2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2015,2016,2017, 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.
37 * \brief This file contains function definitions necessary for
38 * managing automatic load balance of PME calculations (Coulomb and
41 * \author Berk Hess <hess@kth.se>
42 * \ingroup module_ewald
46 #include "pme-load-balancing.h"
56 #include "gromacs/domdec/domdec.h"
57 #include "gromacs/domdec/domdec_network.h"
58 #include "gromacs/domdec/domdec_struct.h"
59 #include "gromacs/ewald/pme.h"
60 #include "gromacs/fft/calcgrid.h"
61 #include "gromacs/gmxlib/network.h"
62 #include "gromacs/math/functions.h"
63 #include "gromacs/math/vec.h"
64 #include "gromacs/mdlib/forcerec.h"
65 #include "gromacs/mdlib/nbnxn_gpu_data_mgmt.h"
66 #include "gromacs/mdlib/sim_util.h"
67 #include "gromacs/mdtypes/commrec.h"
68 #include "gromacs/mdtypes/inputrec.h"
69 #include "gromacs/mdtypes/md_enums.h"
70 #include "gromacs/mdtypes/state.h"
71 #include "gromacs/pbcutil/pbc.h"
72 #include "gromacs/timing/wallcycle.h"
73 #include "gromacs/utility/cstringutil.h"
74 #include "gromacs/utility/fatalerror.h"
75 #include "gromacs/utility/gmxassert.h"
76 #include "gromacs/utility/logger.h"
77 #include "gromacs/utility/smalloc.h"
79 #include "pme-internal.h"
81 /*! \brief Parameters and settings for one PP-PME setup */
83 real rcut_coulomb; /**< Coulomb cut-off */
84 real rlist; /**< pair-list cut-off */
85 real spacing; /**< (largest) PME grid spacing */
86 ivec grid; /**< the PME grid dimensions */
87 real grid_efficiency; /**< ineffiency factor for non-uniform grids <= 1 */
88 real ewaldcoeff_q; /**< Electrostatic Ewald coefficient */
89 real ewaldcoeff_lj; /**< LJ Ewald coefficient, only for the call to send_switchgrid */
90 struct gmx_pme_t *pmedata; /**< the data structure used in the PME code */
91 int count; /**< number of times this setup has been timed */
92 double cycles; /**< the fastest time for this setup in cycles */
95 /*! \brief After 50 nstlist periods of not observing imbalance: never tune PME */
96 const int PMETunePeriod = 50;
97 /*! \brief Trigger PME load balancing at more than 5% PME overload */
98 const real loadBalanceTriggerFactor = 1.05;
99 /*! \brief In the initial scan, step by grids that are at least a factor 0.8 coarser */
100 const real gridScaleFactor = 0.8;
101 /*! \brief In the initial scan, try to skip grids with uneven x/y/z spacing,
102 * checking if the "efficiency" is more than 5% worse than the previous grid.
104 const real relativeEfficiencyFactor = 1.05;
105 /*! \brief Rerun until a run is 12% slower setups than the fastest run so far */
106 const real maxRelativeSlowdownAccepted = 1.12;
107 /*! \brief If setups get more than 2% faster, do another round to avoid
108 * choosing a slower setup due to acceleration or fluctuations.
110 const real maxFluctuationAccepted = 1.02;
112 /*! \brief Enumeration whose values describe the effect limiting the load balancing */
114 epmelblimNO, epmelblimBOX, epmelblimDD, epmelblimPMEGRID, epmelblimNR
117 /*! \brief Descriptive strings matching ::epmelb */
118 const char *pmelblim_str[epmelblimNR] =
119 { "no", "box size", "domain decompostion", "PME grid restriction" };
121 struct pme_load_balancing_t {
122 gmx_bool bSepPMERanks; /**< do we have separate PME ranks? */
123 gmx_bool bActive; /**< is PME tuning active? */
124 gmx_int64_t step_rel_stop; /**< stop the tuning after this value of step_rel */
125 gmx_bool bTriggerOnDLB; /**< trigger balancing only on DD DLB */
126 gmx_bool bBalance; /**< are we in the balancing phase, i.e. trying different setups? */
127 int nstage; /**< the current maximum number of stages */
129 real cut_spacing; /**< the minimum cutoff / PME grid spacing ratio */
130 real rcut_vdw; /**< Vdw cutoff (does not change) */
131 real rcut_coulomb_start; /**< Initial electrostatics cutoff */
132 real rbuf_coulomb; /**< the pairlist buffer size */
133 real rbuf_vdw; /**< the pairlist buffer size */
134 matrix box_start; /**< the initial simulation box */
135 int n; /**< the count of setup as well as the allocation size */
136 pme_setup_t *setup; /**< the PME+cutoff setups */
137 int cur; /**< the inex (in setup) of the current setup */
138 int fastest; /**< index of the fastest setup up till now */
139 int lower_limit; /**< don't go below this setup index */
140 int start; /**< start of setup index range to consider in stage>0 */
141 int end; /**< end of setup index range to consider in stage>0 */
142 int elimited; /**< was the balancing limited, uses enum above */
143 int cutoff_scheme; /**< Verlet or group cut-offs */
145 int stage; /**< the current stage */
147 int cycles_n; /**< step cycle counter cummulative count */
148 double cycles_c; /**< step cycle counter cummulative cycles */
151 /* TODO The code in this file should call this getter, rather than
152 * read bActive anywhere */
153 bool pme_loadbal_is_active(const pme_load_balancing_t *pme_lb)
155 return pme_lb != nullptr && pme_lb->bActive;
158 void pme_loadbal_init(pme_load_balancing_t **pme_lb_p,
160 const gmx::MDLogger &mdlog,
161 const t_inputrec *ir,
163 const interaction_const_t *ic,
168 pme_load_balancing_t *pme_lb;
172 // Note that we don't (yet) support PME load balancing with LJ-PME only.
