#include <algorithm>
-#include "gromacs/bonded/bonded.h"
#include "gromacs/fileio/gmxfio.h"
#include "gromacs/fileio/pdbio.h"
#include "gromacs/imd/imd.h"
#include "gromacs/legacyheaders/pme.h"
#include "gromacs/legacyheaders/shellfc.h"
#include "gromacs/legacyheaders/typedefs.h"
+#include "gromacs/listed-forces/bonded.h"
#include "gromacs/math/vec.h"
#include "gromacs/mdlib/nb_verlet.h"
#include "gromacs/mdlib/nbnxn_search.h"
/* The DLB option */
int eDLB;
+ /* Is eDLB=edlbAUTO locked such that we currently can't turn it on? */
+ gmx_bool bDLB_locked;
/* Are we actually using DLB? */
gmx_bool bDynLoadBal;
int eFlop;
double flop;
int flop_n;
- /* Have often have did we have load measurements */
+ /* How many times have did we have load measurements */
int n_load_have;
- /* Have often have we collected the load measurements */
+ /* How many times have we collected the load measurements */
int n_load_collect;
/* Statistics */
cell_size[i] = 1.0/ncd;
}
}
- else if (dd_load_count(comm))
+ else if (dd_load_count(comm) > 0)
{
load_aver = comm->load[d].sum_m/ncd;
change_max = 0;
static void print_cg_move(FILE *fplog,
gmx_domdec_t *dd,
gmx_int64_t step, int cg, int dim, int dir,
- gmx_bool bHaveLimitdAndCMOld, real limitd,
+ gmx_bool bHaveCgcmOld, real limitd,
rvec cm_old, rvec cm_new, real pos_d)
{
gmx_domdec_comm_t *comm;
comm = dd->comm;
fprintf(fplog, "\nStep %s:\n", gmx_step_str(step, buf));
- if (bHaveLimitdAndCMOld)
+ if (limitd > 0)
{
- fprintf(fplog, "The charge group starting at atom %d moved more than the distance allowed by the domain decomposition (%f) in direction %c\n",
+ fprintf(fplog, "%s %d moved more than the distance allowed by the domain decomposition (%f) in direction %c\n",
+ dd->comm->bCGs ? "The charge group starting at atom" : "Atom",
ddglatnr(dd, dd->cgindex[cg]), limitd, dim2char(dim));
}
else
{
- fprintf(fplog, "The charge group starting at atom %d moved than the distance allowed by the domain decomposition in direction %c\n",
+ /* We don't have a limiting distance available: don't print it */
+ fprintf(fplog, "%s %d moved more than the distance allowed by the domain decomposition in direction %c\n",
+ dd->comm->bCGs ? "The charge group starting at atom" : "Atom",
ddglatnr(dd, dd->cgindex[cg]), dim2char(dim));
}
fprintf(fplog, "distance out of cell %f\n",
dir == 1 ? pos_d - comm->cell_x1[dim] : pos_d - comm->cell_x0[dim]);
- if (bHaveLimitdAndCMOld)
+ if (bHaveCgcmOld)
{
fprintf(fplog, "Old coordinates: %8.3f %8.3f %8.3f\n",
cm_old[XX], cm_old[YY], cm_old[ZZ]);
static void cg_move_error(FILE *fplog,
gmx_domdec_t *dd,
gmx_int64_t step, int cg, int dim, int dir,
- gmx_bool bHaveLimitdAndCMOld, real limitd,
+ gmx_bool bHaveCgcmOld, real limitd,
rvec cm_old, rvec cm_new, real pos_d)
{
if (fplog)
{
print_cg_move(fplog, dd, step, cg, dim, dir,
- bHaveLimitdAndCMOld, limitd, cm_old, cm_new, pos_d);
+ bHaveCgcmOld, limitd, cm_old, cm_new, pos_d);
}
print_cg_move(stderr, dd, step, cg, dim, dir,
- bHaveLimitdAndCMOld, limitd, cm_old, cm_new, pos_d);
+ bHaveCgcmOld, limitd, cm_old, cm_new, pos_d);
gmx_fatal(FARGS,
- "A charge group moved too far between two domain decomposition steps\n"
- "This usually means that your system is not well equilibrated");
+ "%s moved too far between two domain decomposition steps\n"
+ "This usually means that your system is not well equilibrated",
+ dd->comm->bCGs ? "A charge group" : "An atom");
}
