2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2015,2016,2017,2018,2019, 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"
53 #include "gromacs/domdec/dlb.h"
54 #include "gromacs/domdec/domdec.h"
55 #include "gromacs/domdec/domdec_network.h"
56 #include "gromacs/domdec/domdec_struct.h"
57 #include "gromacs/domdec/partition.h"
58 #include "gromacs/ewald/ewald_utils.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/sim_util.h"
66 #include "gromacs/mdtypes/commrec.h"
67 #include "gromacs/mdtypes/inputrec.h"
68 #include "gromacs/mdtypes/md_enums.h"
69 #include "gromacs/mdtypes/state.h"
70 #include "gromacs/nbnxm/gpu_data_mgmt.h"
71 #include "gromacs/nbnxm/nbnxm.h"
72 #include "gromacs/pbcutil/pbc.h"
73 #include "gromacs/timing/wallcycle.h"
74 #include "gromacs/utility/cstringutil.h"
75 #include "gromacs/utility/fatalerror.h"
76 #include "gromacs/utility/gmxassert.h"
77 #include "gromacs/utility/logger.h"
78 #include "gromacs/utility/smalloc.h"
79 #include "gromacs/utility/strconvert.h"
81 #include "pme_internal.h"
83 /*! \brief Parameters and settings for one PP-PME setup */
85 real rcut_coulomb; /**< Coulomb cut-off */
86 real rlistOuter; /**< cut-off for the outer pair-list */
87 real rlistInner; /**< cut-off for the inner pair-list */
88 real spacing; /**< (largest) PME grid spacing */
89 ivec grid; /**< the PME grid dimensions */
90 real grid_efficiency; /**< ineffiency factor for non-uniform grids <= 1 */
91 real ewaldcoeff_q; /**< Electrostatic Ewald coefficient */
92 real ewaldcoeff_lj; /**< LJ Ewald coefficient, only for the call to send_switchgrid */
93 struct gmx_pme_t *pmedata; /**< the data structure used in the PME code */
94 int count; /**< number of times this setup has been timed */
95 double cycles; /**< the fastest time for this setup in cycles */
98 /*! \brief After 50 nstlist periods of not observing imbalance: never tune PME */
99 const int PMETunePeriod = 50;
100 /*! \brief Trigger PME load balancing at more than 5% PME overload */
101 const real loadBalanceTriggerFactor = 1.05;
102 /*! \brief Scale the grid by a most at factor 1.7.
104 * This still leaves room for about 4-4.5x decrease in grid spacing while limiting the cases where
105 * large imbalance leads to extreme cutoff scaling for marginal benefits.
107 * This should help to avoid:
108 * - large increase in power consumption for little performance gain
109 * - increasing communication volume
112 const real c_maxSpacingScaling = 1.7;
113 /*! \brief In the initial scan, step by grids that are at least a factor 0.8 coarser */
114 const real gridpointsScaleFactor = 0.8;
115 /*! \brief In the initial scan, try to skip grids with uneven x/y/z spacing,
116 * checking if the "efficiency" is more than 5% worse than the previous grid.
118 const real relativeEfficiencyFactor = 1.05;
119 /*! \brief Rerun until a run is 12% slower setups than the fastest run so far */
120 const real maxRelativeSlowdownAccepted = 1.12;
121 /*! \brief If setups get more than 2% faster, do another round to avoid
122 * choosing a slower setup due to acceleration or fluctuations.
124 const real maxFluctuationAccepted = 1.02;
126 /*! \brief Enumeration whose values describe the effect limiting the load balancing */
128 epmelblimNO, epmelblimBOX, epmelblimDD, epmelblimPMEGRID, epmelblimMAXSCALING, epmelblimNR
131 /*! \brief Descriptive strings matching ::epmelb */
132 static const char *pmelblim_str[epmelblimNR] =
133 { "no", "box size", "domain decompostion", "PME grid restriction", "maximum allowed grid scaling" };
135 struct pme_load_balancing_t {
136 gmx_bool bSepPMERanks; /**< do we have separate PME ranks? */
137 gmx_bool bActive; /**< is PME tuning active? */
138 int64_t step_rel_stop; /**< stop the tuning after this value of step_rel */
139 gmx_bool bTriggerOnDLB; /**< trigger balancing only on DD DLB */
140 gmx_bool bBalance; /**< are we in the balancing phase, i.e. trying different setups? */
141 int nstage; /**< the current maximum number of stages */
143 real cut_spacing; /**< the minimum cutoff / PME grid spacing ratio */
144 real rcut_vdw; /**< Vdw cutoff (does not change) */
145 real rcut_coulomb_start; /**< Initial electrostatics cutoff */
146 real rbufOuter_coulomb; /**< the outer pairlist buffer size */
