2128 { efXVG, "-field", "field", ffOPTWR },
2129 { efXVG, "-table", "table", ffOPTRD },
2130 { efXVG, "-tabletf", "tabletf", ffOPTRD },
2131 { efXVG, "-tablep", "tablep", ffOPTRD },
2132 { efXVG, "-tableb", "table", ffOPTRD },
2133 { efTRX, "-rerun", "rerun", ffOPTRD },
2134 { efXVG, "-tpi", "tpi", ffOPTWR },
2135 { efXVG, "-tpid", "tpidist", ffOPTWR },
2136 { efEDI, "-ei", "sam", ffOPTRD },
2137 { efXVG, "-eo", "edsam", ffOPTWR },
2138 { efXVG, "-devout", "deviatie", ffOPTWR },
2139 { efXVG, "-runav", "runaver", ffOPTWR },
2140 { efXVG, "-px", "pullx", ffOPTWR },
2141 { efXVG, "-pf", "pullf", ffOPTWR },
2142 { efXVG, "-ro", "rotation", ffOPTWR },
2143 { efLOG, "-ra", "rotangles", ffOPTWR },
2144 { efLOG, "-rs", "rotslabs", ffOPTWR },
2145 { efLOG, "-rt", "rottorque", ffOPTWR },
2146 { efMTX, "-mtx", "nm", ffOPTWR },
2147 { efNDX, "-dn", "dipole", ffOPTWR },
2148 { efXVG, "-swap", "swapions", ffOPTWR },
2149 /* Output files that are deleted after each benchmark run */
2150 { efTRN, "-bo", "bench", ffWRITE },
2151 { efXTC, "-bx", "bench", ffWRITE },
2152 { efCPT, "-bcpo", "bench", ffWRITE },
2153 { efSTO, "-bc", "bench", ffWRITE },
2154 { efEDR, "-be", "bench", ffWRITE },
2155 { efLOG, "-bg", "bench", ffWRITE },
2156 { efXVG, "-beo", "benchedo", ffOPTWR },
2157 { efXVG, "-bdhdl", "benchdhdl", ffOPTWR },
2158 { efXVG, "-bfield", "benchfld", ffOPTWR },
2159 { efXVG, "-btpi", "benchtpi", ffOPTWR },
2160 { efXVG, "-btpid", "benchtpid", ffOPTWR },
2161 { efXVG, "-bdevout", "benchdev", ffOPTWR },
2162 { efXVG, "-brunav", "benchrnav", ffOPTWR },
2163 { efXVG, "-bpx", "benchpx", ffOPTWR },
2164 { efXVG, "-bpf", "benchpf", ffOPTWR },
2165 { efXVG, "-bro", "benchrot", ffOPTWR },
2166 { efLOG, "-bra", "benchrota", ffOPTWR },
2167 { efLOG, "-brs", "benchrots", ffOPTWR },
2168 { efLOG, "-brt", "benchrott", ffOPTWR },
2169 { efMTX, "-bmtx", "benchn", ffOPTWR },
2170 { efNDX, "-bdn", "bench", ffOPTWR },
2171 { efXVG, "-bswap", "benchswp", ffOPTWR }
2174 gmx_bool bThreads = FALSE;
2178 const char *procstring[] =
2179 { NULL, "-np", "-n", "none", NULL };
2180 const char *npmevalues_opt[] =
2181 { NULL, "auto", "all", "subset", NULL };
2183 gmx_bool bAppendFiles = TRUE;
2184 gmx_bool bKeepAndNumCPT = FALSE;
2185 gmx_bool bResetCountersHalfWay = FALSE;
2186 gmx_bool bBenchmark = TRUE;
2188 output_env_t oenv = NULL;
2191 /***********************/
2192 /* g_tune_pme options: */
2193 /***********************/
2194 { "-np", FALSE, etINT, {&nnodes},
2195 "Number of ranks to run the tests on (must be > 2 for separate PME ranks)" },
2196 { "-npstring", FALSE, etENUM, {procstring},
2197 "Specify the number of ranks to [TT]$MPIRUN[tt] using this string"},
2198 { "-ntmpi", FALSE, etINT, {&nthreads},
2199 "Number of MPI-threads to run the tests on (turns MPI & mpirun off)"},
2200 { "-r", FALSE, etINT, {&repeats},
2201 "Repeat each test this often" },
2202 { "-max", FALSE, etREAL, {&maxPMEfraction},
2203 "Max fraction of PME ranks to test with" },
2204 { "-min", FALSE, etREAL, {&minPMEfraction},
2205 "Min fraction of PME ranks to test with" },
2206 { "-npme", FALSE, etENUM, {npmevalues_opt},
2207 "Within -min and -max, benchmark all possible values for [TT]-npme[tt], or just a reasonable subset. "
2208 "Auto neglects -min and -max and chooses reasonable values around a guess for npme derived from the .tpr"},
