1 /* -*- mode: c; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4; c-file-style: "stroustrup"; -*-
4 * This source code is part of
8 * GROningen MAchine for Chemical Simulations
11 * Written by David van der Spoel, Erik Lindahl, Berk Hess, and others.
12 * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
13 * Copyright (c) 2001-2011, The GROMACS development team,
14 * check out http://www.gromacs.org for more information.
16 * This program is free software; you can redistribute it and/or
17 * modify it under the terms of the GNU General Public License
18 * as published by the Free Software Foundation; either version 2
19 * of the License, or (at your option) any later version.
21 * If you want to redistribute modifications, please consider that
22 * scientific software is very special. Version control is crucial -
23 * bugs must be traceable. We will be happy to consider code for
24 * inclusion in the official distribution, but derived work must not
25 * be called official GROMACS. Details are found in the README & COPYING
26 * files - if they are missing, get the official version at www.gromacs.org.
28 * To help us fund GROMACS development, we humbly ask that you cite
29 * the papers on the package - you can find them in the top README file.
31 * For more info, check our website at http://www.gromacs.org
34 * Gallium Rubidium Oxygen Manganese Argon Carbon Silicon
46 #include "nbnxn_cuda_data_mgmt.h"
49 #include "md_logging.h"
50 #include "pme_loadbal.h"
52 /* Parameters and setting for one PP-PME setup */
54 real rcut_coulomb; /* Coulomb cut-off */
55 real rlist; /* pair-list cut-off */
56 real rlistlong; /* LR pair-list cut-off */
57 int nstcalclr; /* frequency of evaluating long-range forces for group scheme */
58 real spacing; /* (largest) PME grid spacing */
59 ivec grid; /* the PME grid dimensions */
60 real grid_efficiency; /* ineffiency factor for non-uniform grids <= 1 */
61 real ewaldcoeff; /* the Ewald coefficient */
62 gmx_pme_t pmedata; /* the data structure used in the PME code */
64 int count; /* number of times this setup has been timed */
65 double cycles; /* the fastest time for this setup in cycles */
68 /* In the initial scan, step by grids that are at least a factor 0.8 coarser */
69 #define PME_LB_GRID_SCALE_FAC 0.8
70 /* In the initial scan, try to skip grids with uneven x/y/z spacing,
71 * checking if the "efficiency" is more than 5% worse than the previous grid.
73 #define PME_LB_GRID_EFFICIENCY_REL_FAC 1.05
74 /* Rerun up till 12% slower setups than the fastest up till now */
75 #define PME_LB_SLOW_FAC 1.12
76 /* If setups get more than 2% faster, do another round to avoid
77 * choosing a slower setup due to acceleration or fluctuations.
79 #define PME_LB_ACCEL_TOL 1.02
82 epmelblimNO, epmelblimBOX, epmelblimDD, epmelblimNR
85 const char *pmelblim_str[epmelblimNR] =
86 { "no", "box size", "domain decompostion" };
88 struct pme_load_balancing {
89 int nstage; /* the current maximum number of stages */
91 real cut_spacing; /* the minimum cutoff / PME grid spacing ratio */