147 real rbufOuter_vdw; /**< the outer pairlist buffer size */
148 real rbufInner_coulomb; /**< the inner pairlist buffer size */
149 real rbufInner_vdw; /**< the inner pairlist buffer size */
150 matrix box_start; /**< the initial simulation box */
151 std::vector<pme_setup_t> setup; /**< the PME+cutoff setups */
152 int cur; /**< the index (in setup) of the current setup */
153 int fastest; /**< index of the fastest setup up till now */
154 int lower_limit; /**< don't go below this setup index */
155 int start; /**< start of setup index range to consider in stage>0 */
156 int end; /**< end of setup index range to consider in stage>0 */
157 int elimited; /**< was the balancing limited, uses enum above */
158 int cutoff_scheme; /**< Verlet or group cut-offs */
160 int stage; /**< the current stage */
162 int cycles_n; /**< step cycle counter cummulative count */
163 double cycles_c; /**< step cycle counter cummulative cycles */
166 /* TODO The code in this file should call this getter, rather than
167 * read bActive anywhere */
168 bool pme_loadbal_is_active(const pme_load_balancing_t *pme_lb)
170 return pme_lb != nullptr && pme_lb->bActive;
173 // TODO Return a unique_ptr to pme_load_balancing_t
174 void pme_loadbal_init(pme_load_balancing_t **pme_lb_p,
176 const gmx::MDLogger &mdlog,
177 const t_inputrec &ir,
179 const interaction_const_t &ic,
180 const NbnxnListParameters &listParams,
185 GMX_RELEASE_ASSERT(ir.cutoff_scheme != ecutsGROUP, "PME tuning is not supported with cutoff-scheme=group (because it contains bugs)");
187 pme_load_balancing_t *pme_lb;
191 // Note that we don't (yet) support PME load balancing with LJ-PME only.
192 GMX_RELEASE_ASSERT(EEL_PME(ir.coulombtype), "pme_loadbal_init called without PME electrostatics");
193 // To avoid complexity, we require a single cut-off with PME for q+LJ.
194 // This is checked by grompp, but it doesn't hurt to check again.
195 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");
197 pme_lb = new pme_load_balancing_t;
199 pme_lb->bSepPMERanks = !thisRankHasDuty(cr, DUTY_PME);
201 /* Initially we turn on balancing directly on based on PP/PME imbalance */
202 pme_lb->bTriggerOnDLB = FALSE;
204 /* Any number of stages >= 2 is supported */
207 pme_lb->cutoff_scheme = ir.cutoff_scheme;
209 pme_lb->rbufOuter_coulomb = listParams.rlistOuter - ic.rcoulomb;
210 pme_lb->rbufOuter_vdw = listParams.rlistOuter - ic.rvdw;
211 pme_lb->rbufInner_coulomb = listParams.rlistInner - ic.rcoulomb;
212 pme_lb->rbufInner_vdw = listParams.rlistInner - ic.rvdw;
214 /* Scale box with Ewald wall factor; note that we pmedata->boxScaler
215 * can't always usedd as it's not available with separate PME ranks.
217 EwaldBoxZScaler boxScaler(ir);
218 boxScaler.scaleBox(box, pme_lb->box_start);
220 pme_lb->setup.resize(1);
222 pme_lb->rcut_vdw = ic.rvdw;
223 pme_lb->rcut_coulomb_start = ir.rcoulomb;
226 pme_lb->setup[0].rcut_coulomb = ic.rcoulomb;
227 pme_lb->setup[0].rlistOuter = listParams.rlistOuter;
228 pme_lb->setup[0].rlistInner = listParams.rlistInner;
229 pme_lb->setup[0].grid[XX] = ir.nkx;
230 pme_lb->setup[0].grid[YY] = ir.nky;
231 pme_lb->setup[0].grid[ZZ] = ir.nkz;
232 pme_lb->setup[0].ewaldcoeff_q = ic.ewaldcoeff_q;
233 pme_lb->setup[0].ewaldcoeff_lj = ic.ewaldcoeff_lj;
235 if (!pme_lb->bSepPMERanks)
237 GMX_RELEASE_ASSERT(pmedata, "On ranks doing both PP and PME we need a valid pmedata object");
238 pme_lb->setup[0].pmedata = pmedata;
242 for (d = 0; d < DIM; d++)
244 sp = norm(pme_lb->box_start[d])/pme_lb->setup[0].grid[d];
250 pme_lb->setup[0].spacing = spm;
252 if (ir.fourier_spacing > 0)
254 pme_lb->cut_spacing = ir.rcoulomb/ir.fourier_spacing;
258 pme_lb->cut_spacing = ir.rcoulomb/pme_lb->setup[0].spacing;
264 pme_lb->lower_limit = 0;
267 pme_lb->elimited = epmelblimNO;
269 pme_lb->cycles_n = 0;
270 pme_lb->cycles_c = 0;
272 if (!wallcycle_have_counter())
274 GMX_LOG(mdlog.warning).asParagraph().appendText("NOTE: Cycle counters unsupported or not enabled in kernel. Cannot use PME-PP balancing.");
277 /* Tune with GPUs and/or separate PME ranks.
278 * When running only on a CPU without PME ranks, PME tuning will only help
279 * with small numbers of atoms in the cut-off sphere.
