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.
41 #ifdef HAVE_SYS_TIME_H
47 #include "gromacs/commandline/pargs.h"
49 #include "types/commrec.h"
50 #include "gromacs/utility/smalloc.h"
53 #include "gromacs/fileio/tpxio.h"
57 #include "checkpoint.h"
63 #include "gromacs/timing/walltime_accounting.h"
66 /* Enum for situations that can occur during log file parsing, the
67 * corresponding string entries can be found in do_the_tests() in
68 * const char* ParseLog[] */
74 eParselogResetProblem,
78 eParselogLargePrimeFactor,
86 int nPMEnodes; /* number of PME-only nodes used in this test */
87 int nx, ny, nz; /* DD grid */
88 int guessPME; /* if nPMEnodes == -1, this is the guessed number of PME nodes */
89 double *Gcycles; /* This can contain more than one value if doing multiple tests */
93 float *PME_f_load; /* PME mesh/force load average*/
94 float PME_f_load_Av; /* Average average ;) ... */
95 char *mdrun_cmd_line; /* Mdrun command line used for this test */
101 int nr_inputfiles; /* The number of tpr and mdp input files */
102 gmx_int64_t orig_sim_steps; /* Number of steps to be done in the real simulation */
103 gmx_int64_t orig_init_step; /* Init step for the real simulation */
104 real *rcoulomb; /* The coulomb radii [0...nr_inputfiles] */
105 real *rvdw; /* The vdW radii */
106 real *rlist; /* Neighbourlist cutoff radius */
108 int *nkx, *nky, *nkz;
109 real *fsx, *fsy, *fsz; /* Fourierspacing in x,y,z dimension */
113 static void sep_line(FILE *fp)
115 fprintf(fp, "\n------------------------------------------------------------\n");
119 /* Wrapper for system calls */
120 static int gmx_system_call(char *command)
123 gmx_fatal(FARGS, "No calls to system(3) supported on this platform. Attempted to call:\n'%s'\n", command);
125 return ( system(command) );
130 /* Check if string starts with substring */
131 static gmx_bool str_starts(const char *string, const char *substring)
133 return ( strncmp(string, substring, strlen(substring)) == 0);
137 static void cleandata(t_perf *perfdata, int test_nr)
139 perfdata->Gcycles[test_nr] = 0.0;
140 perfdata->ns_per_day[test_nr] = 0.0;
141 perfdata->PME_f_load[test_nr] = 0.0;
147 static gmx_bool is_equal(real a, real b)
149 real diff, eps = 1.0e-7;
170 static void finalize(const char *fn_out)
176 fp = fopen(fn_out, "r");
177 fprintf(stdout, "\n\n");
179 while (fgets(buf, STRLEN-1, fp) != NULL)
181 fprintf(stdout, "%s", buf);
184 fprintf(stdout, "\n\n");
189 eFoundNothing, eFoundDDStr, eFoundAccountingStr, eFoundCycleStr
192 static int parse_logfile(const char *logfile, const char *errfile,
193 t_perf *perfdata, int test_nr, int presteps, gmx_int64_t cpt_steps,
197 char line[STRLEN], dumstring[STRLEN], dumstring2[STRLEN];
198 const char matchstrdd[] = "Domain decomposition grid";
199 const char matchstrcr[] = "resetting all time and cycle counters";
200 const char matchstrbal[] = "Average PME mesh/force load:";
201 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";
202 const char errSIG[] = "signal, stopping at the next";
205 float dum1, dum2, dum3, dum4;
208 gmx_int64_t resetsteps = -1;
209 gmx_bool bFoundResetStr = FALSE;
210 gmx_bool bResetChecked = FALSE;
213 if (!gmx_fexist(logfile))
215 fprintf(stderr, "WARNING: Could not find logfile %s.\n", logfile);
216 cleandata(perfdata, test_nr);
217 return eParselogNotFound;
220 fp = fopen(logfile, "r");
221 perfdata->PME_f_load[test_nr] = -1.0;
222 perfdata->guessPME = -1;
224 iFound = eFoundNothing;
227 iFound = eFoundDDStr; /* Skip some case statements */
230 while (fgets(line, STRLEN, fp) != NULL)
232 /* Remove leading spaces */
235 /* Check for TERM and INT signals from user: */
236 if (strstr(line, errSIG) != NULL)
239 cleandata(perfdata, test_nr);
240 return eParselogTerm;
243 /* Check whether cycle resetting worked */
244 if (presteps > 0 && !bFoundResetStr)
246 if (strstr(line, matchstrcr) != NULL)
248 sprintf(dumstring, "step %s", "%"GMX_SCNd64);
249 sscanf(line, dumstring, &resetsteps);
250 bFoundResetStr = TRUE;
251 if (resetsteps == presteps+cpt_steps)
253 bResetChecked = TRUE;
257 sprintf(dumstring, "%"GMX_PRId64, resetsteps);
258 sprintf(dumstring2, "%"GMX_PRId64, presteps+cpt_steps);
259 fprintf(stderr, "WARNING: Time step counters were reset at step %s,\n"
260 " though they were supposed to be reset at step %s!\n",
261 dumstring, dumstring2);
266 /* Look for strings that appear in a certain order in the log file: */
270 /* Look for domain decomp grid and separate PME nodes: */
271 if (str_starts(line, matchstrdd))
273 sscanf(line, "Domain decomposition grid %d x %d x %d, separate PME nodes %d",
274 &(perfdata->nx), &(perfdata->ny), &(perfdata->nz), &npme);
275 if (perfdata->nPMEnodes == -1)
277 perfdata->guessPME = npme;
279 else if (perfdata->nPMEnodes != npme)
281 gmx_fatal(FARGS, "PME nodes from command line and output file are not identical");
283 iFound = eFoundDDStr;
285 /* Catch a few errors that might have occured: */
286 else if (str_starts(line, "There is no domain decomposition for"))
289 return eParselogNoDDGrid;
291 else if (str_starts(line, "The number of nodes you selected"))
294 return eParselogLargePrimeFactor;
296 else if (str_starts(line, "reading tpx file"))
299 return eParselogTPXVersion;
301 else if (str_starts(line, "The -dd or -npme option request a parallel simulation"))
304 return eParselogNotParallel;
308 /* Look for PME mesh/force balance (not necessarily present, though) */
309 if (str_starts(line, matchstrbal))
311 sscanf(&line[strlen(matchstrbal)], "%f", &(perfdata->PME_f_load[test_nr]));
313 /* Look for matchstring */
314 if (str_starts(line, matchstring))
316 iFound = eFoundAccountingStr;
319 case eFoundAccountingStr:
320 /* Already found matchstring - look for cycle data */
321 if (str_starts(line, "Total "))
323 sscanf(line, "Total %d %lf", &procs, &(perfdata->Gcycles[test_nr]));
324 iFound = eFoundCycleStr;
328 /* Already found cycle data - look for remaining performance info and return */
329 if (str_starts(line, "Performance:"))
331 ndum = sscanf(line, "%s %f %f %f %f", dumstring, &dum1, &dum2, &dum3, &dum4);
332 /* (ns/day) is the second last entry, depending on whether GMX_DETAILED_PERF_STATS was set in print_perf(), nrnb.c */
333 perfdata->ns_per_day[test_nr] = (ndum == 5) ? dum3 : dum1;
335 if (bResetChecked || presteps == 0)
341 return eParselogResetProblem;
348 /* Close the log file */
351 /* Check why there is no performance data in the log file.