92 real rcut_vdw; /* Vdw cutoff (does not change) */
93 real rcut_coulomb_start; /* Initial electrostatics cutoff */
94 int nstcalclr_start; /* Initial electrostatics cutoff */
95 real rbuf_coulomb; /* the pairlist buffer size */
96 real rbuf_vdw; /* the pairlist buffer size */
97 matrix box_start; /* the initial simulation box */
98 int n; /* the count of setup as well as the allocation size */
99 pme_setup_t *setup; /* the PME+cutoff setups */
100 int cur; /* the current setup */
101 int fastest; /* fastest setup up till now */
102 int start; /* start of setup range to consider in stage>0 */
103 int end; /* end of setup range to consider in stage>0 */
104 int elimited; /* was the balancing limited, uses enum above */
105 int cutoff_scheme; /* Verlet or group cut-offs */
107 int stage; /* the current stage */
110 void pme_loadbal_init(pme_load_balancing_t *pme_lb_p,
111 const t_inputrec *ir, matrix box,
112 const interaction_const_t *ic,
115 pme_load_balancing_t pme_lb;
121 /* Any number of stages >= 2 is supported */
124 pme_lb->cutoff_scheme = ir->cutoff_scheme;
126 if (pme_lb->cutoff_scheme == ecutsVERLET)
128 pme_lb->rbuf_coulomb = ic->rlist - ic->rcoulomb;
129 pme_lb->rbuf_vdw = pme_lb->rbuf_coulomb;
133 if (ic->rcoulomb > ic->rlist)
135 pme_lb->rbuf_coulomb = ic->rlistlong - ic->rcoulomb;
139 pme_lb->rbuf_coulomb = ic->rlist - ic->rcoulomb;
141 if (ic->rvdw > ic->rlist)
143 pme_lb->rbuf_vdw = ic->rlistlong - ic->rvdw;
147 pme_lb->rbuf_vdw = ic->rlist - ic->rvdw;
151 copy_mat(box, pme_lb->box_start);
152 if (ir->ePBC == epbcXY && ir->nwall == 2)
154 svmul(ir->wall_ewald_zfac, pme_lb->box_start[ZZ], pme_lb->box_start[ZZ]);
158 snew(pme_lb->setup, pme_lb->n);
160 pme_lb->rcut_vdw = ic->rvdw;
161 pme_lb->rcut_coulomb_start = ir->rcoulomb;
162 pme_lb->nstcalclr_start = ir->nstcalclr;
165 pme_lb->setup[0].rcut_coulomb = ic->rcoulomb;
166 pme_lb->setup[0].rlist = ic->rlist;
167 pme_lb->setup[0].rlistlong = ic->rlistlong;
168 pme_lb->setup[0].nstcalclr = ir->nstcalclr;
169 pme_lb->setup[0].grid[XX] = ir->nkx;
170 pme_lb->setup[0].grid[YY] = ir->nky;
171 pme_lb->setup[0].grid[ZZ] = ir->nkz;
172 pme_lb->setup[0].ewaldcoeff = ic->ewaldcoeff;
174 pme_lb->setup[0].pmedata = pmedata;
177 for (d = 0; d < DIM; d++)
179 sp = norm(pme_lb->box_start[d])/pme_lb->setup[0].grid[d];
185 pme_lb->setup[0].spacing = spm;
187 if (ir->fourier_spacing > 0)
189 pme_lb->cut_spacing = ir->rcoulomb/ir->fourier_spacing;
193 pme_lb->cut_spacing = ir->rcoulomb/pme_lb->setup[0].spacing;
201 pme_lb->elimited = epmelblimNO;
206 static gmx_bool pme_loadbal_increase_cutoff(pme_load_balancing_t pme_lb,
211 real tmpr_coulomb, tmpr_vdw;
214 /* Try to add a new setup with next larger cut-off to the list */
216 srenew(pme_lb->setup, pme_lb->n);
217 set = &pme_lb->setup[pme_lb->n-1];
224 clear_ivec(set->grid);
225 sp = calc_grid(NULL, pme_lb->box_start,
226 fac*pme_lb->setup[pme_lb->cur].spacing,
231 /* In parallel we can't have grids smaller than 2*pme_order,
232 * and we would anyhow not gain much speed at these grid sizes.