281 pme_lb->bActive = (wallcycle_have_counter() && (bUseGPU ||
282 pme_lb->bSepPMERanks));
284 /* With GPUs and no separate PME ranks we can't measure the PP/PME
285 * imbalance, so we start balancing right away.
286 * Otherwise we only start balancing after we observe imbalance.
288 pme_lb->bBalance = (pme_lb->bActive && (bUseGPU && !pme_lb->bSepPMERanks));
290 pme_lb->step_rel_stop = PMETunePeriod*ir.nstlist;
292 /* Delay DD load balancing when GPUs are used */
293 if (pme_lb->bActive && DOMAINDECOMP(cr) && cr->dd->nnodes > 1 && bUseGPU)
295 /* Lock DLB=auto to off (does nothing when DLB=yes/no.
296 * With GPUs and separate PME nodes, we want to first
297 * do PME tuning without DLB, since DLB might limit
298 * the cut-off, which never improves performance.
299 * We allow for DLB + PME tuning after a first round of tuning.
302 if (dd_dlb_is_locked(cr->dd))
304 GMX_LOG(mdlog.warning).asParagraph().appendText("NOTE: DLB will not turn on during the first phase of PME tuning");
310 *bPrinting = pme_lb->bBalance;
313 /*! \brief Try to increase the cutoff during load balancing */
314 static gmx_bool pme_loadbal_increase_cutoff(pme_load_balancing_t *pme_lb,
316 const gmx_domdec_t *dd)
319 real tmpr_coulomb, tmpr_vdw;
323 /* Try to add a new setup with next larger cut-off to the list */
326 set.pmedata = nullptr;
328 NumPmeDomains numPmeDomains = getNumPmeDomains(dd);
333 /* Avoid infinite while loop, which can occur at the minimum grid size.
334 * Note that in practice load balancing will stop before this point.
335 * The factor 2.1 allows for the extreme case in which only grids
336 * of powers of 2 are allowed (the current code supports more grids).
344 clear_ivec(set.grid);
345 sp = calcFftGrid(nullptr, pme_lb->box_start,
346 fac*pme_lb->setup[pme_lb->cur].spacing,
347 minimalPmeGridSize(pme_order),
352 /* As here we can't easily check if one of the PME ranks
353 * uses threading, we do a conservative grid check.
354 * This means we can't use pme_order or less grid lines
355 * per PME rank along x, which is not a strong restriction.
357 grid_ok = gmx_pme_check_restrictions(pme_order,
358 set.grid[XX], set.grid[YY], set.grid[ZZ],
363 while (sp <= 1.001*pme_lb->setup[pme_lb->cur].spacing || !grid_ok);
365 set.rcut_coulomb = pme_lb->cut_spacing*sp;
366 if (set.rcut_coulomb < pme_lb->rcut_coulomb_start)
368 /* This is unlikely, but can happen when e.g. continuing from
369 * a checkpoint after equilibration where the box shrank a lot.
370 * We want to avoid rcoulomb getting smaller than rvdw
371 * and there might be more issues with decreasing rcoulomb.
373 set.rcut_coulomb = pme_lb->rcut_coulomb_start;
376 if (pme_lb->cutoff_scheme == ecutsVERLET)
378 /* Never decrease the Coulomb and VdW list buffers */
379 set.rlistOuter = std::max(set.rcut_coulomb + pme_lb->rbufOuter_coulomb,
380 pme_lb->rcut_vdw + pme_lb->rbufOuter_vdw);
381 set.rlistInner = std::max(set.rcut_coulomb + pme_lb->rbufInner_coulomb,
382 pme_lb->rcut_vdw + pme_lb->rbufInner_vdw);
386 /* TODO Remove these lines and pme_lb->cutoff_scheme */
387 tmpr_coulomb = set.rcut_coulomb + pme_lb->rbufOuter_coulomb;
388 tmpr_vdw = pme_lb->rcut_vdw + pme_lb->rbufOuter_vdw;
389 /* Two (known) bugs with cutoff-scheme=group here:
390 * - This modification of rlist results in incorrect DD comunication.
391 * - We should set fr->bTwinRange = (fr->rlistlong > fr->rlist).