2209 { "-fix", FALSE, etINT, {&npme_fixed},
2210 "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."},
2211 { "-rmax", FALSE, etREAL, {&rmax},
2212 "If >0, maximal rcoulomb for -ntpr>1 (rcoulomb upscaling results in fourier grid downscaling)" },
2213 { "-rmin", FALSE, etREAL, {&rmin},
2214 "If >0, minimal rcoulomb for -ntpr>1" },
2215 { "-scalevdw", FALSE, etBOOL, {&bScaleRvdw},
2216 "Scale rvdw along with rcoulomb"},
2217 { "-ntpr", FALSE, etINT, {&ntprs},
2218 "Number of [TT].tpr[tt] files to benchmark. Create this many files with different rcoulomb scaling factors depending on -rmin and -rmax. "
2219 "If < 1, automatically choose the number of [TT].tpr[tt] files to test" },
2220 { "-steps", FALSE, etINT64, {&bench_nsteps},
2221 "Take timings for this many steps in the benchmark runs" },
2222 { "-resetstep", FALSE, etINT, {&presteps},
2223 "Let dlb equilibrate this many steps before timings are taken (reset cycle counters after this many steps)" },
2224 { "-simsteps", FALSE, etINT64, {&new_sim_nsteps},
2225 "If non-negative, perform this many steps in the real run (overwrites nsteps from [TT].tpr[tt], add [TT].cpt[tt] steps)" },
2226 { "-launch", FALSE, etBOOL, {&bLaunch},
2227 "Launch the real simulation after optimization" },
2228 { "-bench", FALSE, etBOOL, {&bBenchmark},
2229 "Run the benchmarks or just create the input [TT].tpr[tt] files?" },
2230 /******************/
2231 /* mdrun options: */
2232 /******************/
2233 /* We let g_tune_pme parse and understand these options, because we need to
2234 * prevent that they appear on the mdrun command line for the benchmarks */
2235 { "-append", FALSE, etBOOL, {&bAppendFiles},
2236 "Append to previous output files when continuing from checkpoint instead of adding the simulation part number to all file names (for launch only)" },
2237 { "-cpnum", FALSE, etBOOL, {&bKeepAndNumCPT},
2238 "Keep and number checkpoint files (launch only)" },
2239 { "-resethway", FALSE, etBOOL, {&bResetCountersHalfWay},
2240 "HIDDENReset the cycle counters after half the number of steps or halfway [TT]-maxh[tt] (launch only)" }
2243 #define NFILE asize(fnm)
2245 seconds = gmx_gettime();
2247 if (!parse_common_args(&argc, argv, PCA_NOEXIT_ON_ARGS,
2248 NFILE, fnm, asize(pa), pa, asize(desc), desc,
2254 /* Store the remaining unparsed command line entries in a string which
2255 * is then attached to the mdrun command line */
2257 ExtraArgs[0] = '\0';
2258 for (i = 1; i < argc; i++) /* argc will now be 1 if everything was understood */
2260 add_to_string(&ExtraArgs, argv[i]);
2261 add_to_string(&ExtraArgs, " ");
2264 if (opt2parg_bSet("-ntmpi", asize(pa), pa))
2267 if (opt2parg_bSet("-npstring", asize(pa), pa))
2269 fprintf(stderr, "WARNING: -npstring has no effect when using threads.\n");
2274 gmx_fatal(FARGS, "Can't run multi-threaded MPI simulation yet!");
2276 /* and now we just set this; a bit of an ugly hack*/
2279 /* Check for PME:PP ratio and whether tpr triggers additional output files */
2280 guessPMEratio = inspect_tpr(NFILE, fnm, &rcoulomb);
2282 /* Automatically set -beo options if -eo is set etc. */
2283 couple_files_options(NFILE, fnm);
2285 /* Construct the command line arguments for benchmark runs
2286 * as well as for the simulation run */
2289 sprintf(bbuf, " -ntmpi %d ", nthreads);