234 for (d = 0; d < DIM; d++)
236 if (set->grid[d] <= 2*pme_order)
244 while (sp <= 1.001*pme_lb->setup[pme_lb->cur].spacing);
246 set->rcut_coulomb = pme_lb->cut_spacing*sp;
248 if (pme_lb->cutoff_scheme == ecutsVERLET)
250 set->rlist = set->rcut_coulomb + pme_lb->rbuf_coulomb;
251 /* We dont use LR lists with Verlet, but this avoids if-statements in further checks */
252 set->rlistlong = set->rlist;
256 tmpr_coulomb = set->rcut_coulomb + pme_lb->rbuf_coulomb;
257 tmpr_vdw = pme_lb->rcut_vdw + pme_lb->rbuf_vdw;
258 set->rlist = min(tmpr_coulomb, tmpr_vdw);
259 set->rlistlong = max(tmpr_coulomb, tmpr_vdw);
261 /* Set the long-range update frequency */
262 if (set->rlist == set->rlistlong)
264 /* No long-range interactions if the short-/long-range cutoffs are identical */
267 else if (pme_lb->nstcalclr_start == 0 || pme_lb->nstcalclr_start == 1)
269 /* We were not doing long-range before, but now we are since rlist!=rlistlong */
274 /* We were already doing long-range interactions from the start */
275 if (pme_lb->rcut_vdw > pme_lb->rcut_coulomb_start)
277 /* We were originally doing long-range VdW-only interactions.
278 * If rvdw is still longer than rcoulomb we keep the original nstcalclr,
279 * but if the coulomb cutoff has become longer we should update the long-range
282 set->nstcalclr = (tmpr_vdw > tmpr_coulomb) ? pme_lb->nstcalclr_start : 1;
286 /* We were not doing any long-range interaction from the start,
287 * since it is not possible to do twin-range coulomb for the PME interaction.
295 /* The grid efficiency is the size wrt a grid with uniform x/y/z spacing */
296 set->grid_efficiency = 1;
297 for (d = 0; d < DIM; d++)
299 set->grid_efficiency *= (set->grid[d]*sp)/norm(pme_lb->box_start[d]);
301 /* The Ewald coefficient is inversly proportional to the cut-off */
303 pme_lb->setup[0].ewaldcoeff*pme_lb->setup[0].rcut_coulomb/set->rcut_coulomb;
310 fprintf(debug, "PME loadbal: grid %d %d %d, coulomb cutoff %f\n",
311 set->grid[XX], set->grid[YY], set->grid[ZZ], set->rcut_coulomb);
316 static void print_grid(FILE *fp_err, FILE *fp_log,
319 const pme_setup_t *set,
322 char buf[STRLEN], buft[STRLEN];
326 sprintf(buft, ": %.1f M-cycles", cycles*1e-6);
332 sprintf(buf, "%-11s%10s pme grid %d %d %d, coulomb cutoff %.3f%s",
334 desc, set->grid[XX], set->grid[YY], set->grid[ZZ], set->rcut_coulomb,
338 fprintf(fp_err, "\r%s\n", buf);
342 fprintf(fp_log, "%s\n", buf);
346 static int pme_loadbal_end(pme_load_balancing_t pme_lb)
348 /* In the initial stage only n is set; end is not set yet */
359 static void print_loadbal_limited(FILE *fp_err, FILE *fp_log,
360 gmx_large_int_t step,
361 pme_load_balancing_t pme_lb)
363 char buf[STRLEN], sbuf[22];
365 sprintf(buf, "step %4s: the %s limited the PME load balancing to a coulomb cut-off of %.3f",
366 gmx_step_str(step, sbuf),
367 pmelblim_str[pme_lb->elimited],
368 pme_lb->setup[pme_loadbal_end(pme_lb)-1].rcut_coulomb);
371 fprintf(fp_err, "\r%s\n", buf);
375 fprintf(fp_log, "%s\n", buf);
379 static void switch_to_stage1(pme_load_balancing_t pme_lb)
382 while (pme_lb->start+1 < pme_lb->n &&
383 (pme_lb->setup[pme_lb->start].count == 0 ||
384 pme_lb->setup[pme_lb->start].cycles >
385 pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC))
389 while (pme_lb->start > 0 && pme_lb->setup[pme_lb->start-1].cycles == 0)
394 pme_lb->end = pme_lb->n;
395 if (pme_lb->setup[pme_lb->end-1].count > 0 &&
396 pme_lb->setup[pme_lb->end-1].cycles >
397 pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC)
404 /* Next we want to choose setup pme_lb->start, but as we will increase
405 * pme_ln->cur by one right after returning, we subtract 1 here.