352 * Did a fatal errors occur? */
353 if (gmx_fexist(errfile))
355 fp = fopen(errfile, "r");
356 while (fgets(line, STRLEN, fp) != NULL)
358 if (str_starts(line, "Fatal error:") )
360 if (fgets(line, STRLEN, fp) != NULL)
362 fprintf(stderr, "\nWARNING: An error occured during this benchmark:\n"
366 cleandata(perfdata, test_nr);
367 return eParselogFatal;
374 fprintf(stderr, "WARNING: Could not find stderr file %s.\n", errfile);
377 /* Giving up ... we could not find out why there is no performance data in
379 fprintf(stdout, "No performance data in log file.\n");
380 cleandata(perfdata, test_nr);
382 return eParselogNoPerfData;
386 static gmx_bool analyze_data(
395 int *index_tpr, /* OUT: Nr of mdp file with best settings */
396 int *npme_optimal) /* OUT: Optimal number of PME nodes */
399 int line = 0, line_win = -1;
400 int k_win = -1, i_win = -1, winPME;
401 double s = 0.0; /* standard deviation */
404 char str_PME_f_load[13];
405 gmx_bool bCanUseOrigTPR;
406 gmx_bool bRefinedCoul, bRefinedVdW, bRefinedGrid;
412 fprintf(fp, "Summary of successful runs:\n");
413 fprintf(fp, "Line tpr PME nodes Gcycles Av. Std.dev. ns/day PME/f");
416 fprintf(fp, " DD grid");
422 for (k = 0; k < ntprs; k++)
424 for (i = 0; i < ntests; i++)
426 /* Select the right dataset: */
427 pd = &(perfdata[k][i]);
429 pd->Gcycles_Av = 0.0;
430 pd->PME_f_load_Av = 0.0;
431 pd->ns_per_day_Av = 0.0;
433 if (pd->nPMEnodes == -1)
435 sprintf(strbuf, "(%3d)", pd->guessPME);
439 sprintf(strbuf, " ");
442 /* Get the average run time of a setting */
443 for (j = 0; j < nrepeats; j++)
445 pd->Gcycles_Av += pd->Gcycles[j];
446 pd->PME_f_load_Av += pd->PME_f_load[j];
448 pd->Gcycles_Av /= nrepeats;
449 pd->PME_f_load_Av /= nrepeats;
451 for (j = 0; j < nrepeats; j++)
453 if (pd->ns_per_day[j] > 0.0)
455 pd->ns_per_day_Av += pd->ns_per_day[j];
459 /* Somehow the performance number was not aquired for this run,
460 * therefor set the average to some negative value: */
461 pd->ns_per_day_Av = -1.0f*nrepeats;
465 pd->ns_per_day_Av /= nrepeats;
468 if (pd->PME_f_load_Av > 0.0)
470 sprintf(str_PME_f_load, "%12.3f", pd->PME_f_load_Av);
474 sprintf(str_PME_f_load, "%s", " - ");
478 /* We assume we had a successful run if both averages are positive */
479 if (pd->Gcycles_Av > 0.0 && pd->ns_per_day_Av > 0.0)
481 /* Output statistics if repeats were done */
484 /* Calculate the standard deviation */
486 for (j = 0; j < nrepeats; j++)
488 s += pow( pd->Gcycles[j] - pd->Gcycles_Av, 2 );
493 fprintf(fp, "%4d %3d %4d%s %12.3f %12.3f %12.3f %s",
494 line, k, pd->nPMEnodes, strbuf, pd->Gcycles_Av, s,
495 pd->ns_per_day_Av, str_PME_f_load);
498 fprintf(fp, " %3d %3d %3d", pd->nx, pd->ny, pd->nz);
502 /* Store the index of the best run found so far in 'winner': */
503 if ( (k_win == -1) || (pd->Gcycles_Av < perfdata[k_win][i_win].Gcycles_Av) )
516 gmx_fatal(FARGS, "None of the runs was successful! Check %s for problems.", fn);
521 winPME = perfdata[k_win][i_win].nPMEnodes;
525 /* We stuck to a fixed number of PME-only nodes */
526 sprintf(strbuf, "settings No. %d", k_win);
530 /* We have optimized the number of PME-only nodes */
533 sprintf(strbuf, "%s", "the automatic number of PME nodes");
537 sprintf(strbuf, "%d PME nodes", winPME);
540 fprintf(fp, "Best performance was achieved with %s", strbuf);
541 if ((nrepeats > 1) && (ntests > 1))
543 fprintf(fp, " (see line %d)", line_win);
547 /* Only mention settings if they were modified: */
548 bRefinedCoul = !is_equal(info->rcoulomb[k_win], info->rcoulomb[0]);
549 bRefinedVdW = !is_equal(info->rvdw[k_win], info->rvdw[0] );
550 bRefinedGrid = !(info->nkx[k_win] == info->nkx[0] &&
551 info->nky[k_win] == info->nky[0] &&
552 info->nkz[k_win] == info->nkz[0]);
554 if (bRefinedCoul || bRefinedVdW || bRefinedGrid)
556 fprintf(fp, "Optimized PME settings:\n");
557 bCanUseOrigTPR = FALSE;
561 bCanUseOrigTPR = TRUE;
566 fprintf(fp, " New Coulomb radius: %f nm (was %f nm)\n", info->rcoulomb[k_win], info->rcoulomb[0]);
571 fprintf(fp, " New Van der Waals radius: %f nm (was %f nm)\n", info->rvdw[k_win], info->rvdw[0]);
576 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],
577 info->nkx[0], info->nky[0], info->nkz[0]);
580 if (bCanUseOrigTPR && ntprs > 1)
582 fprintf(fp, "and original PME settings.\n");
587 /* Return the index of the mdp file that showed the highest performance
588 * and the optimal number of PME nodes */
590 *npme_optimal = winPME;
592 return bCanUseOrigTPR;
596 /* Get the commands we need to set up the runs from environment variables */
597 static void get_program_paths(gmx_bool bThreads, char *cmd_mpirun[], char *cmd_mdrun[])
601 const char def_mpirun[] = "mpirun";
602 const char def_mdrun[] = "mdrun";
604 const char empty_mpirun[] = "";
606 /* Get the commands we need to set up the runs from environment variables */
609 if ( (cp = getenv("MPIRUN")) != NULL)
611 *cmd_mpirun = strdup(cp);
615 *cmd_mpirun = strdup(def_mpirun);
620 *cmd_mpirun = strdup(empty_mpirun);
623 if ( (cp = getenv("MDRUN" )) != NULL)
625 *cmd_mdrun = strdup(cp);
629 *cmd_mdrun = strdup(def_mdrun);
633 /* Check that the commands will run mdrun (perhaps via mpirun) by
634 * running a very quick test simulation. Requires MPI environment to
635 * be available if applicable. */
636 static void check_mdrun_works(gmx_bool bThreads,
637 const char *cmd_mpirun,
639 const char *cmd_mdrun)
641 char *command = NULL;
645 const char filename[] = "benchtest.log";
647 /* This string should always be identical to the one in copyrite.c,
648 * gmx_print_version_info() in the defined(GMX_MPI) section */
649 const char match_mpi[] = "MPI library: MPI";
650 const char match_mdrun[] = "Program: ";
651 gmx_bool bMdrun = FALSE;
652 gmx_bool bMPI = FALSE;
654 /* Run a small test to see whether mpirun + mdrun work */
655 fprintf(stdout, "Making sure that mdrun can be executed. ");
658 snew(command, strlen(cmd_mdrun) + strlen(cmd_np) + strlen(filename) + 50);
659 sprintf(command, "%s%s-version -maxh 0.001 1> %s 2>&1", cmd_mdrun, cmd_np, filename);
663 snew(command, strlen(cmd_mpirun) + strlen(cmd_np) + strlen(cmd_mdrun) + strlen(filename) + 50);
664 sprintf(command, "%s%s%s -version -maxh 0.001 1> %s 2>&1", cmd_mpirun, cmd_np, cmd_mdrun, filename);
666 fprintf(stdout, "Trying '%s' ... ", command);
667 make_backup(filename);
668 gmx_system_call(command);
670 /* Check if we find the characteristic string in the output: */
671 if (!gmx_fexist(filename))
673 gmx_fatal(FARGS, "Output from test run could not be found.");
676 fp = fopen(filename, "r");
677 /* We need to scan the whole output file, since sometimes the queuing system
678 * also writes stuff to stdout/err */
681 cp = fgets(line, STRLEN, fp);
684 if (str_starts(line, match_mdrun) )
688 if (str_starts(line, match_mpi) )
700 gmx_fatal(FARGS, "Need a threaded version of mdrun. This one\n"
702 "seems to have been compiled with MPI instead.",
710 gmx_fatal(FARGS, "Need an MPI-enabled version of mdrun. This one\n"
712 "seems to have been compiled without MPI support.",
719 gmx_fatal(FARGS, "Cannot execute mdrun. Please check %s for problems!",
723 fprintf(stdout, "passed.\n");
731 static void launch_simulation(
732 gmx_bool bLaunch, /* Should the simulation be launched? */
733 FILE *fp, /* General log file */
734 gmx_bool bThreads, /* whether to use threads */
735 char *cmd_mpirun, /* Command for mpirun */
736 char *cmd_np, /* Switch for -np or -ntmpi or empty */
737 char *cmd_mdrun, /* Command for mdrun */
738 char *args_for_mdrun, /* Arguments for mdrun */
739 const char *simulation_tpr, /* This tpr will be simulated */
740 int nPMEnodes) /* Number of PME nodes to use */
745 /* Make enough space for the system call command,
746 * (100 extra chars for -npme ... etc. options should suffice): */
747 snew(command, strlen(cmd_mpirun)+strlen(cmd_mdrun)+strlen(cmd_np)+strlen(args_for_mdrun)+strlen(simulation_tpr)+100);
749 /* Note that the -passall options requires args_for_mdrun to be at the end
750 * of the command line string */
753 sprintf(command, "%s%s-npme %d -s %s %s",
754 cmd_mdrun, cmd_np, nPMEnodes, simulation_tpr, args_for_mdrun);
758 sprintf(command, "%s%s%s -npme %d -s %s %s",
759 cmd_mpirun, cmd_np, cmd_mdrun, nPMEnodes, simulation_tpr, args_for_mdrun);
762 fprintf(fp, "%s this command line to launch the simulation:\n\n%s", bLaunch ? "Using" : "Please use", command);