static void rotate_state_atom(t_state *state, int a)
{
if (pos_d >= limit1[d])
{
- cg_move_error(fplog, dd, step, cg, d, 1, TRUE, limitd[d],
+ cg_move_error(fplog, dd, step, cg, d, 1,
+ cg_cm != state->x, limitd[d],
cg_cm[cg], cm_new, pos_d);
}
dev[d] = 1;
{
if (pos_d < limit0[d])
{
- cg_move_error(fplog, dd, step, cg, d, -1, TRUE, limitd[d],
+ cg_move_error(fplog, dd, step, cg, d, -1,
+ cg_cm != state->x, limitd[d],
cg_cm[cg], cm_new, pos_d);
}
dev[d] = -1;
{
cg_move_error(fplog, dd, step, cg, dim,
(flag & DD_FLAG_FW(d)) ? 1 : 0,
- FALSE, 0,
+ fr->cutoff_scheme == ecutsGROUP, 0,
comm->vbuf.v[buf_pos],
comm->vbuf.v[buf_pos],
comm->vbuf.v[buf_pos][dim]);
/* Initialize to GPU share count to 0, might change later */
comm->nrank_gpu_shared = 0;
- comm->eDLB = check_dlb_support(fplog, cr, dlb_opt, comm->bRecordLoad, Flags, ir);
+ comm->eDLB = check_dlb_support(fplog, cr, dlb_opt, comm->bRecordLoad, Flags, ir);
+ comm->bDLB_locked = FALSE;
comm->bDynLoadBal = (comm->eDLB == edlbYES);
if (fplog)
comm->cellsize_limit = 0;
comm->bBondComm = FALSE;
+ /* Atoms should be able to move by up to half the list buffer size (if > 0)
+ * within nstlist steps. Since boundaries are allowed to displace by half
+ * a cell size, DD cells should be at least the size of the list buffer.
+ */
+ comm->cellsize_limit = std::max(comm->cellsize_limit,
+ ir->rlistlong - std::max(ir->rvdw, ir->rcoulomb));
+
if (comm->bInterCGBondeds)
{
if (comm_distance_min > 0)
comm->PMELoadBal_max_cutoff = comm->cutoff;
}
+gmx_bool dd_dlb_is_locked(const gmx_domdec_t *dd)
+{
+ return dd->comm->bDLB_locked;
+}
+
+void dd_dlb_set_lock(gmx_domdec_t *dd, gmx_bool bValue)
+{
+ /* We can only lock the DLB when it is set to auto, otherwise don't lock */
+ if (dd->comm->eDLB == edlbAUTO)
+ {
+ dd->comm->bDLB_locked = bValue;
+ }
+}
+
static void merge_cg_buffers(int ncell,
gmx_domdec_comm_dim_t *cd, int pulse,
int *ncg_cell,
}
/* Check if we have recorded loads on the nodes */
- if (comm->bRecordLoad && dd_load_count(comm))
+ if (comm->bRecordLoad && dd_load_count(comm) > 0)
{
- if (comm->eDLB == edlbAUTO && !comm->bDynLoadBal)
+ if (comm->eDLB == edlbAUTO && !comm->bDynLoadBal && !dd_dlb_is_locked(dd))
{
/* Check if we should use DLB at the second partitioning
* and every 100 partitionings,
* so the extra communication cost is negligible.
*/
- n = std::max(100, nstglobalcomm);
+ const int nddp_chk_dlb = 100;
bCheckDLB = (comm->n_load_collect == 0 ||
- comm->n_load_have % n == n-1);
+ comm->n_load_have % nddp_chk_dlb == nddp_chk_dlb - 1);
}
else
{
/* Since the timings are node dependent, the master decides */
if (DDMASTER(dd))
{
- bTurnOnDLB =
- (dd_force_imb_perf_loss(dd) >= DD_PERF_LOSS_DLB_ON);
+ /* Here we check if the max PME rank load is more than 0.98
+ * the max PP force load. If so, PP DLB will not help,
+ * since we are (almost) limited by PME. Furthermore,
+ * DLB will cause a significant extra x/f redistribution
+ * cost on the PME ranks, which will then surely result
+ * in lower total performance.
+ * This check might be fragile, since one measurement
+ * below 0.98 (although only done once every 100 DD part.)
+ * could turn on DLB for the rest of the run.
+ */
+ if (cr->npmenodes > 0 &&
+ dd_pme_f_ratio(dd) > 1 - DD_PERF_LOSS_DLB_ON)
+ {
+ bTurnOnDLB = FALSE;
+ }
+ else
+ {
+ bTurnOnDLB =
+ (dd_force_imb_perf_loss(dd) >= DD_PERF_LOSS_DLB_ON);
+ }
if (debug)
{
fprintf(debug, "step %s, imb loss %f\n",