407 pme_lb->cur = pme_lb->start - 1;
410 gmx_bool pme_load_balance(pme_load_balancing_t pme_lb,
417 interaction_const_t *ic,
418 nonbonded_verlet_t *nbv,
420 gmx_large_int_t step)
425 char buf[STRLEN], sbuf[22];
427 gmx_bool bUsesSimpleTables = TRUE;
429 if (pme_lb->stage == pme_lb->nstage)
436 gmx_sumd(1, &cycles, cr);
437 cycles /= cr->nnodes;
440 set = &pme_lb->setup[pme_lb->cur];
443 rtab = ir->rlistlong + ir->tabext;
445 if (set->count % 2 == 1)
447 /* Skip the first cycle, because the first step after a switch
448 * is much slower due to allocation and/or caching effects.
453 sprintf(buf, "step %4s: ", gmx_step_str(step, sbuf));
454 print_grid(fp_err, fp_log, buf, "timed with", set, cycles);
458 set->cycles = cycles;
462 if (cycles*PME_LB_ACCEL_TOL < set->cycles &&
463 pme_lb->stage == pme_lb->nstage - 1)
465 /* The performance went up a lot (due to e.g. DD load balancing).
466 * Add a stage, keep the minima, but rescan all setups.
472 fprintf(debug, "The performance for grid %d %d %d went from %.3f to %.1f M-cycles, this is more than %f\n"
473 "Increased the number stages to %d"
474 " and ignoring the previous performance\n",
475 set->grid[XX], set->grid[YY], set->grid[ZZ],
476 cycles*1e-6, set->cycles*1e-6, PME_LB_ACCEL_TOL,
480 set->cycles = min(set->cycles, cycles);
483 if (set->cycles < pme_lb->setup[pme_lb->fastest].cycles)
485 pme_lb->fastest = pme_lb->cur;
487 if (DOMAINDECOMP(cr))
489 /* We found a new fastest setting, ensure that with subsequent
490 * shorter cut-off's the dynamic load balancing does not make
491 * the use of the current cut-off impossible. This solution is
492 * a trade-off, as the PME load balancing and DD domain size
493 * load balancing can interact in complex ways.
494 * With the Verlet kernels, DD load imbalance will usually be
495 * mainly due to bonded interaction imbalance, which will often
496 * quickly push the domain boundaries beyond the limit for the
497 * optimal, PME load balanced, cut-off. But it could be that
498 * better overal performance can be obtained with a slightly
499 * shorter cut-off and better DD load balancing.
501 change_dd_dlb_cutoff_limit(cr);
504 cycles_fast = pme_lb->setup[pme_lb->fastest].cycles;
506 /* Check in stage 0 if we should stop scanning grids.
507 * Stop when the time is more than SLOW_FAC longer than the fastest.
509 if (pme_lb->stage == 0 && pme_lb->cur > 0 &&
510 cycles > pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC)
512 pme_lb->n = pme_lb->cur + 1;
513 /* Done with scanning, go to stage 1 */
514 switch_to_stage1(pme_lb);
517 if (pme_lb->stage == 0)
521 gridsize_start = set->grid[XX]*set->grid[YY]*set->grid[ZZ];
525 if (pme_lb->cur+1 < pme_lb->n)
527 /* We had already generated the next setup */
532 /* Find the next setup */
533 OK = pme_loadbal_increase_cutoff(pme_lb, ir->pme_order);
536 if (OK && ir->ePBC != epbcNONE)
538 OK = (sqr(pme_lb->setup[pme_lb->cur+1].rlistlong)
539 <= max_cutoff2(ir->ePBC, state->box));
542 pme_lb->elimited = epmelblimBOX;
550 if (DOMAINDECOMP(cr))
552 OK = change_dd_cutoff(cr, state, ir,
553 pme_lb->setup[pme_lb->cur].rlistlong);
556 /* Failed: do not use this setup */
558 pme_lb->elimited = epmelblimDD;
564 /* We hit the upper limit for the cut-off,
565 * the setup should not go further than cur.