393 set.rlistOuter = std::min(tmpr_coulomb, tmpr_vdw);
394 set.rlistInner = set.rlistOuter;
398 /* The grid efficiency is the size wrt a grid with uniform x/y/z spacing */
399 set.grid_efficiency = 1;
400 for (d = 0; d < DIM; d++)
402 set.grid_efficiency *= (set.grid[d]*sp)/norm(pme_lb->box_start[d]);
404 /* The Ewald coefficient is inversly proportional to the cut-off */
406 pme_lb->setup[0].ewaldcoeff_q*pme_lb->setup[0].rcut_coulomb/set.rcut_coulomb;
407 /* We set ewaldcoeff_lj in set, even when LJ-PME is not used */
409 pme_lb->setup[0].ewaldcoeff_lj*pme_lb->setup[0].rcut_coulomb/set.rcut_coulomb;
416 fprintf(debug, "PME loadbal: grid %d %d %d, coulomb cutoff %f\n",
417 set.grid[XX], set.grid[YY], set.grid[ZZ], set.rcut_coulomb);
419 pme_lb->setup.push_back(set);
423 /*! \brief Print the PME grid */
424 static void print_grid(FILE *fp_err, FILE *fp_log,
427 const pme_setup_t *set,
430 auto buf = gmx::formatString("%-11s%10s pme grid %d %d %d, coulomb cutoff %.3f",
432 set->grid[XX], set->grid[YY], set->grid[ZZ], set->rcut_coulomb);
435 buf += gmx::formatString(": %.1f M-cycles", cycles*1e-6);
437 if (fp_err != nullptr)
439 fprintf(fp_err, "\r%s\n", buf.c_str());
442 if (fp_log != nullptr)
444 fprintf(fp_log, "%s\n", buf.c_str());
448 /*! \brief Return the index of the last setup used in PME load balancing */
449 static int pme_loadbal_end(pme_load_balancing_t *pme_lb)
451 /* In the initial stage only n is set; end is not set yet */
458 return pme_lb->setup.size();
462 /*! \brief Print descriptive string about what limits PME load balancing */
463 static void print_loadbal_limited(FILE *fp_err, FILE *fp_log,
465 pme_load_balancing_t *pme_lb)
467 auto buf = gmx::formatString("step %4s: the %s limits the PME load balancing to a coulomb cut-off of %.3f",
468 gmx::int64ToString(step).c_str(),
469 pmelblim_str[pme_lb->elimited],
470 pme_lb->setup[pme_loadbal_end(pme_lb)-1].rcut_coulomb);
471 if (fp_err != nullptr)
473 fprintf(fp_err, "\r%s\n", buf.c_str());
476 if (fp_log != nullptr)
478 fprintf(fp_log, "%s\n", buf.c_str());
482 /*! \brief Switch load balancing to stage 1
484 * In this stage, only reasonably fast setups are run again. */
485 static void switch_to_stage1(pme_load_balancing_t *pme_lb)
487 /* Increase start until we find a setup that is not slower than
488 * maxRelativeSlowdownAccepted times the fastest setup.
490 pme_lb->start = pme_lb->lower_limit;
491 while (pme_lb->start + 1 < gmx::ssize(pme_lb->setup) &&
492 (pme_lb->setup[pme_lb->start].count == 0 ||
493 pme_lb->setup[pme_lb->start].cycles >
494 pme_lb->setup[pme_lb->fastest].cycles*maxRelativeSlowdownAccepted))
498 /* While increasing start, we might have skipped setups that we did not
499 * time during stage 0. We want to extend the range for stage 1 to include
500 * any skipped setups that lie between setups that were measured to be
501 * acceptably fast and too slow.
503 while (pme_lb->start > pme_lb->lower_limit &&
504 pme_lb->setup[pme_lb->start - 1].count == 0)
509 /* Decrease end only with setups that we timed and that are slow. */
510 pme_lb->end = pme_lb->setup.size();
511 if (pme_lb->setup[pme_lb->end - 1].count > 0 &&
512 pme_lb->setup[pme_lb->end - 1].cycles >
513 pme_lb->setup[pme_lb->fastest].cycles*maxRelativeSlowdownAccepted)
520 /* Next we want to choose setup pme_lb->end-1, but as we will decrease
521 * pme_lb->cur by one right after returning, we set cur to end.
523 pme_lb->cur = pme_lb->end;
526 /*! \brief Process the timings and try to adjust the PME grid and Coulomb cut-off
528 * The adjustment is done to generate a different non-bonded PP and PME load.
529 * With separate PME ranks (PP and PME on different processes) or with
530 * a GPU (PP on GPU, PME on CPU), PP and PME run on different resources
531 * and changing the load will affect the load balance and performance.
532 * The total time for a set of integration steps is monitored and a range
533 * of grid/cut-off setups is scanned. After calling pme_load_balance many
534 * times and acquiring enough statistics, the best performing setup is chosen.
535 * Here we try to take into account fluctuations and changes due to external
536 * factors as well as DD load balancing.
539 pme_load_balance(pme_load_balancing_t *pme_lb,
543 const gmx::MDLogger &mdlog,
544 const t_inputrec &ir,
545 const t_state &state,
547 interaction_const_t *ic,
548 struct nonbonded_verlet_t *nbv,
549 struct gmx_pme_t ** pmedata,
555 char buf[STRLEN], sbuf[22];
560 gmx_sumd(1, &cycles, cr);
561 cycles /= cr->nnodes;
564 set = &pme_lb->setup[pme_lb->cur];
567 rtab = ir.rlist + ir.tabext;
569 if (set->count % 2 == 1)
571 /* Skip the first cycle, because the first step after a switch
572 * is much slower due to allocation and/or caching effects.