173 GMX_RELEASE_ASSERT(EEL_PME(ir->coulombtype), "pme_loadbal_init called without PME electrostatics");
174 // To avoid complexity, we require a single cut-off with PME for q+LJ.
175 // This is checked by grompp, but it doesn't hurt to check again.
176 GMX_RELEASE_ASSERT(!(EEL_PME(ir->coulombtype) && EVDW_PME(ir->vdwtype) && ir->rcoulomb != ir->rvdw), "With Coulomb and LJ PME, rcoulomb should be equal to rvdw");
180 pme_lb->bSepPMERanks = !(cr->duty & DUTY_PME);
182 /* Initially we turn on balancing directly on based on PP/PME imbalance */
183 pme_lb->bTriggerOnDLB = FALSE;
185 /* Any number of stages >= 2 is supported */
188 pme_lb->cutoff_scheme = ir->cutoff_scheme;
190 pme_lb->rbuf_coulomb = ic->rlist - ic->rcoulomb;
191 pme_lb->rbuf_vdw = ic->rlist - ic->rvdw;
193 copy_mat(box, pme_lb->box_start);
194 if (ir->ePBC == epbcXY && ir->nwall == 2)
196 svmul(ir->wall_ewald_zfac, pme_lb->box_start[ZZ], pme_lb->box_start[ZZ]);
200 snew(pme_lb->setup, pme_lb->n);
202 pme_lb->rcut_vdw = ic->rvdw;
203 pme_lb->rcut_coulomb_start = ir->rcoulomb;
206 pme_lb->setup[0].rcut_coulomb = ic->rcoulomb;
207 pme_lb->setup[0].rlist = ic->rlist;
208 pme_lb->setup[0].grid[XX] = ir->nkx;
209 pme_lb->setup[0].grid[YY] = ir->nky;
210 pme_lb->setup[0].grid[ZZ] = ir->nkz;
211 pme_lb->setup[0].ewaldcoeff_q = ic->ewaldcoeff_q;
212 pme_lb->setup[0].ewaldcoeff_lj = ic->ewaldcoeff_lj;
214 pme_lb->setup[0].pmedata = pmedata;
217 for (d = 0; d < DIM; d++)
219 sp = norm(pme_lb->box_start[d])/pme_lb->setup[0].grid[d];
225 pme_lb->setup[0].spacing = spm;
227 if (ir->fourier_spacing > 0)
229 pme_lb->cut_spacing = ir->rcoulomb/ir->fourier_spacing;
233 pme_lb->cut_spacing = ir->rcoulomb/pme_lb->setup[0].spacing;
239 pme_lb->lower_limit = 0;
242 pme_lb->elimited = epmelblimNO;
244 pme_lb->cycles_n = 0;
245 pme_lb->cycles_c = 0;
247 if (!wallcycle_have_counter())
249 GMX_LOG(mdlog.warning).asParagraph().appendText("NOTE: Cycle counters unsupported or not enabled in kernel. Cannot use PME-PP balancing.");
252 /* Tune with GPUs and/or separate PME ranks.
253 * When running only on a CPU without PME ranks, PME tuning will only help
254 * with small numbers of atoms in the cut-off sphere.
256 pme_lb->bActive = (wallcycle_have_counter() && (bUseGPU ||
257 pme_lb->bSepPMERanks));
259 /* With GPUs and no separate PME ranks we can't measure the PP/PME
260 * imbalance, so we start balancing right away.
261 * Otherwise we only start balancing after we observe imbalance.
263 pme_lb->bBalance = (pme_lb->bActive && (bUseGPU && !pme_lb->bSepPMERanks));
265 pme_lb->step_rel_stop = PMETunePeriod*ir->nstlist;
267 /* Delay DD load balancing when GPUs are used */
268 if (pme_lb->bActive && DOMAINDECOMP(cr) && cr->dd->nnodes > 1 && bUseGPU)
270 /* Lock DLB=auto to off (does nothing when DLB=yes/no.