567 pme_lb->n = pme_lb->cur + 1;
568 print_loadbal_limited(fp_err, fp_log, step, pme_lb);
569 /* Switch to the next stage */
570 switch_to_stage1(pme_lb);
574 !(pme_lb->setup[pme_lb->cur].grid[XX]*
575 pme_lb->setup[pme_lb->cur].grid[YY]*
576 pme_lb->setup[pme_lb->cur].grid[ZZ] <
577 gridsize_start*PME_LB_GRID_SCALE_FAC
579 pme_lb->setup[pme_lb->cur].grid_efficiency <
580 pme_lb->setup[pme_lb->cur-1].grid_efficiency*PME_LB_GRID_EFFICIENCY_REL_FAC));
583 if (pme_lb->stage > 0 && pme_lb->end == 1)
586 pme_lb->stage = pme_lb->nstage;
588 else if (pme_lb->stage > 0 && pme_lb->end > 1)
590 /* If stage = nstage-1:
591 * scan over all setups, rerunning only those setups
592 * which are not much slower than the fastest
599 if (pme_lb->cur == pme_lb->end)
602 pme_lb->cur = pme_lb->start;
605 while (pme_lb->stage == pme_lb->nstage - 1 &&
606 pme_lb->setup[pme_lb->cur].count > 0 &&
607 pme_lb->setup[pme_lb->cur].cycles > cycles_fast*PME_LB_SLOW_FAC);
609 if (pme_lb->stage == pme_lb->nstage)
611 /* We are done optimizing, use the fastest setup we found */
612 pme_lb->cur = pme_lb->fastest;
616 if (DOMAINDECOMP(cr) && pme_lb->stage > 0)
618 OK = change_dd_cutoff(cr, state, ir, pme_lb->setup[pme_lb->cur].rlistlong);
621 /* Failsafe solution */
622 if (pme_lb->cur > 1 && pme_lb->stage == pme_lb->nstage)
628 pme_lb->end = pme_lb->cur;
629 pme_lb->cur = pme_lb->start;
630 pme_lb->elimited = epmelblimDD;
631 print_loadbal_limited(fp_err, fp_log, step, pme_lb);
635 /* Change the Coulomb cut-off and the PME grid */
637 set = &pme_lb->setup[pme_lb->cur];
639 ic->rcoulomb = set->rcut_coulomb;
640 ic->rlist = set->rlist;
641 ic->rlistlong = set->rlistlong;
642 ir->nstcalclr = set->nstcalclr;
643 ic->ewaldcoeff = set->ewaldcoeff;
645 bUsesSimpleTables = uses_simple_tables(ir->cutoff_scheme, nbv, 0);
646 if (pme_lb->cutoff_scheme == ecutsVERLET &&
647 nbv->grp[0].kernel_type == nbnxnk8x8x8_CUDA)
649 nbnxn_cuda_pme_loadbal_update_param(nbv->cu_nbv, ic);
653 init_interaction_const_tables(NULL, ic, bUsesSimpleTables,
657 if (pme_lb->cutoff_scheme == ecutsVERLET && nbv->ngrp > 1)
659 init_interaction_const_tables(NULL, ic, bUsesSimpleTables,
663 if (cr->duty & DUTY_PME)
665 if (pme_lb->setup[pme_lb->cur].pmedata == NULL)
667 /* Generate a new PME data structure,
668 * copying part of the old pointers.