766 /* Now the real thing! */
769 fprintf(stdout, "\nLaunching simulation with best parameters now.\nExecuting '%s'", command);
772 gmx_system_call(command);
777 static void modify_PMEsettings(
778 gmx_int64_t simsteps, /* Set this value as number of time steps */
779 gmx_int64_t init_step, /* Set this value as init_step */
780 const char *fn_best_tpr, /* tpr file with the best performance */
781 const char *fn_sim_tpr) /* name of tpr file to be launched */
789 read_tpx_state(fn_best_tpr, ir, &state, NULL, &mtop);
791 /* Reset nsteps and init_step to the value of the input .tpr file */
792 ir->nsteps = simsteps;
793 ir->init_step = init_step;
795 /* Write the tpr file which will be launched */
796 sprintf(buf, "Writing optimized simulation file %s with nsteps=%s.\n", fn_sim_tpr, "%"GMX_PRId64);
797 fprintf(stdout, buf, ir->nsteps);
799 write_tpx_state(fn_sim_tpr, ir, &state, &mtop);
805 #define EPME_SWITCHED(e) ((e) == eelPMESWITCH || (e) == eelPMEUSERSWITCH)
807 /* Make additional TPR files with more computational load for the
808 * direct space processors: */
809 static void make_benchmark_tprs(
810 const char *fn_sim_tpr, /* READ : User-provided tpr file */
811 char *fn_bench_tprs[], /* WRITE: Names of benchmark tpr files */
812 gmx_int64_t benchsteps, /* Number of time steps for benchmark runs */
813 gmx_int64_t statesteps, /* Step counter in checkpoint file */
814 real rmin, /* Minimal Coulomb radius */
815 real rmax, /* Maximal Coulomb radius */
816 real bScaleRvdw, /* Scale rvdw along with rcoulomb */
817 int *ntprs, /* No. of TPRs to write, each with a different
818 rcoulomb and fourierspacing */
819 t_inputinfo *info, /* Contains information about mdp file options */
820 FILE *fp) /* Write the output here */
826 real nlist_buffer; /* Thickness of the buffer regions for PME-switch potentials */
829 gmx_bool bNote = FALSE;
830 real add; /* Add this to rcoul for the next test */
831 real fac = 1.0; /* Scaling factor for Coulomb radius */
832 real fourierspacing; /* Basic fourierspacing from tpr */
835 sprintf(buf, "Making benchmark tpr file%s with %s time step%s",
836 *ntprs > 1 ? "s" : "", "%"GMX_PRId64, benchsteps > 1 ? "s" : "");
837 fprintf(stdout, buf, benchsteps);
840 sprintf(buf, " (adding %s steps from checkpoint file)", "%"GMX_PRId64);
841 fprintf(stdout, buf, statesteps);
842 benchsteps += statesteps;
844 fprintf(stdout, ".\n");
848 read_tpx_state(fn_sim_tpr, ir, &state, NULL, &mtop);
850 /* Check if some kind of PME was chosen */
851 if (EEL_PME(ir->coulombtype) == FALSE)
853 gmx_fatal(FARGS, "Can only do optimizations for simulations with %s electrostatics.",
857 /* Check if rcoulomb == rlist, which is necessary for plain PME. */
858 if ( (ir->cutoff_scheme != ecutsVERLET) &&
859 (eelPME == ir->coulombtype) && !(ir->rcoulomb == ir->rlist))
861 gmx_fatal(FARGS, "%s requires rcoulomb (%f) to be equal to rlist (%f).",
862 EELTYPE(eelPME), ir->rcoulomb, ir->rlist);
864 /* For other PME types, rcoulomb is allowed to be smaller than rlist */
865 else if (ir->rcoulomb > ir->rlist)
867 gmx_fatal(FARGS, "%s requires rcoulomb (%f) to be equal to or smaller than rlist (%f)",
868 EELTYPE(ir->coulombtype), ir->rcoulomb, ir->rlist);
871 if (bScaleRvdw && ir->rvdw != ir->rcoulomb)
873 fprintf(stdout, "NOTE: input rvdw != rcoulomb, will not scale rvdw\n");
877 /* Reduce the number of steps for the benchmarks */
878 info->orig_sim_steps = ir->nsteps;
879 ir->nsteps = benchsteps;
880 /* We must not use init_step from the input tpr file for the benchmarks */
881 info->orig_init_step = ir->init_step;
884 /* For PME-switch potentials, keep the radial distance of the buffer region */
885 nlist_buffer = ir->rlist - ir->rcoulomb;
887 /* Determine length of triclinic box vectors */
888 for (d = 0; d < DIM; d++)
891 for (i = 0; i < DIM; i++)
893 box_size[d] += state.box[d][i]*state.box[d][i];
895 box_size[d] = sqrt(box_size[d]);
898 if (ir->fourier_spacing > 0)
900 info->fsx[0] = ir->fourier_spacing;
901 info->fsy[0] = ir->fourier_spacing;
902 info->fsz[0] = ir->fourier_spacing;
906 /* Reconstruct fourierspacing per dimension from the number of grid points and box size */
907 info->fsx[0] = box_size[XX]/ir->nkx;
908 info->fsy[0] = box_size[YY]/ir->nky;
909 info->fsz[0] = box_size[ZZ]/ir->nkz;
912 /* If no value for the fourierspacing was provided on the command line, we
913 * use the reconstruction from the tpr file */
914 if (ir->fourier_spacing > 0)
916 /* Use the spacing from the tpr */
917 fourierspacing = ir->fourier_spacing;
921 /* Use the maximum observed spacing */
922 fourierspacing = max(max(info->fsx[0], info->fsy[0]), info->fsz[0]);
925 fprintf(stdout, "Calculating PME grid points on the basis of a fourierspacing of %f nm\n", fourierspacing);
927 /* For performance comparisons the number of particles is useful to have */
928 fprintf(fp, " Number of particles : %d\n", mtop.natoms);
930 /* Print information about settings of which some are potentially modified: */
931 fprintf(fp, " Coulomb type : %s\n", EELTYPE(ir->coulombtype));
932 fprintf(fp, " Grid spacing x y z : %f %f %f\n",
933 box_size[XX]/ir->nkx, box_size[YY]/ir->nky, box_size[ZZ]/ir->nkz);
934 fprintf(fp, " Van der Waals type : %s\n", EVDWTYPE(ir->vdwtype));
935 if (ir_vdw_switched(ir))
937 fprintf(fp, " rvdw_switch : %f nm\n", ir->rvdw_switch);
939 if (EPME_SWITCHED(ir->coulombtype))
941 fprintf(fp, " rlist : %f nm\n", ir->rlist);
943 if (ir->rlistlong != max_cutoff(ir->rvdw, ir->rcoulomb))
945 fprintf(fp, " rlistlong : %f nm\n", ir->rlistlong);
948 /* Print a descriptive line about the tpr settings tested */
949 fprintf(fp, "\nWill try these real/reciprocal workload settings:\n");
950 fprintf(fp, " No. scaling rcoulomb");
951 fprintf(fp, " nkx nky nkz");
952 fprintf(fp, " spacing");
953 if (evdwCUT == ir->vdwtype)
955 fprintf(fp, " rvdw");
957 if (EPME_SWITCHED(ir->coulombtype))
959 fprintf(fp, " rlist");
961 if (ir->rlistlong != max_cutoff(ir->rlist, max_cutoff(ir->rvdw, ir->rcoulomb)) )
963 fprintf(fp, " rlistlong");
965 fprintf(fp, " tpr file\n");
967 /* Loop to create the requested number of tpr input files */
968 for (j = 0; j < *ntprs; j++)
970 /* The first .tpr is the provided one, just need to modify nsteps,
971 * so skip the following block */
974 /* Determine which Coulomb radii rc to use in the benchmarks */
975 add = (rmax-rmin)/(*ntprs-1);
976 if (is_equal(rmin, info->rcoulomb[0]))
978 ir->rcoulomb = rmin + j*add;
980 else if (is_equal(rmax, info->rcoulomb[0]))
982 ir->rcoulomb = rmin + (j-1)*add;
986 /* rmin != rcoul != rmax, ergo test between rmin and rmax */
987 add = (rmax-rmin)/(*ntprs-2);
988 ir->rcoulomb = rmin + (j-1)*add;
991 /* Determine the scaling factor fac */
992 fac = ir->rcoulomb/info->rcoulomb[0];
994 /* Scale the Fourier grid spacing */
995 ir->nkx = ir->nky = ir->nkz = 0;
996 calc_grid(NULL, state.box, fourierspacing*fac, &ir->nkx, &ir->nky, &ir->nkz);
998 /* Adjust other radii since various conditions need to be fulfilled */
999 if (eelPME == ir->coulombtype)
1001 /* plain PME, rcoulomb must be equal to rlist */
1002 ir->rlist = ir->rcoulomb;
1006 /* rlist must be >= rcoulomb, we keep the size of the buffer region */
1007 ir->rlist = ir->rcoulomb + nlist_buffer;
1010 if (bScaleRvdw && evdwCUT == ir->vdwtype)
1012 if (ecutsVERLET == ir->cutoff_scheme)
1014 /* With Verlet, the van der Waals radius must always equal the Coulomb radius */
1015 ir->rvdw = ir->rcoulomb;
1019 /* For vdw cutoff, rvdw >= rlist */
1020 ir->rvdw = max(info->rvdw[0], ir->rlist);
1024 ir->rlistlong = max_cutoff(ir->rlist, max_cutoff(ir->rvdw, ir->rcoulomb));
1026 } /* end of "if (j != 0)" */
1028 /* for j==0: Save the original settings
1029 * for j >0: Save modified radii and Fourier grids */
1030 info->rcoulomb[j] = ir->rcoulomb;
1031 info->rvdw[j] = ir->rvdw;
1032 info->nkx[j] = ir->nkx;
1033 info->nky[j] = ir->nky;
1034 info->nkz[j] = ir->nkz;
1035 info->rlist[j] = ir->rlist;
1036 info->rlistlong[j] = ir->rlistlong;
1037 info->fsx[j] = fac*fourierspacing;
1038 info->fsy[j] = fac*fourierspacing;
1039 info->fsz[j] = fac*fourierspacing;
1041 /* Write the benchmark tpr file */
1042 strncpy(fn_bench_tprs[j], fn_sim_tpr, strlen(fn_sim_tpr)-strlen(".tpr"));
1043 sprintf(buf, "_bench%.2d.tpr", j);
1044 strcat(fn_bench_tprs[j], buf);
1045 fprintf(stdout, "Writing benchmark tpr %s with nsteps=", fn_bench_tprs[j]);