271 * With GPUs and separate PME nodes, we want to first
272 * do PME tuning without DLB, since DLB might limit
273 * the cut-off, which never improves performance.
274 * We allow for DLB + PME tuning after a first round of tuning.
277 if (dd_dlb_is_locked(cr->dd))
279 GMX_LOG(mdlog.warning).asParagraph().appendText("NOTE: DLB will not turn on during the first phase of PME tuning");
285 *bPrinting = pme_lb->bBalance;
288 /*! \brief Try to increase the cutoff during load balancing */
289 static gmx_bool pme_loadbal_increase_cutoff(pme_load_balancing_t *pme_lb,
291 const gmx_domdec_t *dd)
294 int npmeranks_x, npmeranks_y;
296 real tmpr_coulomb, tmpr_vdw;
300 /* Try to add a new setup with next larger cut-off to the list */
302 srenew(pme_lb->setup, pme_lb->n);
303 set = &pme_lb->setup[pme_lb->n-1];
304 set->pmedata = nullptr;
306 get_pme_nnodes(dd, &npmeranks_x, &npmeranks_y);
311 /* Avoid infinite while loop, which can occur at the minimum grid size.
312 * Note that in practice load balancing will stop before this point.
313 * The factor 2.1 allows for the extreme case in which only grids
314 * of powers of 2 are allowed (the current code supports more grids).
324 clear_ivec(set->grid);
325 sp = calcFftGrid(nullptr, pme_lb->box_start,
326 fac*pme_lb->setup[pme_lb->cur].spacing,
327 minimalPmeGridSize(pme_order),
332 /* As here we can't easily check if one of the PME ranks
333 * uses threading, we do a conservative grid check.
334 * This means we can't use pme_order or less grid lines
335 * per PME rank along x, which is not a strong restriction.
337 grid_ok = gmx_pme_check_restrictions(pme_order,
338 set->grid[XX], set->grid[YY], set->grid[ZZ],
343 while (sp <= 1.001*pme_lb->setup[pme_lb->cur].spacing || !grid_ok);
345 set->rcut_coulomb = pme_lb->cut_spacing*sp;
346 if (set->rcut_coulomb < pme_lb->rcut_coulomb_start)
348 /* This is unlikely, but can happen when e.g. continuing from
349 * a checkpoint after equilibration where the box shrank a lot.
350 * We want to avoid rcoulomb getting smaller than rvdw
351 * and there might be more issues with decreasing rcoulomb.
353 set->rcut_coulomb = pme_lb->rcut_coulomb_start;
356 if (pme_lb->cutoff_scheme == ecutsVERLET)
358 /* Never decrease the Coulomb and VdW list buffers */
359 set->rlist = std::max(set->rcut_coulomb + pme_lb->rbuf_coulomb,
360 pme_lb->rcut_vdw + pme_lb->rbuf_vdw);
364 tmpr_coulomb = set->rcut_coulomb + pme_lb->rbuf_coulomb;
365 tmpr_vdw = pme_lb->rcut_vdw + pme_lb->rbuf_vdw;
366 set->rlist = std::min(tmpr_coulomb, tmpr_vdw);
370 /* The grid efficiency is the size wrt a grid with uniform x/y/z spacing */
371 set->grid_efficiency = 1;
372 for (d = 0; d < DIM; d++)
374 set->grid_efficiency *= (set->grid[d]*sp)/norm(pme_lb->box_start[d]);
376 /* The Ewald coefficient is inversly proportional to the cut-off */
378 pme_lb->setup[0].ewaldcoeff_q*pme_lb->setup[0].rcut_coulomb/set->rcut_coulomb;
379 /* We set ewaldcoeff_lj in set, even when LJ-PME is not used */
381 pme_lb->setup[0].ewaldcoeff_lj*pme_lb->setup[0].rcut_coulomb/set->rcut_coulomb;
388 fprintf(debug, "PME loadbal: grid %d %d %d, coulomb cutoff %f\n",
389 set->grid[XX], set->grid[YY], set->grid[ZZ], set->rcut_coulomb);
394 /*! \brief Print the PME grid */
395 static void print_grid(FILE *fp_err, FILE *fp_log,
398 const pme_setup_t *set,
401 char buf[STRLEN], buft[STRLEN];
405 sprintf(buft, ": %.1f M-cycles", cycles*1e-6);
411 sprintf(buf, "%-11s%10s pme grid %d %d %d, coulomb cutoff %.3f%s",
413 desc, set->grid[XX], set->grid[YY], set->grid[ZZ], set->rcut_coulomb,
415 if (fp_err != nullptr)
417 fprintf(fp_err, "\r%s\n", buf);
420 if (fp_log != nullptr)
422 fprintf(fp_log, "%s\n", buf);
426 /*! \brief Return the index of the last setup used in PME load balancing */
427 static int pme_loadbal_end(pme_load_balancing_t *pme_lb)
429 /* In the initial stage only n is set; end is not set yet */
440 /*! \brief Print descriptive string about what limits PME load balancing */
441 static void print_loadbal_limited(FILE *fp_err, FILE *fp_log,
443 pme_load_balancing_t *pme_lb)
445 char buf[STRLEN], sbuf[22];
447 sprintf(buf, "step %4s: the %s limits the PME load balancing to a coulomb cut-off of %.3f",
448 gmx_step_str(step, sbuf),
449 pmelblim_str[pme_lb->elimited],
450 pme_lb->setup[pme_loadbal_end(pme_lb)-1].rcut_coulomb);
451 if (fp_err != nullptr)
453 fprintf(fp_err, "\r%s\n", buf);
456 if (fp_log != nullptr)
458 fprintf(fp_log, "%s\n", buf);
462 /*! \brief Switch load balancing to stage 1
464 * In this stage, only reasonably fast setups are run again. */
465 static void switch_to_stage1(pme_load_balancing_t *pme_lb)
467 /* Increase start until we find a setup that is not slower than
468 * maxRelativeSlowdownAccepted times the fastest setup.
