64-BitBrainstorm_2026/AMSS-NCKU
8
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Optimize MPI Sync with merged transfers, caching, and async overlap

Browse Source 
Phase 1: Merge N+1 transfer() calls into a single transfer() per
Sync(PatchList), reducing N+1 MPI_Waitall barriers to 1 via new
Sync_merged() that collects all intra-patch and inter-patch grid
segment lists into combined per-rank arrays.

Phase 2: Cache grid segment lists and reuse grow-only communication
buffers across RK4 substeps via SyncCache struct. Caches are per-level
and per-variable-list (predictor/corrector), invalidated on regrid.
Eliminates redundant build_ghost_gsl/build_owned_gsl0/build_gstl
rebuilds and malloc/free cycles between regrids.

Phase 3: Split Sync into async Sync_start/Sync_finish to overlap
Cartesian ghost zone exchange (MPI_Isend/Irecv) with Shell patch
synchronization. Uses MPI tag 2 to avoid conflicts with SH->Synch()
which uses transfer() with tag 1.

Also updates makefile.inc paths and flags for local build environment.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
CGH0S7
2026-02-09 21:03:37 +08:00
parent e9d321fd00
commit 42b9cf1ad9
4 changed files with 564 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

53
AMSS_NCKU_source/bssn_class.C
View File

@@ -730,6 +730,10 @@ void bssn_class::Initialize()
PhysTime = StartTime;
Setup_Black_Hole_position();
}
// Initialize sync caches (per-level, for predictor and corrector)
sync_cache_pre = new Parallel::SyncCache[GH->levels];
sync_cache_cor = new Parallel::SyncCache[GH->levels];
}
//================================================================================================
@@ -981,6 +985,20 @@ bssn_class::~bssn_class()
delete Azzz;
#endif
// Destroy sync caches before GH
if (sync_cache_pre)
{
for (int i = 0; i < GH->levels; i++)
sync_cache_pre[i].destroy();
delete[] sync_cache_pre;
}
if (sync_cache_cor)
{
for (int i = 0; i < GH->levels; i++)
sync_cache_cor[i].destroy();
delete[] sync_cache_cor;
}
delete GH;
#ifdef WithShell
delete SH;
@@ -2181,6 +2199,7 @@ void bssn_class::Evolve(int Steps)
GH->Regrid(Symmetry, BH_num, Porgbr, Porg0,
SynchList_cor, OldStateList, StateList, SynchList_pre,
fgt(PhysTime - dT_mon, StartTime, dT_mon / 2), ErrorMonitor);
for (int il = 0; il < GH->levels; il++) { sync_cache_pre[il].invalidate(); sync_cache_cor[il].invalidate(); }
#endif
#if (REGLEV == 0 && (PSTR == 1 || PSTR == 2))
@@ -2396,6 +2415,7 @@ void bssn_class::RecursiveStep(int lev)
GH->Regrid_Onelevel(lev, Symmetry, BH_num, Porgbr, Porg0,
SynchList_cor, OldStateList, StateList, SynchList_pre,
fgt(PhysTime - dT_lev, StartTime, dT_lev / 2), ErrorMonitor);
for (int il = 0; il < GH->levels; il++) { sync_cache_pre[il].invalidate(); sync_cache_cor[il].invalidate(); }
#endif
}
@@ -2574,6 +2594,7 @@ void bssn_class::ParallelStep()
GH->Regrid_Onelevel(GH->mylev, Symmetry, BH_num, Porgbr, Porg0,
SynchList_cor, OldStateList, StateList, SynchList_pre,
fgt(PhysTime - dT_lev, StartTime, dT_lev / 2), ErrorMonitor);
for (int il = 0; il < GH->levels; il++) { sync_cache_pre[il].invalidate(); sync_cache_cor[il].invalidate(); }
#endif
}
@@ -2740,6 +2761,7 @@ void bssn_class::ParallelStep()
GH->Regrid_Onelevel(lev + 1, Symmetry, BH_num, Porgbr, Porg0,
SynchList_cor, OldStateList, StateList, SynchList_pre,
fgt(PhysTime - dT_levp1, StartTime, dT_levp1 / 2), ErrorMonitor);
for (int il = 0; il < GH->levels; il++) { sync_cache_pre[il].invalidate(); sync_cache_cor[il].invalidate(); }
// a_stream.clear();
// a_stream.str("");
@@ -2754,6 +2776,7 @@ void bssn_class::ParallelStep()
GH->Regrid_Onelevel(lev, Symmetry, BH_num, Porgbr, Porg0,
SynchList_cor, OldStateList, StateList, SynchList_pre,
fgt(PhysTime - dT_lev, StartTime, dT_lev / 2), ErrorMonitor);
for (int il = 0; il < GH->levels; il++) { sync_cache_pre[il].invalidate(); sync_cache_cor[il].invalidate(); }
// a_stream.clear();