1046 fprintf(stdout, "%"GMX_PRId64, ir->nsteps);
1049 fprintf(stdout, ", scaling factor %f\n", fac);
1053 fprintf(stdout, ", unmodified settings\n");
1056 write_tpx_state(fn_bench_tprs[j], ir, &state, &mtop);
1058 /* Write information about modified tpr settings to log file */
1059 fprintf(fp, "%4d%10f%10f", j, fac, ir->rcoulomb);
1060 fprintf(fp, "%5d%5d%5d", ir->nkx, ir->nky, ir->nkz);
1061 fprintf(fp, " %9f ", info->fsx[j]);
1062 if (evdwCUT == ir->vdwtype)
1064 fprintf(fp, "%10f", ir->rvdw);
1066 if (EPME_SWITCHED(ir->coulombtype))
1068 fprintf(fp, "%10f", ir->rlist);
1070 if (info->rlistlong[0] != max_cutoff(info->rlist[0], max_cutoff(info->rvdw[0], info->rcoulomb[0])) )
1072 fprintf(fp, "%10f", ir->rlistlong);
1074 fprintf(fp, " %-14s\n", fn_bench_tprs[j]);
1076 /* Make it clear to the user that some additional settings were modified */
1077 if (!is_equal(ir->rvdw, info->rvdw[0])
1078 || !is_equal(ir->rlistlong, info->rlistlong[0]) )
1085 fprintf(fp, "\nNote that in addition to the Coulomb radius and the Fourier grid\n"
1086 "other input settings were also changed (see table above).\n"
1087 "Please check if the modified settings are appropriate.\n");
1095 /* Rename the files we want to keep to some meaningful filename and
1096 * delete the rest */
1097 static void cleanup(const t_filenm *fnm, int nfile, int k, int nnodes,
1098 int nPMEnodes, int nr, gmx_bool bKeepStderr)
1100 char numstring[STRLEN];
1101 char newfilename[STRLEN];
1102 const char *fn = NULL;
1107 fprintf(stdout, "Cleaning up, deleting benchmark temp files ...\n");
1109 for (i = 0; i < nfile; i++)
1111 opt = (char *)fnm[i].opt;
1112 if (strcmp(opt, "-p") == 0)
1114 /* do nothing; keep this file */
1117 else if (strcmp(opt, "-bg") == 0)
1119 /* Give the log file a nice name so one can later see which parameters were used */
1120 numstring[0] = '\0';
1123 sprintf(numstring, "_%d", nr);
1125 sprintf(newfilename, "%s_no%d_np%d_npme%d%s", opt2fn("-bg", nfile, fnm), k, nnodes, nPMEnodes, numstring);
1126 if (gmx_fexist(opt2fn("-bg", nfile, fnm)))
1128 fprintf(stdout, "renaming log file to %s\n", newfilename);
1129 make_backup(newfilename);
1130 rename(opt2fn("-bg", nfile, fnm), newfilename);
1133 else if (strcmp(opt, "-err") == 0)
1135 /* This file contains the output of stderr. We want to keep it in
1136 * cases where there have been problems. */
1137 fn = opt2fn(opt, nfile, fnm);
1138 numstring[0] = '\0';
1141 sprintf(numstring, "_%d", nr);
1143 sprintf(newfilename, "%s_no%d_np%d_npme%d%s", fn, k, nnodes, nPMEnodes, numstring);
1148 fprintf(stdout, "Saving stderr output in %s\n", newfilename);
1149 make_backup(newfilename);
1150 rename(fn, newfilename);
1154 fprintf(stdout, "Deleting %s\n", fn);
1159 /* Delete the files which are created for each benchmark run: (options -b*) */
1160 else if ( (0 == strncmp(opt, "-b", 2)) && (opt2bSet(opt, nfile, fnm) || !is_optional(&fnm[i])) )
1162 fn = opt2fn(opt, nfile, fnm);
1165 fprintf(stdout, "Deleting %s\n", fn);
1173 /* Returns the largest common factor of n1 and n2 */
1174 static int largest_common_factor(int n1, int n2)
1179 for (factor = nmax; factor > 0; factor--)
1181 if (0 == (n1 % factor) && 0 == (n2 % factor) )
1186 return 0; /* one for the compiler */
1190 eNpmeAuto, eNpmeAll, eNpmeReduced, eNpmeSubset, eNpmeNr
1193 /* Create a list of numbers of PME nodes to test */
1194 static void make_npme_list(
1195 const char *npmevalues_opt, /* Make a complete list with all
1196 * possibilities or a short list that keeps only
1197 * reasonable numbers of PME nodes */
1198 int *nentries, /* Number of entries we put in the nPMEnodes list */
1199 int *nPMEnodes[], /* Each entry contains the value for -npme */
1200 int nnodes, /* Total number of nodes to do the tests on */
1201 int minPMEnodes, /* Minimum number of PME nodes */
1202 int maxPMEnodes) /* Maximum number of PME nodes */
1205 int min_factor = 1; /* We request that npp and npme have this minimal
1206 * largest common factor (depends on npp) */
1207 int nlistmax; /* Max. list size */
1208 int nlist; /* Actual number of entries in list */
1212 /* Do we need to check all possible values for -npme or is a reduced list enough? */
1213 if (0 == strcmp(npmevalues_opt, "all") )
1217 else if (0 == strcmp(npmevalues_opt, "subset") )
1219 eNPME = eNpmeSubset;
1221 else /* "auto" or "range" */
1227 else if (nnodes < 128)
1229 eNPME = eNpmeReduced;
1233 eNPME = eNpmeSubset;
1237 /* Calculate how many entries we could possibly have (in case of -npme all) */
1240 nlistmax = maxPMEnodes - minPMEnodes + 3;
1241 if (0 == minPMEnodes)
1251 /* Now make the actual list which is at most of size nlist */
1252 snew(*nPMEnodes, nlistmax);
1253 nlist = 0; /* start counting again, now the real entries in the list */
1254 for (i = 0; i < nlistmax - 2; i++)
1256 npme = maxPMEnodes - i;
1267 /* For 2d PME we want a common largest factor of at least the cube
1268 * root of the number of PP nodes */
1269 min_factor = (int) pow(npp, 1.0/3.0);
1272 gmx_fatal(FARGS, "Unknown option for eNPME in make_npme_list");
1275 if (largest_common_factor(npp, npme) >= min_factor)
1277 (*nPMEnodes)[nlist] = npme;
1281 /* We always test 0 PME nodes and the automatic number */
1282 *nentries = nlist + 2;
1283 (*nPMEnodes)[nlist ] = 0;
1284 (*nPMEnodes)[nlist+1] = -1;
1286 fprintf(stderr, "Will try the following %d different values for -npme:\n", *nentries);
1287 for (i = 0; i < *nentries-1; i++)
1289 fprintf(stderr, "%d, ", (*nPMEnodes)[i]);
1291 fprintf(stderr, "and %d (auto).\n", (*nPMEnodes)[*nentries-1]);
1295 /* Allocate memory to store the performance data */
1296 static void init_perfdata(t_perf *perfdata[], int ntprs, int datasets, int repeats)
1301 for (k = 0; k < ntprs; k++)
1303 snew(perfdata[k], datasets);
1304 for (i = 0; i < datasets; i++)
1306 for (j = 0; j < repeats; j++)
1308 snew(perfdata[k][i].Gcycles, repeats);
1309 snew(perfdata[k][i].ns_per_day, repeats);
1310 snew(perfdata[k][i].PME_f_load, repeats);
1317 /* Check for errors on mdrun -h */
1318 static void make_sure_it_runs(char *mdrun_cmd_line, int length, FILE *fp)
1320 char *command, *msg;
1324 snew(command, length + 15);
1325 snew(msg, length + 500);
1327 fprintf(stdout, "Making sure the benchmarks can be executed ...\n");
1328 /* FIXME: mdrun -h no longer actually does anything useful.
1329 * It unconditionally prints the help, ignoring all other options. */
1330 sprintf(command, "%s-h -quiet", mdrun_cmd_line);
1331 ret = gmx_system_call(command);
1335 /* To prevent confusion, do not again issue a gmx_fatal here since we already
1336 * get the error message from mdrun itself */
1337 sprintf(msg, "Cannot run the benchmark simulations! Please check the error message of\n"
1338 "mdrun for the source of the problem. Did you provide a command line\n"
1339 "argument that neither g_tune_pme nor mdrun understands? Offending command:\n"
1340 "\n%s\n\n", command);
1342 fprintf(stderr, "%s", msg);
1344 fprintf(fp, "%s", msg);
1354 static void do_the_tests(
1355 FILE *fp, /* General g_tune_pme output file */
1356 char **tpr_names, /* Filenames of the input files to test */
1357 int maxPMEnodes, /* Max fraction of nodes to use for PME */
1358 int minPMEnodes, /* Min fraction of nodes to use for PME */
1359 int npme_fixed, /* If >= -1, test fixed number of PME
1361 const char *npmevalues_opt, /* Which -npme values should be tested */
1362 t_perf **perfdata, /* Here the performace data is stored */
1363 int *pmeentries, /* Entries in the nPMEnodes list */
1364 int repeats, /* Repeat each test this often */
1365 int nnodes, /* Total number of nodes = nPP + nPME */
1366 int nr_tprs, /* Total number of tpr files to test */
1367 gmx_bool bThreads, /* Threads or MPI? */
1368 char *cmd_mpirun, /* mpirun command string */
1369 char *cmd_np, /* "-np", "-n", whatever mpirun needs */
1370 char *cmd_mdrun, /* mdrun command string */
1371 char *cmd_args_bench, /* arguments for mdrun in a string */
1372 const t_filenm *fnm, /* List of filenames from command line */
1373 int nfile, /* Number of files specified on the cmdl. */
1374 int presteps, /* DLB equilibration steps, is checked */
1375 gmx_int64_t cpt_steps) /* Time step counter in the checkpoint */
1377 int i, nr, k, ret, count = 0, totaltests;
1378 int *nPMEnodes = NULL;
1381 char *command, *cmd_stub;
1383 gmx_bool bResetProblem = FALSE;