470 pme_lb->start = pme_lb->lower_limit;
471 while (pme_lb->start + 1 < pme_lb->n &&
472 (pme_lb->setup[pme_lb->start].count == 0 ||
473 pme_lb->setup[pme_lb->start].cycles >
474 pme_lb->setup[pme_lb->fastest].cycles*maxRelativeSlowdownAccepted))
478 /* While increasing start, we might have skipped setups that we did not
479 * time during stage 0. We want to extend the range for stage 1 to include
480 * any skipped setups that lie between setups that were measured to be
481 * acceptably fast and too slow.
483 while (pme_lb->start > pme_lb->lower_limit &&
484 pme_lb->setup[pme_lb->start - 1].count == 0)
489 /* Decrease end only with setups that we timed and that are slow. */
490 pme_lb->end = pme_lb->n;
491 if (pme_lb->setup[pme_lb->end - 1].count > 0 &&
492 pme_lb->setup[pme_lb->end - 1].cycles >
493 pme_lb->setup[pme_lb->fastest].cycles*maxRelativeSlowdownAccepted)
500 /* Next we want to choose setup pme_lb->end-1, but as we will decrease
501 * pme_lb->cur by one right after returning, we set cur to end.
503 pme_lb->cur = pme_lb->end;
506 /*! \brief Process the timings and try to adjust the PME grid and Coulomb cut-off
508 * The adjustment is done to generate a different non-bonded PP and PME load.
509 * With separate PME ranks (PP and PME on different processes) or with
510 * a GPU (PP on GPU, PME on CPU), PP and PME run on different resources
511 * and changing the load will affect the load balance and performance.
512 * The total time for a set of integration steps is monitored and a range
513 * of grid/cut-off setups is scanned. After calling pme_load_balance many
514 * times and acquiring enough statistics, the best performing setup is chosen.
515 * Here we try to take into account fluctuations and changes due to external
516 * factors as well as DD load balancing.
519 pme_load_balance(pme_load_balancing_t *pme_lb,
523 const gmx::MDLogger &mdlog,
524 const t_inputrec *ir,
527 interaction_const_t *ic,
528 struct nonbonded_verlet_t *nbv,
529 struct gmx_pme_t ** pmedata,
535 char buf[STRLEN], sbuf[22];
540 gmx_sumd(1, &cycles, cr);
541 cycles /= cr->nnodes;
544 set = &pme_lb->setup[pme_lb->cur];
547 rtab = ir->rlist + ir->tabext;
549 if (set->count % 2 == 1)
551 /* Skip the first cycle, because the first step after a switch
552 * is much slower due to allocation and/or caching effects.
557 sprintf(buf, "step %4s: ", gmx_step_str(step, sbuf));
558 print_grid(fp_err, fp_log, buf, "timed with", set, cycles);
562 set->cycles = cycles;
566 if (cycles*maxFluctuationAccepted < set->cycles &&
567 pme_lb->stage == pme_lb->nstage - 1)
569 /* The performance went up a lot (due to e.g. DD load balancing).
570 * Add a stage, keep the minima, but rescan all setups.