// a_stream.str("");
@@ -2772,6 +2795,7 @@ void bssn_class::ParallelStep()
GH->Regrid_Onelevel(lev - 1, Symmetry, BH_num, Porgbr, Porg0,
SynchList_cor, OldStateList, StateList, SynchList_pre,
fgt(PhysTime - dT_lev, StartTime, dT_levm1 / 2), ErrorMonitor);
for (int il = 0; il < GH->levels; il++) { sync_cache_pre[il].invalidate(); sync_cache_cor[il].invalidate(); }
// a_stream.clear();
// a_stream.str("");
@@ -2787,6 +2811,7 @@ void bssn_class::ParallelStep()
GH->Regrid_Onelevel(lev - 1, Symmetry, BH_num, Porgbr, Porg0,
SynchList_cor, OldStateList, StateList, SynchList_pre,
fgt(PhysTime - dT_lev, StartTime, dT_levm1 / 2), ErrorMonitor);
for (int il = 0; il < GH->levels; il++) { sync_cache_pre[il].invalidate(); sync_cache_cor[il].invalidate(); }
// a_stream.clear();
// a_stream.str("");
@@ -3310,7 +3335,8 @@ void bssn_class::Step(int lev, int YN)
}
#endif
Parallel::Sync(GH->PatL[lev], SynchList_pre, Symmetry);
Parallel::AsyncSyncState async_pre;
Parallel::Sync_start(GH->PatL[lev], SynchList_pre, Symmetry, sync_cache_pre[lev], async_pre);
#ifdef WithShell
if (lev == 0)
@@ -3328,7 +3354,10 @@ void bssn_class::Step(int lev, int YN)
<< " seconds! " << endl;
}
}
#endif
Parallel::Sync_finish(sync_cache_pre[lev], async_pre, SynchList_pre, Symmetry);
#ifdef WithShell
// Complete non-blocking error reduction and check
MPI_Wait(&err_req, MPI_STATUS_IGNORE);
if (ERROR)
@@ -3666,7 +3695,8 @@ void bssn_class::Step(int lev, int YN)
}
#endif
Parallel::Sync(GH->PatL[lev], SynchList_cor, Symmetry);
Parallel::AsyncSyncState async_cor;
Parallel::Sync_start(GH->PatL[lev], SynchList_cor, Symmetry, sync_cache_cor[lev], async_cor);
#ifdef WithShell
if (lev == 0)
@@ -3684,7 +3714,10 @@ void bssn_class::Step(int lev, int YN)
<< " seconds! " << endl;
}
}
#endif
Parallel::Sync_finish(sync_cache_cor[lev], async_cor, SynchList_cor, Symmetry);
#ifdef WithShell
// Complete non-blocking error reduction and check
MPI_Wait(&err_req_cor, MPI_STATUS_IGNORE);
if (ERROR)
@@ -4161,7 +4194,8 @@ void bssn_class::Step(int lev, int YN)
}
#endif
Parallel::Sync(GH->PatL[lev], SynchList_pre, Symmetry);
Parallel::AsyncSyncState async_pre;
Parallel::Sync_start(GH->PatL[lev], SynchList_pre, Symmetry, sync_cache_pre[lev], async_pre);
#ifdef WithShell
if (lev == 0)
@@ -4179,7 +4213,10 @@ void bssn_class::Step(int lev, int YN)
<< " seconds! " << endl;
}
}
#endif
Parallel::Sync_finish(sync_cache_pre[lev], async_pre, SynchList_pre, Symmetry);
#ifdef WithShell
// Complete non-blocking error reduction and check
MPI_Wait(&err_req, MPI_STATUS_IGNORE);
if (ERROR)
@@ -4502,7 +4539,8 @@ void bssn_class::Step(int lev, int YN)
}
#endif
Parallel::Sync(GH->PatL[lev], SynchList_cor, Symmetry);
Parallel::AsyncSyncState async_cor;
Parallel::Sync_start(GH->PatL[lev], SynchList_cor, Symmetry, sync_cache_cor[lev], async_cor);
#ifdef WithShell
if (lev == 0)
@@ -4520,7 +4558,10 @@ void bssn_class::Step(int lev, int YN)
<< " seconds! " << endl;
}
}
#endif
Parallel::Sync_finish(sync_cache_cor[lev], async_cor, SynchList_cor, Symmetry);
#ifdef WithShell
// Complete non-blocking error reduction and check
MPI_Wait(&err_req_cor, MPI_STATUS_IGNORE);
if (ERROR)
@@ -4910,7 +4951,7 @@ void bssn_class::Step(int lev, int YN)
// misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"before Predictor sync");
Parallel::Sync(GH->PatL[lev], SynchList_pre, Symmetry);
Parallel::Sync_cached(GH->PatL[lev], SynchList_pre, Symmetry, sync_cache_pre[lev]);
// Complete non-blocking error reduction and check
MPI_Wait(&err_req, MPI_STATUS_IGNORE);
@@ -5111,7 +5152,7 @@ void bssn_class::Step(int lev, int YN)
// misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"before Corrector sync");
Parallel::Sync(GH->PatL[lev], SynchList_cor, Symmetry);
Parallel::Sync_cached(GH->PatL[lev], SynchList_cor, Symmetry, sync_cache_cor[lev]);
// misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"after Corrector sync");