577 sprintf(buf, "step %4s: ", gmx_step_str(step, sbuf));
578 print_grid(fp_err, fp_log, buf, "timed with", set, cycles);
582 set->cycles = cycles;
586 if (cycles*maxFluctuationAccepted < set->cycles &&
587 pme_lb->stage == pme_lb->nstage - 1)
589 /* The performance went up a lot (due to e.g. DD load balancing).
590 * Add a stage, keep the minima, but rescan all setups.
596 fprintf(debug, "The performance for grid %d %d %d went from %.3f to %.1f M-cycles, this is more than %f\n"
597 "Increased the number stages to %d"
598 " and ignoring the previous performance\n",
599 set->grid[XX], set->grid[YY], set->grid[ZZ],
600 set->cycles*1e-6, cycles*1e-6, maxFluctuationAccepted,
604 set->cycles = std::min(set->cycles, cycles);
607 if (set->cycles < pme_lb->setup[pme_lb->fastest].cycles)
609 pme_lb->fastest = pme_lb->cur;
611 if (DOMAINDECOMP(cr))
613 /* We found a new fastest setting, ensure that with subsequent
614 * shorter cut-off's the dynamic load balancing does not make
615 * the use of the current cut-off impossible. This solution is
616 * a trade-off, as the PME load balancing and DD domain size
617 * load balancing can interact in complex ways.
618 * With the Verlet kernels, DD load imbalance will usually be
619 * mainly due to bonded interaction imbalance, which will often
620 * quickly push the domain boundaries beyond the limit for the
621 * optimal, PME load balanced, cut-off. But it could be that
622 * better overal performance can be obtained with a slightly
623 * shorter cut-off and better DD load balancing.
625 set_dd_dlb_max_cutoff(cr, pme_lb->setup[pme_lb->fastest].rlistOuter);
628 cycles_fast = pme_lb->setup[pme_lb->fastest].cycles;
630 /* Check in stage 0 if we should stop scanning grids.
631 * Stop when the time is more than maxRelativeSlowDownAccepted longer than the fastest.
633 if (pme_lb->stage == 0 && pme_lb->cur > 0 &&
634 cycles > pme_lb->setup[pme_lb->fastest].cycles*maxRelativeSlowdownAccepted)
636 pme_lb->setup.resize(pme_lb->cur + 1);
637 /* Done with scanning, go to stage 1 */
638 switch_to_stage1(pme_lb);
641 if (pme_lb->stage == 0)
645 gridsize_start = set->grid[XX]*set->grid[YY]*set->grid[ZZ];
649 if (pme_lb->cur+1 < gmx::ssize(pme_lb->setup))
651 /* We had already generated the next setup */
656 /* Find the next setup */
657 OK = pme_loadbal_increase_cutoff(pme_lb, ir.pme_order, cr->dd);
661 pme_lb->elimited = epmelblimPMEGRID;
666 pme_lb->setup[pme_lb->cur+1].spacing > c_maxSpacingScaling*pme_lb->setup[0].spacing)
669 pme_lb->elimited = epmelblimMAXSCALING;
672 if (OK && ir.ePBC != epbcNONE)
674 OK = (gmx::square(pme_lb->setup[pme_lb->cur+1].rlistOuter)
675 <= max_cutoff2(ir.ePBC, state.box));
678 pme_lb->elimited = epmelblimBOX;
686 if (DOMAINDECOMP(cr))
688 OK = change_dd_cutoff(cr, state,
689 pme_lb->setup[pme_lb->cur].rlistOuter);
692 /* Failed: do not use this setup */
694 pme_lb->elimited = epmelblimDD;
700 /* We hit the upper limit for the cut-off,
701 * the setup should not go further than cur.
703 pme_lb->setup.resize(pme_lb->cur + 1);
704 print_loadbal_limited(fp_err, fp_log, step, pme_lb);
705 /* Switch to the next stage */
706 switch_to_stage1(pme_lb);
710 !(pme_lb->setup[pme_lb->cur].grid[XX]*
711 pme_lb->setup[pme_lb->cur].grid[YY]*
712 pme_lb->setup[pme_lb->cur].grid[ZZ] <
713 gridsize_start*gridpointsScaleFactor
715 pme_lb->setup[pme_lb->cur].grid_efficiency <
716 pme_lb->setup[pme_lb->cur-1].grid_efficiency*relativeEfficiencyFactor));
719 if (pme_lb->stage > 0 && pme_lb->end == 1)
721 pme_lb->cur = pme_lb->lower_limit;
722 pme_lb->stage = pme_lb->nstage;
724 else if (pme_lb->stage > 0 && pme_lb->end > 1)
726 /* If stage = nstage-1:
727 * scan over all setups, rerunning only those setups
728 * which are not much slower than the fastest
731 * Note that we loop backward to minimize the risk of the cut-off
732 * getting limited by DD DLB, since the DLB cut-off limit is set
733 * to the fastest PME setup.