2293 /* This string will be used for MPI runs and will appear after the
2294 * mpirun command. */
2295 if (strcmp(procstring[0], "none") != 0)
2297 sprintf(bbuf, " %s %d ", procstring[0], nnodes);
2307 create_command_line_snippets(bAppendFiles, bKeepAndNumCPT, bResetCountersHalfWay, presteps,
2308 NFILE, fnm, &cmd_args_bench, &cmd_args_launch, ExtraArgs);
2310 /* Read in checkpoint file if requested */
2312 if (opt2bSet("-cpi", NFILE, fnm))
2315 cr->duty = DUTY_PP; /* makes the following routine happy */
2316 read_checkpoint_simulation_part(opt2fn("-cpi", NFILE, fnm),
2317 &sim_part, &cpt_steps, cr,
2318 FALSE, NFILE, fnm, NULL, NULL);
2321 /* sim_part will now be 1 if no checkpoint file was found */
2324 gmx_fatal(FARGS, "Checkpoint file %s not found!", opt2fn("-cpi", NFILE, fnm));
2328 /* Open performance output file and write header info */
2329 fp = gmx_ffopen(opt2fn("-p", NFILE, fnm), "w");
2331 /* Make a quick consistency check of command line parameters */
2332 check_input(nnodes, repeats, &ntprs, &rmin, rcoulomb, &rmax,
2333 maxPMEfraction, minPMEfraction, npme_fixed,
2334 bench_nsteps, fnm, NFILE, sim_part, presteps,
2337 /* Determine the maximum and minimum number of PME nodes to test,
2338 * the actual list of settings is build in do_the_tests(). */
2339 if ((nnodes > 2) && (npme_fixed < -1))
2341 if (0 == strcmp(npmevalues_opt[0], "auto"))
2343 /* Determine the npme range automatically based on the PME:PP load guess */
2344 if (guessPMEratio > 1.0)
2346 /* More PME than PP work, probably we do not need separate PME nodes at all! */
2347 maxPMEnodes = nnodes/2;
2348 minPMEnodes = nnodes/2;
2352 /* PME : PP load is in the range 0..1, let's test around the guess */
2353 guessPMEnodes = nnodes/(1.0 + 1.0/guessPMEratio);
2354 minPMEnodes = floor(0.7*guessPMEnodes);
2355 maxPMEnodes = ceil(1.6*guessPMEnodes);
2356 maxPMEnodes = min(maxPMEnodes, nnodes/2);
2361 /* Determine the npme range based on user input */
2362 maxPMEnodes = floor(maxPMEfraction*nnodes);
2363 minPMEnodes = max(floor(minPMEfraction*nnodes), 0);
2364 fprintf(stdout, "Will try runs with %d ", minPMEnodes);
2365 if (maxPMEnodes != minPMEnodes)
2367 fprintf(stdout, "- %d ", maxPMEnodes);
2369 fprintf(stdout, "PME-only ranks.\n Note that the automatic number of PME-only ranks and no separate PME ranks are always tested.\n");
2378 /* Get the commands we need to set up the runs from environment variables */
2379 get_program_paths(bThreads, &cmd_mpirun, &cmd_mdrun);
2380 if (bBenchmark && repeats > 0)
2382 check_mdrun_works(bThreads, cmd_mpirun, cmd_np, cmd_mdrun);
2385 /* Print some header info to file */
2387 fprintf(fp, "\n P E R F O R M A N C E R E S U L T S\n");
2389 fprintf(fp, "%s for Gromacs %s\n", ShortProgram(), GromacsVersion());
2392 fprintf(fp, "Number of ranks : %d\n", nnodes);
2393 fprintf(fp, "The mpirun command is : %s\n", cmd_mpirun);
2394 if (strcmp(procstring[0], "none") != 0)
2396 fprintf(fp, "Passing # of ranks via : %s\n", procstring[0]);
2400 fprintf(fp, "Not setting number of ranks in system call\n");
2405 fprintf(fp, "Number of threads : %d\n", nnodes);
2408 fprintf(fp, "The mdrun command is : %s\n", cmd_mdrun);
2409 fprintf(fp, "mdrun args benchmarks : %s\n", cmd_args_bench);
2410 fprintf(fp, "Benchmark steps : ");
2411 fprintf(fp, "%"GMX_PRId64, bench_nsteps);