1384 gmx_bool bFirst = TRUE;
1387 /* This string array corresponds to the eParselog enum type at the start
1389 const char* ParseLog[] = {
1391 "Logfile not found!",
1392 "No timings, logfile truncated?",
1393 "Run was terminated.",
1394 "Counters were not reset properly.",
1395 "No DD grid found for these settings.",
1396 "TPX version conflict!",
1397 "mdrun was not started in parallel!",
1398 "Number of PP nodes has a prime factor that is too large.",
1401 char str_PME_f_load[13];
1404 /* Allocate space for the mdrun command line. 100 extra characters should
1405 be more than enough for the -npme etcetera arguments */
1406 cmdline_length = strlen(cmd_mpirun)
1409 + strlen(cmd_args_bench)
1410 + strlen(tpr_names[0]) + 100;
1411 snew(command, cmdline_length);
1412 snew(cmd_stub, cmdline_length);
1414 /* Construct the part of the command line that stays the same for all tests: */
1417 sprintf(cmd_stub, "%s%s", cmd_mdrun, cmd_np);
1421 sprintf(cmd_stub, "%s%s%s ", cmd_mpirun, cmd_np, cmd_mdrun);
1424 /* Create a list of numbers of PME nodes to test */
1425 if (npme_fixed < -1)
1427 make_npme_list(npmevalues_opt, pmeentries, &nPMEnodes,
1428 nnodes, minPMEnodes, maxPMEnodes);
1434 nPMEnodes[0] = npme_fixed;
1435 fprintf(stderr, "Will use a fixed number of %d PME-only nodes.\n", nPMEnodes[0]);
1440 fprintf(fp, "\nNo benchmarks done since number of repeats (-r) is 0.\n");
1442 finalize(opt2fn("-p", nfile, fnm));
1446 /* Allocate one dataset for each tpr input file: */
1447 init_perfdata(perfdata, nr_tprs, *pmeentries, repeats);
1449 /*****************************************/
1450 /* Main loop over all tpr files to test: */
1451 /*****************************************/
1452 totaltests = nr_tprs*(*pmeentries)*repeats;
1453 for (k = 0; k < nr_tprs; k++)
1455 fprintf(fp, "\nIndividual timings for input file %d (%s):\n", k, tpr_names[k]);
1456 fprintf(fp, "PME nodes Gcycles ns/day PME/f Remark\n");
1457 /* Loop over various numbers of PME nodes: */
1458 for (i = 0; i < *pmeentries; i++)
1460 pd = &perfdata[k][i];
1462 /* Loop over the repeats for each scenario: */
1463 for (nr = 0; nr < repeats; nr++)
1465 pd->nPMEnodes = nPMEnodes[i];
1467 /* Add -npme and -s to the command line and save it. Note that
1468 * the -passall (if set) options requires cmd_args_bench to be
1469 * at the end of the command line string */
1470 snew(pd->mdrun_cmd_line, cmdline_length);
1471 sprintf(pd->mdrun_cmd_line, "%s-npme %d -s %s %s",
1472 cmd_stub, pd->nPMEnodes, tpr_names[k], cmd_args_bench);
1474 /* To prevent that all benchmarks fail due to a show-stopper argument
1475 * on the mdrun command line, we make a quick check with mdrun -h first */
1478 make_sure_it_runs(pd->mdrun_cmd_line, cmdline_length, fp);
1482 /* Do a benchmark simulation: */
1485 sprintf(buf, ", pass %d/%d", nr+1, repeats);
1491 fprintf(stdout, "\n=== Progress %2.0f%%, tpr %d/%d, run %d/%d%s:\n",
1492 (100.0*count)/totaltests,
1493 k+1, nr_tprs, i+1, *pmeentries, buf);
1494 make_backup(opt2fn("-err", nfile, fnm));
1495 sprintf(command, "%s 1> /dev/null 2>%s", pd->mdrun_cmd_line, opt2fn("-err", nfile, fnm));
1496 fprintf(stdout, "%s\n", pd->mdrun_cmd_line);
1497 gmx_system_call(command);
1499 /* Collect the performance data from the log file; also check stderr
1500 * for fatal errors */
1501 ret = parse_logfile(opt2fn("-bg", nfile, fnm), opt2fn("-err", nfile, fnm),
1502 pd, nr, presteps, cpt_steps, nnodes);
1503 if ((presteps > 0) && (ret == eParselogResetProblem))
1505 bResetProblem = TRUE;
1508 if (-1 == pd->nPMEnodes)
1510 sprintf(buf, "(%3d)", pd->guessPME);
1518 if (pd->PME_f_load[nr] > 0.0)
1520 sprintf(str_PME_f_load, "%12.3f", pd->PME_f_load[nr]);
1524 sprintf(str_PME_f_load, "%s", " - ");
1527 /* Write the data we got to disk */
1528 fprintf(fp, "%4d%s %12.3f %12.3f %s %s", pd->nPMEnodes,
1529 buf, pd->Gcycles[nr], pd->ns_per_day[nr], str_PME_f_load, ParseLog[ret]);
1530 if (!(ret == eParselogOK || ret == eParselogNoDDGrid || ret == eParselogNotFound) )
1532 fprintf(fp, " Check %s file for problems.", ret == eParselogFatal ? "err" : "log");
1538 /* Do some cleaning up and delete the files we do not need any more */
1539 cleanup(fnm, nfile, k, nnodes, pd->nPMEnodes, nr, ret == eParselogFatal);
1541 /* If the first run with this number of processors already failed, do not try again: */
1542 if (pd->Gcycles[0] <= 0.0 && repeats > 1)
1544 fprintf(stdout, "Skipping remaining passes of unsuccessful setting, see log file for details.\n");
1545 count += repeats-(nr+1);
1548 } /* end of repeats loop */
1549 } /* end of -npme loop */
1550 } /* end of tpr file loop */
1555 fprintf(fp, "WARNING: The cycle and time step counters could not be reset properly. ");
1563 static void check_input(
1570 real maxPMEfraction,
1571 real minPMEfraction,
1573 gmx_int64_t bench_nsteps,
1574 const t_filenm *fnm,
1584 /* Make sure the input file exists */
1585 if (!gmx_fexist(opt2fn("-s", nfile, fnm)))
1587 gmx_fatal(FARGS, "File %s not found.", opt2fn("-s", nfile, fnm));
1590 /* Make sure that the checkpoint file is not overwritten during benchmarking */
1591 if ( (0 == strcmp(opt2fn("-cpi", nfile, fnm), opt2fn("-bcpo", nfile, fnm)) ) && (sim_part > 1) )
1593 gmx_fatal(FARGS, "Checkpoint input (-cpi) and benchmark checkpoint output (-bcpo) files must not be identical.\n"
1594 "The checkpoint input file must not be overwritten during the benchmarks.\n");
1597 /* Make sure that repeats is >= 0 (if == 0, only write tpr files) */
1600 gmx_fatal(FARGS, "Number of repeats < 0!");
1603 /* Check number of nodes */
1606 gmx_fatal(FARGS, "Number of nodes/threads must be a positive integer.");
1609 /* Automatically choose -ntpr if not set */
1619 /* Set a reasonable scaling factor for rcoulomb */
1622 *rmax = rcoulomb * 1.2;
1625 fprintf(stderr, "Will test %d tpr file%s.\n", *ntprs, *ntprs == 1 ? "" : "s");
1631 fprintf(stderr, "Note: Choose ntpr>1 to shift PME load between real and reciprocal space.\n");
1635 /* Make shure that rmin <= rcoulomb <= rmax */
1644 if (!(*rmin <= *rmax) )
1646 gmx_fatal(FARGS, "Please choose the Coulomb radii such that rmin <= rmax.\n"
1647 "rmin = %g, rmax = %g, actual rcoul from .tpr file = %g\n", *rmin, *rmax, rcoulomb);
1649 /* Add test scenarios if rmin or rmax were set */
1652 if (!is_equal(*rmin, rcoulomb) && (*ntprs == 1) )
1655 fprintf(stderr, "NOTE: Setting -rmin to %g changed -ntpr to %d\n",
1658 if (!is_equal(*rmax, rcoulomb) && (*ntprs == 1) )
1661 fprintf(stderr, "NOTE: Setting -rmax to %g changed -ntpr to %d\n",
1666 /* If one of rmin, rmax is set, we need 2 tpr files at minimum */
1667 if (!is_equal(*rmax, rcoulomb) || !is_equal(*rmin, rcoulomb) )
1669 *ntprs = max(*ntprs, 2);
1672 /* If both rmin, rmax are set, we need 3 tpr files at minimum */
1673 if (!is_equal(*rmax, rcoulomb) && !is_equal(*rmin, rcoulomb) )
1675 *ntprs = max(*ntprs, 3);
1680 fprintf(stderr, "NOTE: Your rmin, rmax setting changed -ntpr to %d\n", *ntprs);
1685 if (is_equal(*rmin, rcoulomb) && is_equal(rcoulomb, *rmax)) /* We have just a single rc */
1687 fprintf(stderr, "WARNING: Resetting -ntpr to 1 since no Coulomb radius scaling is requested.\n"
1688 "Please set rmin < rmax to test Coulomb radii in the [rmin, rmax] interval\n"
1689 "with correspondingly adjusted PME grid settings\n");
1694 /* Check whether max and min fraction are within required values */
1695 if (maxPMEfraction > 0.5 || maxPMEfraction < 0)
1697 gmx_fatal(FARGS, "-max must be between 0 and 0.5");
1699 if (minPMEfraction > 0.5 || minPMEfraction < 0)
1701 gmx_fatal(FARGS, "-min must be between 0 and 0.5");
1703 if (maxPMEfraction < minPMEfraction)
1705 gmx_fatal(FARGS, "-max must be larger or equal to -min");
1708 /* Check whether the number of steps - if it was set - has a reasonable value */
1709 if (bench_nsteps < 0)
1711 gmx_fatal(FARGS, "Number of steps must be positive.");
1714 if (bench_nsteps > 10000 || bench_nsteps < 100)
1716 fprintf(stderr, "WARNING: steps=");
1717 fprintf(stderr, "%"GMX_PRId64, bench_nsteps);
1718 fprintf(stderr, ". Are you sure you want to perform so %s steps for each benchmark?\n", (bench_nsteps < 100) ? "few" : "many");
1723 gmx_fatal(FARGS, "Cannot have a negative number of presteps.\n");
1726 /* Check for rcoulomb scaling if more than one .tpr file is tested */
1729 if (*rmin/rcoulomb < 0.75 || *rmax/rcoulomb > 1.25)