737 if (pme_lb->cur > pme_lb->start)
745 pme_lb->cur = pme_lb->end - 1;
748 while (pme_lb->stage == pme_lb->nstage - 1 &&
749 pme_lb->setup[pme_lb->cur].count > 0 &&
750 pme_lb->setup[pme_lb->cur].cycles > cycles_fast*maxRelativeSlowdownAccepted);
752 if (pme_lb->stage == pme_lb->nstage)
754 /* We are done optimizing, use the fastest setup we found */
755 pme_lb->cur = pme_lb->fastest;
759 if (DOMAINDECOMP(cr) && pme_lb->stage > 0)
761 OK = change_dd_cutoff(cr, state, pme_lb->setup[pme_lb->cur].rlistOuter);
764 /* For some reason the chosen cut-off is incompatible with DD.
765 * We should continue scanning a more limited range of cut-off's.
767 if (pme_lb->cur > 1 && pme_lb->stage == pme_lb->nstage)
769 /* stage=nstage says we're finished, but we should continue
770 * balancing, so we set back stage which was just incremented.
774 if (pme_lb->cur <= pme_lb->fastest)
776 /* This should not happen, as we set limits on the DLB bounds.
777 * But we implement a complete failsafe solution anyhow.
779 GMX_LOG(mdlog.warning).asParagraph().appendTextFormatted(
780 "The fastest PP/PME load balancing setting (cutoff %.3d nm) is no longer available due to DD DLB or box size limitations", pme_lb->fastest);
781 pme_lb->fastest = pme_lb->lower_limit;
782 pme_lb->start = pme_lb->lower_limit;
784 /* Limit the range to below the current cut-off, scan from start */
785 pme_lb->end = pme_lb->cur;
786 pme_lb->cur = pme_lb->start;
787 pme_lb->elimited = epmelblimDD;
788 print_loadbal_limited(fp_err, fp_log, step, pme_lb);
792 /* Change the Coulomb cut-off and the PME grid */
794 set = &pme_lb->setup[pme_lb->cur];
796 NbnxnListParameters *listParams = nbv->listParams.get();
798 ic->rcoulomb = set->rcut_coulomb;
799 listParams->rlistOuter = set->rlistOuter;
800 listParams->rlistInner = set->rlistInner;
801 ic->ewaldcoeff_q = set->ewaldcoeff_q;
802 /* TODO: centralize the code that sets the potentials shifts */
803 if (ic->coulomb_modifier == eintmodPOTSHIFT)
805 GMX_RELEASE_ASSERT(ic->rcoulomb != 0, "Cutoff radius cannot be zero");
806 ic->sh_ewald = std::erfc(ic->ewaldcoeff_q*ic->rcoulomb) / ic->rcoulomb;
808 if (EVDW_PME(ic->vdwtype))
810 /* We have PME for both Coulomb and VdW, set rvdw equal to rcoulomb */
811 ic->rvdw = set->rcut_coulomb;
812 ic->ewaldcoeff_lj = set->ewaldcoeff_lj;
813 if (ic->vdw_modifier == eintmodPOTSHIFT)
817 ic->dispersion_shift.cpot = -1.0/gmx::power6(static_cast<double>(ic->rvdw));
818 ic->repulsion_shift.cpot = -1.0/gmx::power12(static_cast<double>(ic->rvdw));
819 ic->sh_invrc6 = -ic->dispersion_shift.cpot;
820 crc2 = gmx::square(ic->ewaldcoeff_lj*ic->rvdw);
821 ic->sh_lj_ewald = (std::exp(-crc2)*(1 + crc2 + 0.5*crc2*crc2) - 1)/gmx::power6(ic->rvdw);
825 /* We always re-initialize the tables whether they are used or not */
826 init_interaction_const_tables(nullptr, ic, rtab);
828 Nbnxm::gpu_pme_loadbal_update_param(nbv, ic, listParams);
830 if (!pme_lb->bSepPMERanks)
833 * CPU PME keeps a list of allocated pmedata's, that's why pme_lb->setup[pme_lb->cur].pmedata is not always nullptr.
834 * GPU PME, however, currently needs the gmx_pme_reinit always called on load balancing
835 * (pme_gpu_reinit might be not sufficiently decoupled from gmx_pme_init).
836 * This can lead to a lot of reallocations for PME GPU.
837 * Would be nicer if the allocated grid list was hidden within a single pmedata structure.
839 if ((pme_lb->setup[pme_lb->cur].pmedata == nullptr) || pme_gpu_task_enabled(pme_lb->setup[pme_lb->cur].pmedata))
841 /* Generate a new PME data structure,
842 * copying part of the old pointers.
844 gmx_pme_reinit(&set->pmedata,
845 cr, pme_lb->setup[0].pmedata, &ir,
846 set->grid, set->ewaldcoeff_q, set->ewaldcoeff_lj);
848 *pmedata = set->pmedata;
852 /* Tell our PME-only rank to switch grid */
853 gmx_pme_send_switchgrid(cr, set->grid, set->ewaldcoeff_q, set->ewaldcoeff_lj);
858 print_grid(nullptr, debug, "", "switched to", set, -1);
861 if (pme_lb->stage == pme_lb->nstage)
863 print_grid(fp_err, fp_log, "", "optimal", set, -1);
867 /*! \brief Prepare for another round of PME load balancing
869 * \param[in,out] pme_lb Pointer to PME load balancing struct
870 * \param[in] bDlbUnlocked TRUE is DLB was locked and is now unlocked
872 * If the conditions (e.g. DLB off/on, CPU/GPU throttling etc.) changed,
873 * the PP/PME balance might change and re-balancing can improve performance.