576 fprintf(debug, "The performance for grid %d %d %d went from %.3f to %.1f M-cycles, this is more than %f\n"
577 "Increased the number stages to %d"
578 " and ignoring the previous performance\n",
579 set->grid[XX], set->grid[YY], set->grid[ZZ],
580 set->cycles*1e-6, cycles*1e-6, maxFluctuationAccepted,
584 set->cycles = std::min(set->cycles, cycles);
587 if (set->cycles < pme_lb->setup[pme_lb->fastest].cycles)
589 pme_lb->fastest = pme_lb->cur;
591 if (DOMAINDECOMP(cr))
593 /* We found a new fastest setting, ensure that with subsequent
594 * shorter cut-off's the dynamic load balancing does not make
595 * the use of the current cut-off impossible. This solution is
596 * a trade-off, as the PME load balancing and DD domain size
597 * load balancing can interact in complex ways.
598 * With the Verlet kernels, DD load imbalance will usually be
599 * mainly due to bonded interaction imbalance, which will often
600 * quickly push the domain boundaries beyond the limit for the
601 * optimal, PME load balanced, cut-off. But it could be that
602 * better overal performance can be obtained with a slightly
603 * shorter cut-off and better DD load balancing.
605 set_dd_dlb_max_cutoff(cr, pme_lb->setup[pme_lb->fastest].rlist);
608 cycles_fast = pme_lb->setup[pme_lb->fastest].cycles;
610 /* Check in stage 0 if we should stop scanning grids.
611 * Stop when the time is more than maxRelativeSlowDownAccepted longer than the fastest.
613 if (pme_lb->stage == 0 && pme_lb->cur > 0 &&
614 cycles > pme_lb->setup[pme_lb->fastest].cycles*maxRelativeSlowdownAccepted)
616 pme_lb->n = pme_lb->cur + 1;
617 /* Done with scanning, go to stage 1 */
618 switch_to_stage1(pme_lb);
621 if (pme_lb->stage == 0)
625 gridsize_start = set->grid[XX]*set->grid[YY]*set->grid[ZZ];
629 if (pme_lb->cur+1 < pme_lb->n)
631 /* We had already generated the next setup */
636 /* Find the next setup */
637 OK = pme_loadbal_increase_cutoff(pme_lb, ir->pme_order, cr->dd);
641 pme_lb->elimited = epmelblimPMEGRID;
645 if (OK && ir->ePBC != epbcNONE)
647 OK = (gmx::square(pme_lb->setup[pme_lb->cur+1].rlist)
648 <= max_cutoff2(ir->ePBC, state->box));
651 pme_lb->elimited = epmelblimBOX;
659 if (DOMAINDECOMP(cr))
661 OK = change_dd_cutoff(cr, state, ir,
662 pme_lb->setup[pme_lb->cur].rlist);
665 /* Failed: do not use this setup */
667 pme_lb->elimited = epmelblimDD;
673 /* We hit the upper limit for the cut-off,
674 * the setup should not go further than cur.
676 pme_lb->n = pme_lb->cur + 1;
677 print_loadbal_limited(fp_err, fp_log, step, pme_lb);
678 /* Switch to the next stage */
679 switch_to_stage1(pme_lb);
683 !(pme_lb->setup[pme_lb->cur].grid[XX]*
684 pme_lb->setup[pme_lb->cur].grid[YY]*
685 pme_lb->setup[pme_lb->cur].grid[ZZ] <
686 gridsize_start*gridScaleFactor
688 pme_lb->setup[pme_lb->cur].grid_efficiency <
689 pme_lb->setup[pme_lb->cur-1].grid_efficiency*relativeEfficiencyFactor));
692 if (pme_lb->stage > 0 && pme_lb->end == 1)
694 pme_lb->cur = pme_lb->lower_limit;
695 pme_lb->stage = pme_lb->nstage;
697 else if (pme_lb->stage > 0 && pme_lb->end > 1)
699 /* If stage = nstage-1:
700 * scan over all setups, rerunning only those setups
701 * which are not much slower than the fastest
704 * Note that we loop backward to minimize the risk of the cut-off
705 * getting limited by DD DLB, since the DLB cut-off limit is set
706 * to the fastest PME setup.
710 if (pme_lb->cur > pme_lb->start)
718 pme_lb->cur = pme_lb->end - 1;
721 while (pme_lb->stage == pme_lb->nstage - 1 &&
722 pme_lb->setup[pme_lb->cur].count > 0 &&
723 pme_lb->setup[pme_lb->cur].cycles > cycles_fast*maxRelativeSlowdownAccepted);
725 if (pme_lb->stage == pme_lb->nstage)
727 /* We are done optimizing, use the fastest setup we found */
728 pme_lb->cur = pme_lb->fastest;
732 if (DOMAINDECOMP(cr) && pme_lb->stage > 0)
734 OK = change_dd_cutoff(cr, state, ir, pme_lb->setup[pme_lb->cur].rlist);
737 /* For some reason the chosen cut-off is incompatible with DD.
738 * We should continue scanning a more limited range of cut-off's.
740 if (pme_lb->cur > 1 && pme_lb->stage == pme_lb->nstage)
742 /* stage=nstage says we're finished, but we should continue
743 * balancing, so we set back stage which was just incremented.