1731 fprintf(stderr, "WARNING: Applying extreme scaling factor. I hope you know what you are doing.\n");
1735 /* If a fixed number of PME nodes is set we do rcoulomb and PME gird tuning
1736 * only. We need to check whether the requested number of PME-only nodes
1738 if (npme_fixed > -1)
1740 /* No more than 50% of all nodes can be assigned as PME-only nodes. */
1741 if (2*npme_fixed > nnodes)
1743 gmx_fatal(FARGS, "Cannot have more than %d PME-only nodes for a total of %d nodes (you chose %d).\n",
1744 nnodes/2, nnodes, npme_fixed);
1746 if ((npme_fixed > 0) && (5*npme_fixed < nnodes))
1748 fprintf(stderr, "WARNING: Only %g percent of the nodes are assigned as PME-only nodes.\n",
1749 100.0*((real)npme_fixed / (real)nnodes));
1751 if (opt2parg_bSet("-min", npargs, pa) || opt2parg_bSet("-max", npargs, pa))
1753 fprintf(stderr, "NOTE: The -min, -max, and -npme options have no effect when a\n"
1754 " fixed number of PME-only nodes is requested with -fix.\n");
1760 /* Returns TRUE when "opt" is needed at launch time */
1761 static gmx_bool is_launch_file(char *opt, gmx_bool bSet)
1763 if (0 == strncmp(opt, "-swap", 5))
1768 /* Apart from the input .tpr and the output log files we need all options that
1769 * were set on the command line and that do not start with -b */
1770 if (0 == strncmp(opt, "-b", 2) || 0 == strncmp(opt, "-s", 2)
1771 || 0 == strncmp(opt, "-err", 4) || 0 == strncmp(opt, "-p", 2) )
1780 /* Returns TRUE when "opt" defines a file which is needed for the benchmarks runs */
1781 static gmx_bool is_bench_file(char *opt, gmx_bool bSet, gmx_bool bOptional, gmx_bool bIsOutput)
1783 /* Apart from the input .tpr, all files starting with "-b" are for
1784 * _b_enchmark files exclusively */
1785 if (0 == strncmp(opt, "-s", 2))
1790 if (0 == strncmp(opt, "-b", 2) || 0 == strncmp(opt, "-s", 2))
1792 if (!bOptional || bSet)
1809 if (bSet) /* These are additional input files like -cpi -ei */
1822 /* Adds 'buf' to 'str' */
1823 static void add_to_string(char **str, char *buf)
1828 len = strlen(*str) + strlen(buf) + 1;
1834 /* Create the command line for the benchmark as well as for the real run */
1835 static void create_command_line_snippets(
1836 gmx_bool bAppendFiles,
1837 gmx_bool bKeepAndNumCPT,
1838 gmx_bool bResetHWay,
1842 char *cmd_args_bench[], /* command line arguments for benchmark runs */
1843 char *cmd_args_launch[], /* command line arguments for simulation run */
1844 char extra_args[]) /* Add this to the end of the command line */
1849 char strbuf[STRLEN];
1852 /* strlen needs at least '\0' as a string: */
1853 snew(*cmd_args_bench, 1);
1854 snew(*cmd_args_launch, 1);
1855 *cmd_args_launch[0] = '\0';
1856 *cmd_args_bench[0] = '\0';
1859 /*******************************************/
1860 /* 1. Process other command line arguments */
1861 /*******************************************/
1864 /* Add equilibration steps to benchmark options */
1865 sprintf(strbuf, "-resetstep %d ", presteps);
1866 add_to_string(cmd_args_bench, strbuf);
1868 /* These switches take effect only at launch time */
1869 if (FALSE == bAppendFiles)
1871 add_to_string(cmd_args_launch, "-noappend ");
1875 add_to_string(cmd_args_launch, "-cpnum ");
1879 add_to_string(cmd_args_launch, "-resethway ");
1882 /********************/
1883 /* 2. Process files */
1884 /********************/
1885 for (i = 0; i < nfile; i++)
1887 opt = (char *)fnm[i].opt;
1888 name = opt2fn(opt, nfile, fnm);
1890 /* Strbuf contains the options, now let's sort out where we need that */
1891 sprintf(strbuf, "%s %s ", opt, name);
1893 if (is_bench_file(opt, opt2bSet(opt, nfile, fnm), is_optional(&fnm[i]), is_output(&fnm[i])) )
1895 /* All options starting with -b* need the 'b' removed,
1896 * therefore overwrite strbuf */
1897 if (0 == strncmp(opt, "-b", 2))
1899 sprintf(strbuf, "-%s %s ", &opt[2], name);
1902 add_to_string(cmd_args_bench, strbuf);
1905 if (is_launch_file(opt, opt2bSet(opt, nfile, fnm)) )
1907 add_to_string(cmd_args_launch, strbuf);
1911 add_to_string(cmd_args_bench, extra_args);
1912 add_to_string(cmd_args_launch, extra_args);
1916 /* Set option opt */
1917 static void setopt(const char *opt, int nfile, t_filenm fnm[])
1921 for (i = 0; (i < nfile); i++)
1923 if (strcmp(opt, fnm[i].opt) == 0)
1925 fnm[i].flag |= ffSET;
1931 /* This routine inspects the tpr file and ...
1932 * 1. checks for output files that get triggered by a tpr option. These output
1933 * files are marked as 'set' to allow for a proper cleanup after each
1935 * 2. returns the PME:PP load ratio
1936 * 3. returns rcoulomb from the tpr */
1937 static float inspect_tpr(int nfile, t_filenm fnm[], real *rcoulomb)
1939 gmx_bool bPull; /* Is pulling requested in .tpr file? */
1940 gmx_bool bTpi; /* Is test particle insertion requested? */
1941 gmx_bool bFree; /* Is a free energy simulation requested? */
1942 gmx_bool bNM; /* Is a normal mode analysis requested? */
1943 gmx_bool bSwap; /* Is water/ion position swapping requested? */
1949 /* Check tpr file for options that trigger extra output files */
1950 read_tpx_state(opt2fn("-s", nfile, fnm), &ir, &state, NULL, &mtop);
1951 bPull = (epullNO != ir.ePull);
1952 bFree = (efepNO != ir.efep );
1953 bNM = (eiNM == ir.eI );
1954 bSwap = (eswapNO != ir.eSwapCoords);
1955 bTpi = EI_TPI(ir.eI);
1957 /* Set these output files on the tuning command-line */
1960 setopt("-pf", nfile, fnm);
1961 setopt("-px", nfile, fnm);
1965 setopt("-dhdl", nfile, fnm);
1969 setopt("-tpi", nfile, fnm);
1970 setopt("-tpid", nfile, fnm);
1974 setopt("-mtx", nfile, fnm);
1978 setopt("-swap", nfile, fnm);
1981 *rcoulomb = ir.rcoulomb;
1983 /* Return the estimate for the number of PME nodes */
1984 return pme_load_estimate(&mtop, &ir, state.box);
1988 static void couple_files_options(int nfile, t_filenm fnm[])
1991 gmx_bool bSet, bBench;
1996 for (i = 0; i < nfile; i++)
1998 opt = (char *)fnm[i].opt;
1999 bSet = ((fnm[i].flag & ffSET) != 0);
2000 bBench = (0 == strncmp(opt, "-b", 2));
2002 /* Check optional files */
2003 /* If e.g. -eo is set, then -beo also needs to be set */
2004 if (is_optional(&fnm[i]) && bSet && !bBench)
2006 sprintf(buf, "-b%s", &opt[1]);
2007 setopt(buf, nfile, fnm);
2009 /* If -beo is set, then -eo also needs to be! */
2010 if (is_optional(&fnm[i]) && bSet && bBench)
2012 sprintf(buf, "-%s", &opt[2]);
2013 setopt(buf, nfile, fnm);
2019 #define BENCHSTEPS (1000)
2021 int gmx_tune_pme(int argc, char *argv[])
2023 const char *desc[] = {
2024 "For a given number [TT]-np[tt] or [TT]-ntmpi[tt] of processors/threads, [THISMODULE] systematically",
2025 "times [gmx-mdrun] with various numbers of PME-only nodes and determines",
2026 "which setting is fastest. It will also test whether performance can",
2027 "be enhanced by shifting load from the reciprocal to the real space",
2028 "part of the Ewald sum. ",
2029 "Simply pass your [TT].tpr[tt] file to [THISMODULE] together with other options",
2030 "for [gmx-mdrun] as needed.[PAR]",
2031 "Which executables are used can be set in the environment variables",
2032 "MPIRUN and MDRUN. If these are not present, 'mpirun' and 'mdrun'",
2033 "will be used as defaults. Note that for certain MPI frameworks you",
2034 "need to provide a machine- or hostfile. This can also be passed",
2035 "via the MPIRUN variable, e.g.[PAR]",
2036 "[TT]export MPIRUN=\"/usr/local/mpirun -machinefile hosts\"[tt][PAR]",
2037 "Please call [THISMODULE] with the normal options you would pass to",
2038 "[gmx-mdrun] and add [TT]-np[tt] for the number of processors to perform the",
2039 "tests on, or [TT]-ntmpi[tt] for the number of threads. You can also add [TT]-r[tt]",
2040 "to repeat each test several times to get better statistics. [PAR]",
2041 "[THISMODULE] can test various real space / reciprocal space workloads",
2042 "for you. With [TT]-ntpr[tt] you control how many extra [TT].tpr[tt] files will be",
2043 "written with enlarged cutoffs and smaller Fourier grids respectively.",
2044 "Typically, the first test (number 0) will be with the settings from the input",
2045 "[TT].tpr[tt] file; the last test (number [TT]ntpr[tt]) will have the Coulomb cutoff",
2046 "specified by [TT]-rmax[tt] with a somwhat smaller PME grid at the same time. ",
2047 "In this last test, the Fourier spacing is multiplied with [TT]rmax[tt]/rcoulomb. ",
2048 "The remaining [TT].tpr[tt] files will have equally-spaced Coulomb radii (and Fourier "