2413 fprintf(fp, "dlb equilibration steps : %d\n", presteps);
2416 fprintf(fp, "Checkpoint time step : ");
2417 fprintf(fp, "%"GMX_PRId64, cpt_steps);
2420 fprintf(fp, "mdrun args at launchtime: %s\n", cmd_args_launch);
2422 if (new_sim_nsteps >= 0)
2425 fprintf(stderr, "Note: Simulation input file %s will have ", opt2fn("-so", NFILE, fnm));
2426 fprintf(stderr, "%"GMX_PRId64, new_sim_nsteps+cpt_steps);
2427 fprintf(stderr, " steps.\n");
2428 fprintf(fp, "Simulation steps : ");
2429 fprintf(fp, "%"GMX_PRId64, new_sim_nsteps);
2434 fprintf(fp, "Repeats for each test : %d\n", repeats);
2437 if (npme_fixed >= -1)
2439 fprintf(fp, "Fixing -npme at : %d\n", npme_fixed);
2442 fprintf(fp, "Input file : %s\n", opt2fn("-s", NFILE, fnm));
2443 fprintf(fp, " PME/PP load estimate : %g\n", guessPMEratio);
2445 /* Allocate memory for the inputinfo struct: */
2447 info->nr_inputfiles = ntprs;
2448 for (i = 0; i < ntprs; i++)
2450 snew(info->rcoulomb, ntprs);
2451 snew(info->rvdw, ntprs);
2452 snew(info->rlist, ntprs);
2453 snew(info->rlistlong, ntprs);
2454 snew(info->nkx, ntprs);
2455 snew(info->nky, ntprs);
2456 snew(info->nkz, ntprs);
2457 snew(info->fsx, ntprs);
2458 snew(info->fsy, ntprs);
2459 snew(info->fsz, ntprs);
2461 /* Make alternative tpr files to test: */
2462 snew(tpr_names, ntprs);
2463 for (i = 0; i < ntprs; i++)
2465 snew(tpr_names[i], STRLEN);
2468 /* It can be that ntprs is reduced by make_benchmark_tprs if not enough
2469 * different grids could be found. */
2470 make_benchmark_tprs(opt2fn("-s", NFILE, fnm), tpr_names, bench_nsteps+presteps,
2471 cpt_steps, rmin, rmax, bScaleRvdw, &ntprs, info, fp);
2473 /********************************************************************************/
2474 /* Main loop over all scenarios we need to test: tpr files, PME nodes, repeats */
2475 /********************************************************************************/
2476 snew(perfdata, ntprs);
2479 do_the_tests(fp, tpr_names, maxPMEnodes, minPMEnodes, npme_fixed, npmevalues_opt[0], perfdata, &pmeentries,
2480 repeats, nnodes, ntprs, bThreads, cmd_mpirun, cmd_np, cmd_mdrun,
2481 cmd_args_bench, fnm, NFILE, presteps, cpt_steps);
2483 fprintf(fp, "\nTuning took%8.1f minutes.\n", (gmx_gettime()-seconds)/60.0);
2485 /* Analyse the results and give a suggestion for optimal settings: */
2486 bKeepTPR = analyze_data(fp, opt2fn("-p", NFILE, fnm), perfdata, nnodes, ntprs, pmeentries,
2487 repeats, info, &best_tpr, &best_npme);
2489 /* Take the best-performing tpr file and enlarge nsteps to original value */
2490 if (bKeepTPR && !bOverwrite)
2492 simulation_tpr = opt2fn("-s", NFILE, fnm);
2496 simulation_tpr = opt2fn("-so", NFILE, fnm);
2497 modify_PMEsettings(bOverwrite ? (new_sim_nsteps+cpt_steps) : info->orig_sim_steps,
2498 info->orig_init_step, tpr_names[best_tpr], simulation_tpr);
2501 /* Let's get rid of the temporary benchmark input files */
2502 for (i = 0; i < ntprs; i++)
2504 fprintf(stdout, "Deleting temporary benchmark input file %s\n", tpr_names[i]);
2505 remove(tpr_names[i]);
2508 /* Now start the real simulation if the user requested it ... */
2509 launch_simulation(bLaunch, fp, bThreads, cmd_mpirun, cmd_np, cmd_mdrun,
2510 cmd_args_launch, simulation_tpr, best_npme);
2514 /* ... or simply print the performance results to screen: */
2517 finalize(opt2fn("-p", NFILE, fnm));