747 if (pme_lb->cur <= pme_lb->fastest)
749 /* This should not happen, as we set limits on the DLB bounds.
750 * But we implement a complete failsafe solution anyhow.
752 GMX_LOG(mdlog.warning).asParagraph().appendTextFormatted(
753 "The fastest PP/PME load balancing setting (cutoff %.3f nm) is no longer available due to DD DLB or box size limitations", pme_lb->fastest);
754 pme_lb->fastest = pme_lb->lower_limit;
755 pme_lb->start = pme_lb->lower_limit;
757 /* Limit the range to below the current cut-off, scan from start */
758 pme_lb->end = pme_lb->cur;
759 pme_lb->cur = pme_lb->start;
760 pme_lb->elimited = epmelblimDD;
761 print_loadbal_limited(fp_err, fp_log, step, pme_lb);
765 /* Change the Coulomb cut-off and the PME grid */
767 set = &pme_lb->setup[pme_lb->cur];
769 ic->rcoulomb = set->rcut_coulomb;
770 ic->rlist = set->rlist;
771 ic->ewaldcoeff_q = set->ewaldcoeff_q;
772 /* TODO: centralize the code that sets the potentials shifts */
773 if (ic->coulomb_modifier == eintmodPOTSHIFT)
775 ic->sh_ewald = std::erfc(ic->ewaldcoeff_q*ic->rcoulomb);
777 if (EVDW_PME(ic->vdwtype))
779 /* We have PME for both Coulomb and VdW, set rvdw equal to rcoulomb */
780 ic->rvdw = set->rcut_coulomb;
781 ic->ewaldcoeff_lj = set->ewaldcoeff_lj;
782 if (ic->vdw_modifier == eintmodPOTSHIFT)
786 ic->dispersion_shift.cpot = -1.0/gmx::power6(static_cast<double>(ic->rvdw));
787 ic->repulsion_shift.cpot = -1.0/gmx::power12(static_cast<double>(ic->rvdw));
788 ic->sh_invrc6 = -ic->dispersion_shift.cpot;
789 crc2 = gmx::square(ic->ewaldcoeff_lj*ic->rvdw);
790 ic->sh_lj_ewald = (std::exp(-crc2)*(1 + crc2 + 0.5*crc2*crc2) - 1)/gmx::power6(ic->rvdw);
794 /* We always re-initialize the tables whether they are used or not */
795 init_interaction_const_tables(nullptr, ic, rtab);
797 nbnxn_gpu_pme_loadbal_update_param(nbv, ic);
799 /* With tMPI + GPUs some ranks may be sharing GPU(s) and therefore
800 * also sharing texture references. To keep the code simple, we don't
801 * treat texture references as shared resources, but this means that
802 * the coulomb_tab texture ref will get updated by multiple threads.
803 * Hence, to ensure that the non-bonded kernels don't start before all
804 * texture binding operations are finished, we need to wait for all ranks
805 * to arrive here before continuing.
807 * Note that we could omit this barrier if GPUs are not shared (or
808 * texture objects are used), but as this is initialization code, there
809 * is not point in complicating things.
812 if (PAR(cr) && use_GPU(nbv))
816 #endif /* GMX_THREAD_MPI */
818 if (!pme_lb->bSepPMERanks)
820 if (pme_lb->setup[pme_lb->cur].pmedata == nullptr)
822 /* Generate a new PME data structure,
823 * copying part of the old pointers.
825 gmx_pme_reinit(&set->pmedata,
826 cr, pme_lb->setup[0].pmedata, ir,
827 set->grid, set->ewaldcoeff_q, set->ewaldcoeff_lj);
829 *pmedata = set->pmedata;
833 /* Tell our PME-only rank to switch grid */
834 gmx_pme_send_switchgrid(cr, set->grid, set->ewaldcoeff_q, set->ewaldcoeff_lj);
839 print_grid(nullptr, debug, "", "switched to", set, -1);
842 if (pme_lb->stage == pme_lb->nstage)
844 print_grid(fp_err, fp_log, "", "optimal", set, -1);
848 /*! \brief Prepare for another round of PME load balancing
850 * \param[in,out] pme_lb Pointer to PME load balancing struct
851 * \param[in] bDlbUnlocked TRUE is DLB was locked and is now unlocked
853 * If the conditions (e.g. DLB off/on, CPU/GPU throttling etc.) changed,
854 * the PP/PME balance might change and re-balancing can improve performance.
855 * This function adds 2 stages and adjusts the considered setup range.
857 static void continue_pme_loadbal(pme_load_balancing_t *pme_lb,
858 gmx_bool bDlbUnlocked)
860 /* Add 2 tuning stages, keep the detected end of the setup range */
862 if (bDlbUnlocked && pme_lb->bSepPMERanks)
864 /* With separate PME ranks, DLB should always lower the PP load and
865 * can only increase the PME load (more communication and imbalance),
866 * so we only need to scan longer cut-off's.