2049 "spacings) between these extremes. [BB]Note[bb] that you can set [TT]-ntpr[tt] to 1",
2050 "if you just seek the optimal number of PME-only nodes; in that case",
2051 "your input [TT].tpr[tt] file will remain unchanged.[PAR]",
2052 "For the benchmark runs, the default of 1000 time steps should suffice for most",
2053 "MD systems. The dynamic load balancing needs about 100 time steps",
2054 "to adapt to local load imbalances, therefore the time step counters",
2055 "are by default reset after 100 steps. For large systems (>1M atoms), as well as ",
2056 "for a higher accuarcy of the measurements, you should set [TT]-resetstep[tt] to a higher value.",
2057 "From the 'DD' load imbalance entries in the md.log output file you",
2058 "can tell after how many steps the load is sufficiently balanced. Example call:[PAR]"
2059 "[TT]gmx tune_pme -np 64 -s protein.tpr -launch[tt][PAR]",
2060 "After calling [gmx-mdrun] several times, detailed performance information",
2061 "is available in the output file [TT]perf.out[tt].",
2062 "[BB]Note[bb] that during the benchmarks, a couple of temporary files are written",
2063 "(options [TT]-b[tt]*), these will be automatically deleted after each test.[PAR]",
2064 "If you want the simulation to be started automatically with the",
2065 "optimized parameters, use the command line option [TT]-launch[tt].[PAR]",
2070 int pmeentries = 0; /* How many values for -npme do we actually test for each tpr file */
2071 real maxPMEfraction = 0.50;
2072 real minPMEfraction = 0.25;
2073 int maxPMEnodes, minPMEnodes;
2074 float guessPMEratio; /* guessed PME:PP ratio based on the tpr file */
2075 float guessPMEnodes;
2076 int npme_fixed = -2; /* If >= -1, use only this number
2077 * of PME-only nodes */
2079 real rmin = 0.0, rmax = 0.0; /* min and max value for rcoulomb if scaling is requested */
2080 real rcoulomb = -1.0; /* Coulomb radius as set in .tpr file */
2081 gmx_bool bScaleRvdw = TRUE;
2082 gmx_int64_t bench_nsteps = BENCHSTEPS;
2083 gmx_int64_t new_sim_nsteps = -1; /* -1 indicates: not set by the user */
2084 gmx_int64_t cpt_steps = 0; /* Step counter in .cpt input file */
2085 int presteps = 100; /* Do a full cycle reset after presteps steps */
2086 gmx_bool bOverwrite = FALSE, bKeepTPR;
2087 gmx_bool bLaunch = FALSE;
2088 char *ExtraArgs = NULL;
2089 char **tpr_names = NULL;
2090 const char *simulation_tpr = NULL;
2091 int best_npme, best_tpr;
2092 int sim_part = 1; /* For benchmarks with checkpoint files */
2095 /* Default program names if nothing else is found */
2096 char *cmd_mpirun = NULL, *cmd_mdrun = NULL;
2097 char *cmd_args_bench, *cmd_args_launch;
2098 char *cmd_np = NULL;
2100 t_perf **perfdata = NULL;
2106 /* Print out how long the tuning took */
2109 static t_filenm fnm[] = {
2111 { efOUT, "-p", "perf", ffWRITE },
2112 { efLOG, "-err", "bencherr", ffWRITE },
2113 { efTPX, "-so", "tuned", ffWRITE },
2115 { efTPX, NULL, NULL, ffREAD },
2116 { efTRN, "-o", NULL, ffWRITE },
2117 { efCOMPRESSED, "-x", NULL, ffOPTWR },
2118 { efCPT, "-cpi", NULL, ffOPTRD },
2119 { efCPT, "-cpo", NULL, ffOPTWR },
2120 { efSTO, "-c", "confout", ffWRITE },
2121 { efEDR, "-e", "ener", ffWRITE },
2122 { efLOG, "-g", "md", ffWRITE },
2123 { efXVG, "-dhdl", "dhdl", ffOPTWR },
2124 { efXVG, "-field", "field", ffOPTWR },
2125 { efXVG, "-table", "table", ffOPTRD },
2126 { efXVG, "-tabletf", "tabletf", ffOPTRD },
2127 { efXVG, "-tablep", "tablep", ffOPTRD },
2128 { efXVG, "-tableb", "table", ffOPTRD },
2129 { efTRX, "-rerun", "rerun", ffOPTRD },
2130 { efXVG, "-tpi", "tpi", ffOPTWR },
2131 { efXVG, "-tpid", "tpidist", ffOPTWR },
2132 { efEDI, "-ei", "sam", ffOPTRD },
2133 { efXVG, "-eo", "edsam", ffOPTWR },
2134 { efXVG, "-devout", "deviatie", ffOPTWR },
2135 { efXVG, "-runav", "runaver", ffOPTWR },
2136 { efXVG, "-px", "pullx", ffOPTWR },
2137 { efXVG, "-pf", "pullf", ffOPTWR },
2138 { efXVG, "-ro", "rotation", ffOPTWR },
2139 { efLOG, "-ra", "rotangles", ffOPTWR },
2140 { efLOG, "-rs", "rotslabs", ffOPTWR },
2141 { efLOG, "-rt", "rottorque", ffOPTWR },
2142 { efMTX, "-mtx", "nm", ffOPTWR },
2143 { efNDX, "-dn", "dipole", ffOPTWR },
2144 { efXVG, "-swap", "swapions", ffOPTWR },
2145 /* Output files that are deleted after each benchmark run */
2146 { efTRN, "-bo", "bench", ffWRITE },
2147 { efXTC, "-bx", "bench", ffWRITE },
2148 { efCPT, "-bcpo", "bench", ffWRITE },
2149 { efSTO, "-bc", "bench", ffWRITE },
2150 { efEDR, "-be", "bench", ffWRITE },
2151 { efLOG, "-bg", "bench", ffWRITE },
2152 { efXVG, "-beo", "benchedo", ffOPTWR },
2153 { efXVG, "-bdhdl", "benchdhdl", ffOPTWR },
2154 { efXVG, "-bfield", "benchfld", ffOPTWR },
2155 { efXVG, "-btpi", "benchtpi", ffOPTWR },
2156 { efXVG, "-btpid", "benchtpid", ffOPTWR },
2157 { efXVG, "-bdevout", "benchdev", ffOPTWR },
2158 { efXVG, "-brunav", "benchrnav", ffOPTWR },
2159 { efXVG, "-bpx", "benchpx", ffOPTWR },
2160 { efXVG, "-bpf", "benchpf", ffOPTWR },
2161 { efXVG, "-bro", "benchrot", ffOPTWR },
2162 { efLOG, "-bra", "benchrota", ffOPTWR },
2163 { efLOG, "-brs", "benchrots", ffOPTWR },
2164 { efLOG, "-brt", "benchrott", ffOPTWR },
2165 { efMTX, "-bmtx", "benchn", ffOPTWR },
2166 { efNDX, "-bdn", "bench", ffOPTWR },
2167 { efXVG, "-bswap", "benchswp", ffOPTWR }
2170 gmx_bool bThreads = FALSE;
2174 const char *procstring[] =
2175 { NULL, "-np", "-n", "none", NULL };
2176 const char *npmevalues_opt[] =
2177 { NULL, "auto", "all", "subset", NULL };
2179 gmx_bool bAppendFiles = TRUE;
2180 gmx_bool bKeepAndNumCPT = FALSE;
2181 gmx_bool bResetCountersHalfWay = FALSE;
2182 gmx_bool bBenchmark = TRUE;
2184 output_env_t oenv = NULL;
2187 /***********************/
2188 /* g_tune_pme options: */
2189 /***********************/
2190 { "-np", FALSE, etINT, {&nnodes},
2191 "Number of nodes to run the tests on (must be > 2 for separate PME nodes)" },
2192 { "-npstring", FALSE, etENUM, {procstring},
2193 "Specify the number of processors to [TT]$MPIRUN[tt] using this string"},
2194 { "-ntmpi", FALSE, etINT, {&nthreads},
2195 "Number of MPI-threads to run the tests on (turns MPI & mpirun off)"},
2196 { "-r", FALSE, etINT, {&repeats},
2197 "Repeat each test this often" },
2198 { "-max", FALSE, etREAL, {&maxPMEfraction},
2199 "Max fraction of PME nodes to test with" },
2200 { "-min", FALSE, etREAL, {&minPMEfraction},
2201 "Min fraction of PME nodes to test with" },
2202 { "-npme", FALSE, etENUM, {npmevalues_opt},
2203 "Within -min and -max, benchmark all possible values for [TT]-npme[tt], or just a reasonable subset. "
2204 "Auto neglects -min and -max and chooses reasonable values around a guess for npme derived from the .tpr"},
2205 { "-fix", FALSE, etINT, {&npme_fixed},
2206 "If >= -1, do not vary the number of PME-only nodes, instead use this fixed value and only vary rcoulomb and the PME grid spacing."},
2207 { "-rmax", FALSE, etREAL, {&rmax},
2208 "If >0, maximal rcoulomb for -ntpr>1 (rcoulomb upscaling results in fourier grid downscaling)" },
2209 { "-rmin", FALSE, etREAL, {&rmin},
2210 "If >0, minimal rcoulomb for -ntpr>1" },
2211 { "-scalevdw", FALSE, etBOOL, {&bScaleRvdw},
2212 "Scale rvdw along with rcoulomb"},
2213 { "-ntpr", FALSE, etINT, {&ntprs},
2214 "Number of [TT].tpr[tt] files to benchmark. Create this many files with different rcoulomb scaling factors depending on -rmin and -rmax. "
2215 "If < 1, automatically choose the number of [TT].tpr[tt] files to test" },
2216 { "-steps", FALSE, etINT64, {&bench_nsteps},
2217 "Take timings for this many steps in the benchmark runs" },
2218 { "-resetstep", FALSE, etINT, {&presteps},
2219 "Let dlb equilibrate this many steps before timings are taken (reset cycle counters after this many steps)" },
2220 { "-simsteps", FALSE, etINT64, {&new_sim_nsteps},
2221 "If non-negative, perform this many steps in the real run (overwrites nsteps from [TT].tpr[tt], add [TT].cpt[tt] steps)" },
2222 { "-launch", FALSE, etBOOL, {&bLaunch},
2223 "Launch the real simulation after optimization" },
2224 { "-bench", FALSE, etBOOL, {&bBenchmark},
2225 "Run the benchmarks or just create the input [TT].tpr[tt] files?" },
2226 /******************/
2227 /* mdrun options: */
2228 /******************/
2229 /* We let g_tune_pme parse and understand these options, because we need to
2230 * prevent that they appear on the mdrun command line for the benchmarks */