670 gmx_pme_reinit(&set->pmedata,
671 cr, pme_lb->setup[0].pmedata, ir,
674 *pmedata = set->pmedata;
678 /* Tell our PME-only node to switch grid */
679 gmx_pme_send_switchgrid(cr, set->grid, set->ewaldcoeff);
684 print_grid(NULL, debug, "", "switched to", set, -1);
687 if (pme_lb->stage == pme_lb->nstage)
689 print_grid(fp_err, fp_log, "", "optimal", set, -1);
695 void restart_pme_loadbal(pme_load_balancing_t pme_lb, int n)
700 static int pme_grid_points(const pme_setup_t *setup)
702 return setup->grid[XX]*setup->grid[YY]*setup->grid[ZZ];
705 static real pme_loadbal_rlist(const pme_setup_t *setup)
707 /* With the group cut-off scheme we can have twin-range either
708 * for Coulomb or for VdW, so we need a check here.
709 * With the Verlet cut-off scheme rlist=rlistlong.
711 if (setup->rcut_coulomb > setup->rlist)
713 return setup->rlistlong;
721 static void print_pme_loadbal_setting(FILE *fplog,
723 const pme_setup_t *setup)
726 " %-7s %6.3f nm %6.3f nm %3d %3d %3d %5.3f nm %5.3f nm\n",
728 setup->rcut_coulomb, pme_loadbal_rlist(setup),
729 setup->grid[XX], setup->grid[YY], setup->grid[ZZ],
730 setup->spacing, 1/setup->ewaldcoeff);
733 static void print_pme_loadbal_settings(pme_load_balancing_t pme_lb,
736 gmx_bool bNonBondedOnGPU)
738 double pp_ratio, grid_ratio;
740 pp_ratio = pow(pme_loadbal_rlist(&pme_lb->setup[pme_lb->cur])/pme_loadbal_rlist(&pme_lb->setup[0]), 3.0);
741 grid_ratio = pme_grid_points(&pme_lb->setup[pme_lb->cur])/
742 (double)pme_grid_points(&pme_lb->setup[0]);
744 fprintf(fplog, "\n");
745 fprintf(fplog, " P P - P M E L O A D B A L A N C I N G\n");
746 fprintf(fplog, "\n");
747 /* Here we only warn when the optimal setting is the last one */
748 if (pme_lb->elimited != epmelblimNO &&
749 pme_lb->cur == pme_loadbal_end(pme_lb)-1)
751 fprintf(fplog, " NOTE: The PP/PME load balancing was limited by the %s,\n",
752 pmelblim_str[pme_lb->elimited]);
753 fprintf(fplog, " you might not have reached a good load balance.\n");
754 if (pme_lb->elimited == epmelblimDD)
756 fprintf(fplog, " Try different mdrun -dd settings or lower the -dds value.\n");
758 fprintf(fplog, "\n");
760 fprintf(fplog, " PP/PME load balancing changed the cut-off and PME settings:\n");
761 fprintf(fplog, " particle-particle PME\n");
762 fprintf(fplog, " rcoulomb rlist grid spacing 1/beta\n");
763 print_pme_loadbal_setting(fplog, "initial", &pme_lb->setup[0]);
764 print_pme_loadbal_setting(fplog, "final", &pme_lb->setup[pme_lb->cur]);
765 fprintf(fplog, " cost-ratio %4.2f %4.2f\n",
766 pp_ratio, grid_ratio);
767 fprintf(fplog, " (note that these numbers concern only part of the total PP and PME load)\n");
769 if (pp_ratio > 1.5 && !bNonBondedOnGPU)
771 md_print_warn(cr, fplog,
772 "NOTE: PME load balancing increased the non-bonded workload by more than 50%%.\n"
773 " For better performance use (more) PME nodes (mdrun -npme),\n"
774 " or in case you are beyond the scaling limit, use less nodes in total.\n");
778 fprintf(fplog, "\n");
782 void pme_loadbal_done(pme_load_balancing_t pme_lb,
783 t_commrec *cr, FILE *fplog,
784 gmx_bool bNonBondedOnGPU)
786 if (fplog != NULL && (pme_lb->cur > 0 || pme_lb->elimited != epmelblimNO))
788 print_pme_loadbal_settings(pme_lb, cr, fplog, bNonBondedOnGPU);
791 /* TODO: Here we should free all pointers in pme_lb,
792 * but as it contains pme data structures,
793 * we need to first make pme.c free all data.