868 pme_lb->lower_limit = pme_lb->cur;
870 pme_lb->start = pme_lb->lower_limit;
873 void pme_loadbal_do(pme_load_balancing_t *pme_lb,
877 const gmx::MDLogger &mdlog,
878 const t_inputrec *ir,
881 gmx_wallcycle_t wcycle,
883 gmx_int64_t step_rel,
889 assert(pme_lb != NULL);
891 if (!pme_lb->bActive)
896 n_prev = pme_lb->cycles_n;
897 cycles_prev = pme_lb->cycles_c;
898 wallcycle_get(wcycle, ewcSTEP, &pme_lb->cycles_n, &pme_lb->cycles_c);
900 /* Before the first step we haven't done any steps yet.
901 * Also handle cases where ir->init_step % ir->nstlist != 0.
903 if (pme_lb->cycles_n < ir->nstlist)
907 /* Sanity check, we expect nstlist cycle counts */
908 if (pme_lb->cycles_n - n_prev != ir->nstlist)
910 /* We could return here, but it's safer to issue an error and quit */
911 gmx_incons("pme_loadbal_do called at an interval != nstlist");
914 /* PME grid + cut-off optimization with GPUs or PME ranks */
915 if (!pme_lb->bBalance && pme_lb->bSepPMERanks)
917 if (pme_lb->bTriggerOnDLB)
919 pme_lb->bBalance = dd_dlb_is_on(cr->dd);
921 /* We should ignore the first timing to avoid timing allocation
922 * overhead. And since the PME load balancing is called just
923 * before DD repartitioning, the ratio returned by dd_pme_f_ratio
924 * is not over the last nstlist steps, but the nstlist steps before
925 * that. So the first useful ratio is available at step_rel=3*nstlist.
927 else if (step_rel >= 3*ir->nstlist)
929 if (DDMASTER(cr->dd))
931 /* If PME rank load is too high, start tuning */
933 (dd_pme_f_ratio(cr->dd) >= loadBalanceTriggerFactor);
935 dd_bcast(cr->dd, sizeof(gmx_bool), &pme_lb->bBalance);
938 pme_lb->bActive = (pme_lb->bBalance ||
939 step_rel <= pme_lb->step_rel_stop);
942 /* The location in the code of this balancing termination is strange.
943 * You would expect to have it after the call to pme_load_balance()
944 * below, since there pme_lb->stage is updated.
945 * But when terminating directly after deciding on and selecting the
946 * optimal setup, DLB will turn on right away if it was locked before.
947 * This might be due to PME reinitialization. So we check stage here
948 * to allow for another nstlist steps with DLB locked to stabilize
951 if (pme_lb->bBalance && pme_lb->stage == pme_lb->nstage)
953 pme_lb->bBalance = FALSE;
955 if (DOMAINDECOMP(cr) && dd_dlb_is_locked(cr->dd))
957 /* Unlock the DLB=auto, DLB is allowed to activate */
958 dd_dlb_unlock(cr->dd);
959 GMX_LOG(mdlog.warning).asParagraph().appendText("NOTE: DLB can now turn on, when beneficial");
961 /* We don't deactivate the tuning yet, since we will balance again
962 * after DLB gets turned on, if it does within PMETune_period.
964 continue_pme_loadbal(pme_lb, TRUE);
965 pme_lb->bTriggerOnDLB = TRUE;
966 pme_lb->step_rel_stop = step_rel + PMETunePeriod*ir->nstlist;
970 /* We're completely done with PME tuning */
971 pme_lb->bActive = FALSE;
974 if (DOMAINDECOMP(cr))
976 /* Set the cut-off limit to the final selected cut-off,
977 * so we don't have artificial DLB limits.
978 * This also ensures that we won't disable the currently
979 * optimal setting during a second round of PME balancing.
981 set_dd_dlb_max_cutoff(cr, fr->ic->rlist);
985 if (pme_lb->bBalance)
987 /* We might not have collected nstlist steps in cycles yet,
988 * since init_step might not be a multiple of nstlist,
989 * but the first data collected is skipped anyhow.
991 pme_load_balance(pme_lb, cr,
992 fp_err, fp_log, mdlog,
993 ir, state, pme_lb->cycles_c - cycles_prev,
994 fr->ic, fr->nbv, &fr->pmedata,
997 /* Update constants in forcerec/inputrec to keep them in sync with fr->ic */
998 fr->ewaldcoeff_q = fr->ic->ewaldcoeff_q;
999 fr->ewaldcoeff_lj = fr->ic->ewaldcoeff_lj;
1000 fr->rlist = fr->ic->rlist;
1001 fr->rcoulomb = fr->ic->rcoulomb;
1002 fr->rvdw = fr->ic->rvdw;
1004 if (ir->eDispCorr != edispcNO)
1006 calc_enervirdiff(nullptr, ir->eDispCorr, fr);
1010 if (!pme_lb->bBalance &&
1011 (!pme_lb->bSepPMERanks || step_rel > pme_lb->step_rel_stop))
1013 /* We have just deactivated the balancing and we're not measuring PP/PME
1014 * imbalance during the first steps of the run: deactivate the tuning.