2231 { "-append", FALSE, etBOOL, {&bAppendFiles},
2232 "Append to previous output files when continuing from checkpoint instead of adding the simulation part number to all file names (for launch only)" },
2233 { "-cpnum", FALSE, etBOOL, {&bKeepAndNumCPT},
2234 "Keep and number checkpoint files (launch only)" },
2235 { "-resethway", FALSE, etBOOL, {&bResetCountersHalfWay},
2236 "HIDDENReset the cycle counters after half the number of steps or halfway [TT]-maxh[tt] (launch only)" }
2239 #define NFILE asize(fnm)
2241 seconds = gmx_gettime();
2243 if (!parse_common_args(&argc, argv, PCA_NOEXIT_ON_ARGS,
2244 NFILE, fnm, asize(pa), pa, asize(desc), desc,
2250 /* Store the remaining unparsed command line entries in a string which
2251 * is then attached to the mdrun command line */
2253 ExtraArgs[0] = '\0';
2254 for (i = 1; i < argc; i++) /* argc will now be 1 if everything was understood */
2256 add_to_string(&ExtraArgs, argv[i]);
2257 add_to_string(&ExtraArgs, " ");
2260 if (opt2parg_bSet("-ntmpi", asize(pa), pa))
2263 if (opt2parg_bSet("-npstring", asize(pa), pa))
2265 fprintf(stderr, "WARNING: -npstring has no effect when using threads.\n");
2270 gmx_fatal(FARGS, "Can't run multi-threaded MPI simulation yet!");
2272 /* and now we just set this; a bit of an ugly hack*/
2275 /* Check for PME:PP ratio and whether tpr triggers additional output files */
2276 guessPMEratio = inspect_tpr(NFILE, fnm, &rcoulomb);
2278 /* Automatically set -beo options if -eo is set etc. */
2279 couple_files_options(NFILE, fnm);
2281 /* Construct the command line arguments for benchmark runs
2282 * as well as for the simulation run */
2285 sprintf(bbuf, " -ntmpi %d ", nthreads);
2289 /* This string will be used for MPI runs and will appear after the
2290 * mpirun command. */
2291 if (strcmp(procstring[0], "none") != 0)
2293 sprintf(bbuf, " %s %d ", procstring[0], nnodes);
2303 create_command_line_snippets(bAppendFiles, bKeepAndNumCPT, bResetCountersHalfWay, presteps,
2304 NFILE, fnm, &cmd_args_bench, &cmd_args_launch, ExtraArgs);
2306 /* Read in checkpoint file if requested */
2308 if (opt2bSet("-cpi", NFILE, fnm))
2311 cr->duty = DUTY_PP; /* makes the following routine happy */
2312 read_checkpoint_simulation_part(opt2fn("-cpi", NFILE, fnm),
2313 &sim_part, &cpt_steps, cr,
2314 FALSE, NFILE, fnm, NULL, NULL);
2317 /* sim_part will now be 1 if no checkpoint file was found */
2320 gmx_fatal(FARGS, "Checkpoint file %s not found!", opt2fn("-cpi", NFILE, fnm));
2324 /* Open performance output file and write header info */
2325 fp = gmx_ffopen(opt2fn("-p", NFILE, fnm), "w");
2327 /* Make a quick consistency check of command line parameters */
2328 check_input(nnodes, repeats, &ntprs, &rmin, rcoulomb, &rmax,
2329 maxPMEfraction, minPMEfraction, npme_fixed,
2330 bench_nsteps, fnm, NFILE, sim_part, presteps,
2333 /* Determine the maximum and minimum number of PME nodes to test,
2334 * the actual list of settings is build in do_the_tests(). */
2335 if ((nnodes > 2) && (npme_fixed < -1))
2337 if (0 == strcmp(npmevalues_opt[0], "auto"))
2339 /* Determine the npme range automatically based on the PME:PP load guess */
2340 if (guessPMEratio > 1.0)
2342 /* More PME than PP work, probably we do not need separate PME nodes at all! */
2343 maxPMEnodes = nnodes/2;
2344 minPMEnodes = nnodes/2;
2348 /* PME : PP load is in the range 0..1, let's test around the guess */
2349 guessPMEnodes = nnodes/(1.0 + 1.0/guessPMEratio);
2350 minPMEnodes = floor(0.7*guessPMEnodes);
2351 maxPMEnodes = ceil(1.6*guessPMEnodes);
2352 maxPMEnodes = min(maxPMEnodes, nnodes/2);
2357 /* Determine the npme range based on user input */
2358 maxPMEnodes = floor(maxPMEfraction*nnodes);
2359 minPMEnodes = max(floor(minPMEfraction*nnodes), 0);
2360 fprintf(stdout, "Will try runs with %d ", minPMEnodes);
2361 if (maxPMEnodes != minPMEnodes)
2363 fprintf(stdout, "- %d ", maxPMEnodes);
2365 fprintf(stdout, "PME-only nodes.\n Note that the automatic number of PME-only nodes and no separate PME nodes are always tested.\n");
2374 /* Get the commands we need to set up the runs from environment variables */
2375 get_program_paths(bThreads, &cmd_mpirun, &cmd_mdrun);
2376 if (bBenchmark && repeats > 0)
2378 check_mdrun_works(bThreads, cmd_mpirun, cmd_np, cmd_mdrun);
2381 /* Print some header info to file */
2383 fprintf(fp, "\n P E R F O R M A N C E R E S U L T S\n");
2385 fprintf(fp, "%s for Gromacs %s\n", ShortProgram(), GromacsVersion());
2388 fprintf(fp, "Number of nodes : %d\n", nnodes);
2389 fprintf(fp, "The mpirun command is : %s\n", cmd_mpirun);
2390 if (strcmp(procstring[0], "none") != 0)
2392 fprintf(fp, "Passing # of nodes via : %s\n", procstring[0]);
2396 fprintf(fp, "Not setting number of nodes in system call\n");
2401 fprintf(fp, "Number of threads : %d\n", nnodes);
2404 fprintf(fp, "The mdrun command is : %s\n", cmd_mdrun);
2405 fprintf(fp, "mdrun args benchmarks : %s\n", cmd_args_bench);
2406 fprintf(fp, "Benchmark steps : ");
2407 fprintf(fp, "%"GMX_PRId64, bench_nsteps);
2409 fprintf(fp, "dlb equilibration steps : %d\n", presteps);
2412 fprintf(fp, "Checkpoint time step : ");
2413 fprintf(fp, "%"GMX_PRId64, cpt_steps);
2416 fprintf(fp, "mdrun args at launchtime: %s\n", cmd_args_launch);
2418 if (new_sim_nsteps >= 0)
2421 fprintf(stderr, "Note: Simulation input file %s will have ", opt2fn("-so", NFILE, fnm));
2422 fprintf(stderr, "%"GMX_PRId64, new_sim_nsteps+cpt_steps);
2423 fprintf(stderr, " steps.\n");
2424 fprintf(fp, "Simulation steps : ");
2425 fprintf(fp, "%"GMX_PRId64, new_sim_nsteps);
2430 fprintf(fp, "Repeats for each test : %d\n", repeats);
2433 if (npme_fixed >= -1)
2435 fprintf(fp, "Fixing -npme at : %d\n", npme_fixed);
2438 fprintf(fp, "Input file : %s\n", opt2fn("-s", NFILE, fnm));
2439 fprintf(fp, " PME/PP load estimate : %g\n", guessPMEratio);
2441 /* Allocate memory for the inputinfo struct: */
2443 info->nr_inputfiles = ntprs;
2444 for (i = 0; i < ntprs; i++)
2446 snew(info->rcoulomb, ntprs);
2447 snew(info->rvdw, ntprs);
2448 snew(info->rlist, ntprs);
2449 snew(info->rlistlong, ntprs);
2450 snew(info->nkx, ntprs);
2451 snew(info->nky, ntprs);
2452 snew(info->nkz, ntprs);
2453 snew(info->fsx, ntprs);
2454 snew(info->fsy, ntprs);
2455 snew(info->fsz, ntprs);
2457 /* Make alternative tpr files to test: */
2458 snew(tpr_names, ntprs);
2459 for (i = 0; i < ntprs; i++)
2461 snew(tpr_names[i], STRLEN);
2464 /* It can be that ntprs is reduced by make_benchmark_tprs if not enough
2465 * different grids could be found. */
2466 make_benchmark_tprs(opt2fn("-s", NFILE, fnm), tpr_names, bench_nsteps+presteps,
2467 cpt_steps, rmin, rmax, bScaleRvdw, &ntprs, info, fp);
2469 /********************************************************************************/
2470 /* Main loop over all scenarios we need to test: tpr files, PME nodes, repeats */
2471 /********************************************************************************/
2472 snew(perfdata, ntprs);
2475 do_the_tests(fp, tpr_names, maxPMEnodes, minPMEnodes, npme_fixed, npmevalues_opt[0], perfdata, &pmeentries,
2476 repeats, nnodes, ntprs, bThreads, cmd_mpirun, cmd_np, cmd_mdrun,
2477 cmd_args_bench, fnm, NFILE, presteps, cpt_steps);
2479 fprintf(fp, "\nTuning took%8.1f minutes.\n", (gmx_gettime()-seconds)/60.0);
2481 /* Analyse the results and give a suggestion for optimal settings: */
2482 bKeepTPR = analyze_data(fp, opt2fn("-p", NFILE, fnm), perfdata, nnodes, ntprs, pmeentries,
2483 repeats, info, &best_tpr, &best_npme);
2485 /* Take the best-performing tpr file and enlarge nsteps to original value */
2486 if (bKeepTPR && !bOverwrite)
2488 simulation_tpr = opt2fn("-s", NFILE, fnm);
2492 simulation_tpr = opt2fn("-so", NFILE, fnm);
2493 modify_PMEsettings(bOverwrite ? (new_sim_nsteps+cpt_steps) : info->orig_sim_steps,
2494 info->orig_init_step, tpr_names[best_tpr], simulation_tpr);
2497 /* Let's get rid of the temporary benchmark input files */
2498 for (i = 0; i < ntprs; i++)
2500 fprintf(stdout, "Deleting temporary benchmark input file %s\n", tpr_names[i]);
2501 remove(tpr_names[i]);
2504 /* Now start the real simulation if the user requested it ... */
2505 launch_simulation(bLaunch, fp, bThreads, cmd_mpirun, cmd_np, cmd_mdrun,
2506 cmd_args_launch, simulation_tpr, best_npme);
2510 /* ... or simply print the performance results to screen: */
2513 finalize(opt2fn("-p", NFILE, fnm));