874 * This function adds 2 stages and adjusts the considered setup range.
876 static void continue_pme_loadbal(pme_load_balancing_t *pme_lb,
877 gmx_bool bDlbUnlocked)
879 /* Add 2 tuning stages, keep the detected end of the setup range */
881 if (bDlbUnlocked && pme_lb->bSepPMERanks)
883 /* With separate PME ranks, DLB should always lower the PP load and
884 * can only increase the PME load (more communication and imbalance),
885 * so we only need to scan longer cut-off's.
887 pme_lb->lower_limit = pme_lb->cur;
889 pme_lb->start = pme_lb->lower_limit;
892 void pme_loadbal_do(pme_load_balancing_t *pme_lb,
896 const gmx::MDLogger &mdlog,
897 const t_inputrec &ir,
899 const t_state &state,
900 gmx_wallcycle_t wcycle,
908 assert(pme_lb != nullptr);
910 if (!pme_lb->bActive)
915 n_prev = pme_lb->cycles_n;
916 cycles_prev = pme_lb->cycles_c;
917 wallcycle_get(wcycle, ewcSTEP, &pme_lb->cycles_n, &pme_lb->cycles_c);
919 /* Before the first step we haven't done any steps yet.
920 * Also handle cases where ir.init_step % ir.nstlist != 0.
922 if (pme_lb->cycles_n < ir.nstlist)
926 /* Sanity check, we expect nstlist cycle counts */
927 if (pme_lb->cycles_n - n_prev != ir.nstlist)
929 /* We could return here, but it's safer to issue an error and quit */
930 gmx_incons("pme_loadbal_do called at an interval != nstlist");
933 /* PME grid + cut-off optimization with GPUs or PME ranks */
934 if (!pme_lb->bBalance && pme_lb->bSepPMERanks)
936 if (pme_lb->bTriggerOnDLB)
938 pme_lb->bBalance = dd_dlb_is_on(cr->dd);
940 /* We should ignore the first timing to avoid timing allocation
941 * overhead. And since the PME load balancing is called just
942 * before DD repartitioning, the ratio returned by dd_pme_f_ratio
943 * is not over the last nstlist steps, but the nstlist steps before
944 * that. So the first useful ratio is available at step_rel=3*nstlist.
946 else if (step_rel >= 3*ir.nstlist)
948 if (DDMASTER(cr->dd))
950 /* If PME rank load is too high, start tuning */
952 (dd_pme_f_ratio(cr->dd) >= loadBalanceTriggerFactor);
954 dd_bcast(cr->dd, sizeof(gmx_bool), &pme_lb->bBalance);
957 pme_lb->bActive = (pme_lb->bBalance ||
958 step_rel <= pme_lb->step_rel_stop);
961 /* The location in the code of this balancing termination is strange.
962 * You would expect to have it after the call to pme_load_balance()
963 * below, since there pme_lb->stage is updated.
964 * But when terminating directly after deciding on and selecting the
965 * optimal setup, DLB will turn on right away if it was locked before.
966 * This might be due to PME reinitialization. So we check stage here
967 * to allow for another nstlist steps with DLB locked to stabilize
970 if (pme_lb->bBalance && pme_lb->stage == pme_lb->nstage)
972 pme_lb->bBalance = FALSE;
974 if (DOMAINDECOMP(cr) && dd_dlb_is_locked(cr->dd))
976 /* Unlock the DLB=auto, DLB is allowed to activate */
977 dd_dlb_unlock(cr->dd);
978 GMX_LOG(mdlog.warning).asParagraph().appendText("NOTE: DLB can now turn on, when beneficial");
980 /* We don't deactivate the tuning yet, since we will balance again
981 * after DLB gets turned on, if it does within PMETune_period.
983 continue_pme_loadbal(pme_lb, TRUE);
984 pme_lb->bTriggerOnDLB = TRUE;
985 pme_lb->step_rel_stop = step_rel + PMETunePeriod*ir.nstlist;
989 /* We're completely done with PME tuning */
990 pme_lb->bActive = FALSE;
993 if (DOMAINDECOMP(cr))
995 /* Set the cut-off limit to the final selected cut-off,
996 * so we don't have artificial DLB limits.
997 * This also ensures that we won't disable the currently
998 * optimal setting during a second round of PME balancing.
1000 set_dd_dlb_max_cutoff(cr, fr->nbv->listParams->rlistOuter);
1004 if (pme_lb->bBalance)
1006 /* We might not have collected nstlist steps in cycles yet,
1007 * since init_step might not be a multiple of nstlist,
1008 * but the first data collected is skipped anyhow.