1016 pme_lb->bActive = FALSE;
1019 if (!(pme_lb->bActive) && DOMAINDECOMP(cr) && dd_dlb_is_locked(cr->dd))
1021 /* Make sure DLB is allowed when we deactivate PME tuning */
1022 dd_dlb_unlock(cr->dd);
1023 GMX_LOG(mdlog.warning).asParagraph().appendText("NOTE: DLB can now turn on, when beneficial");
1026 *bPrinting = pme_lb->bBalance;
1029 /*! \brief Return product of the number of PME grid points in each dimension */
1030 static int pme_grid_points(const pme_setup_t *setup)
1032 return setup->grid[XX]*setup->grid[YY]*setup->grid[ZZ];
1035 /*! \brief Print one load-balancing setting */
1036 static void print_pme_loadbal_setting(FILE *fplog,
1038 const pme_setup_t *setup)
1041 " %-7s %6.3f nm %6.3f nm %3d %3d %3d %5.3f nm %5.3f nm\n",
1043 setup->rcut_coulomb, setup->rlist,
1044 setup->grid[XX], setup->grid[YY], setup->grid[ZZ],
1045 setup->spacing, 1/setup->ewaldcoeff_q);
1048 /*! \brief Print all load-balancing settings */
1049 static void print_pme_loadbal_settings(pme_load_balancing_t *pme_lb,
1051 const gmx::MDLogger &mdlog,
1052 gmx_bool bNonBondedOnGPU)
1054 double pp_ratio, grid_ratio;
1055 real pp_ratio_temporary;
1057 pp_ratio_temporary = pme_lb->setup[pme_lb->cur].rlist / pme_lb->setup[0].rlist;
1058 pp_ratio = gmx::power3(pp_ratio_temporary);
1059 grid_ratio = pme_grid_points(&pme_lb->setup[pme_lb->cur])/
1060 (double)pme_grid_points(&pme_lb->setup[0]);
1062 fprintf(fplog, "\n");
1063 fprintf(fplog, " P P - P M E L O A D B A L A N C I N G\n");
1064 fprintf(fplog, "\n");
1065 /* Here we only warn when the optimal setting is the last one */
1066 if (pme_lb->elimited != epmelblimNO &&
1067 pme_lb->cur == pme_loadbal_end(pme_lb)-1)
1069 fprintf(fplog, " NOTE: The PP/PME load balancing was limited by the %s,\n",
1070 pmelblim_str[pme_lb->elimited]);
1071 fprintf(fplog, " you might not have reached a good load balance.\n");
1072 if (pme_lb->elimited == epmelblimDD)
1074 fprintf(fplog, " Try different mdrun -dd settings or lower the -dds value.\n");
1076 fprintf(fplog, "\n");
1078 fprintf(fplog, " PP/PME load balancing changed the cut-off and PME settings:\n");
1079 fprintf(fplog, " particle-particle PME\n");
1080 fprintf(fplog, " rcoulomb rlist grid spacing 1/beta\n");
1081 print_pme_loadbal_setting(fplog, "initial", &pme_lb->setup[0]);
1082 print_pme_loadbal_setting(fplog, "final", &pme_lb->setup[pme_lb->cur]);
1083 fprintf(fplog, " cost-ratio %4.2f %4.2f\n",
1084 pp_ratio, grid_ratio);
1085 fprintf(fplog, " (note that these numbers concern only part of the total PP and PME load)\n");
1087 if (pp_ratio > 1.5 && !bNonBondedOnGPU)
1089 GMX_LOG(mdlog.warning).asParagraph().appendText(
1090 "NOTE: PME load balancing increased the non-bonded workload by more than 50%.\n"
1091 " For better performance, use (more) PME ranks (mdrun -npme),\n"
1092 " or if you are beyond the scaling limit, use fewer total ranks (or nodes).");
1096 fprintf(fplog, "\n");
1100 void pme_loadbal_done(pme_load_balancing_t *pme_lb,
1102 const gmx::MDLogger &mdlog,
1103 gmx_bool bNonBondedOnGPU)
1105 if (fplog != nullptr && (pme_lb->cur > 0 || pme_lb->elimited != epmelblimNO))
1107 print_pme_loadbal_settings(pme_lb, fplog, mdlog, bNonBondedOnGPU);
1110 /* TODO: Here we should free all pointers in pme_lb,
1111 * but as it contains pme data structures,
1112 * we need to first make pme.c free all data.