1010 pme_load_balance(pme_lb, cr,
1011 fp_err, fp_log, mdlog,
1012 ir, state, pme_lb->cycles_c - cycles_prev,
1013 fr->ic, fr->nbv, &fr->pmedata,
1016 /* Update deprecated rlist in forcerec to stay in sync with fr->nbv */
1017 fr->rlist = fr->nbv->listParams->rlistOuter;
1019 if (ir.eDispCorr != edispcNO)
1021 calc_enervirdiff(nullptr, ir.eDispCorr, fr);
1025 if (!pme_lb->bBalance &&
1026 (!pme_lb->bSepPMERanks || step_rel > pme_lb->step_rel_stop))
1028 /* We have just deactivated the balancing and we're not measuring PP/PME
1029 * imbalance during the first steps of the run: deactivate the tuning.
1031 pme_lb->bActive = FALSE;
1034 if (!(pme_lb->bActive) && DOMAINDECOMP(cr) && dd_dlb_is_locked(cr->dd))
1036 /* Make sure DLB is allowed when we deactivate PME tuning */
1037 dd_dlb_unlock(cr->dd);
1038 GMX_LOG(mdlog.warning).asParagraph().appendText("NOTE: DLB can now turn on, when beneficial");
1041 *bPrinting = pme_lb->bBalance;
1044 /*! \brief Return product of the number of PME grid points in each dimension */
1045 static int pme_grid_points(const pme_setup_t *setup)
1047 return setup->grid[XX]*setup->grid[YY]*setup->grid[ZZ];
1050 /*! \brief Print one load-balancing setting */
1051 static void print_pme_loadbal_setting(FILE *fplog,
1053 const pme_setup_t *setup)
1056 " %-7s %6.3f nm %6.3f nm %3d %3d %3d %5.3f nm %5.3f nm\n",
1058 setup->rcut_coulomb, setup->rlistInner,
1059 setup->grid[XX], setup->grid[YY], setup->grid[ZZ],
1060 setup->spacing, 1/setup->ewaldcoeff_q);
1063 /*! \brief Print all load-balancing settings */
1064 static void print_pme_loadbal_settings(pme_load_balancing_t *pme_lb,
1066 const gmx::MDLogger &mdlog,
1067 gmx_bool bNonBondedOnGPU)
1069 double pp_ratio, grid_ratio;
1070 real pp_ratio_temporary;
1072 pp_ratio_temporary = pme_lb->setup[pme_lb->cur].rlistInner / pme_lb->setup[0].rlistInner;
1073 pp_ratio = gmx::power3(pp_ratio_temporary);
1074 grid_ratio = pme_grid_points(&pme_lb->setup[pme_lb->cur])/
1075 static_cast<double>(pme_grid_points(&pme_lb->setup[0]));
1077 fprintf(fplog, "\n");
1078 fprintf(fplog, " P P - P M E L O A D B A L A N C I N G\n");
1079 fprintf(fplog, "\n");
1080 /* Here we only warn when the optimal setting is the last one */
1081 if (pme_lb->elimited != epmelblimNO &&
1082 pme_lb->cur == pme_loadbal_end(pme_lb)-1)
1084 fprintf(fplog, " NOTE: The PP/PME load balancing was limited by the %s,\n",
1085 pmelblim_str[pme_lb->elimited]);
1086 fprintf(fplog, " you might not have reached a good load balance.\n");
1087 if (pme_lb->elimited == epmelblimDD)
1089 fprintf(fplog, " Try different mdrun -dd settings or lower the -dds value.\n");
1091 fprintf(fplog, "\n");
1093 fprintf(fplog, " PP/PME load balancing changed the cut-off and PME settings:\n");
1094 fprintf(fplog, " particle-particle PME\n");
1095 fprintf(fplog, " rcoulomb rlist grid spacing 1/beta\n");
1096 print_pme_loadbal_setting(fplog, "initial", &pme_lb->setup[0]);
1097 print_pme_loadbal_setting(fplog, "final", &pme_lb->setup[pme_lb->cur]);
1098 fprintf(fplog, " cost-ratio %4.2f %4.2f\n",
1099 pp_ratio, grid_ratio);
1100 fprintf(fplog, " (note that these numbers concern only part of the total PP and PME load)\n");
1102 if (pp_ratio > 1.5 && !bNonBondedOnGPU)
1104 GMX_LOG(mdlog.warning).asParagraph().appendText(
1105 "NOTE: PME load balancing increased the non-bonded workload by more than 50%.\n"
1106 " For better performance, use (more) PME ranks (mdrun -npme),\n"
1107 " or if you are beyond the scaling limit, use fewer total ranks (or nodes).");
1111 fprintf(fplog, "\n");
1115 void pme_loadbal_done(pme_load_balancing_t *pme_lb,
1117 const gmx::MDLogger &mdlog,
1118 gmx_bool bNonBondedOnGPU)
1120 if (fplog != nullptr && (pme_lb->cur > 0 || pme_lb->elimited != epmelblimNO))
1122 print_pme_loadbal_settings(pme_lb, fplog, mdlog, bNonBondedOnGPU);