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

Cache GSL in SyncPlan and apply async Sync to Z4c_class

Browse Source 
Major optimization: Pre-build grid segment lists (GSLs) once per Step() call
via SyncPreparePlan(), then reuse them across all 4 RK4 substep SyncBegin calls
via SyncBeginWithPlan(). This eliminates the O(cpusize * blocks^2) GSL rebuild
cost that was incurred on every ghost zone exchange.

Applied async SyncBegin/SyncEnd overlap pattern to Z4c_class.C (ABEtype==2,
the default configuration), which was still using blocking Parallel::Sync.
Both the regular and CPBC variants of Z4c Step() are now optimized.

Co-authored-by: copilot-swe-agent[bot] <198982749+copilot@users.noreply.github.com>
This commit is contained in:
ianchb
2026-02-08 08:36:21 +00:00
parent a918dc103e
commit afd4006da2
4 changed files with 225 additions and 25 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
AMSS_NCKU_source/bssn_class.C
View File

@@ -3035,6 +3035,12 @@ void bssn_class::Step(int lev, int YN)
int ERROR = 0;
MyList<ss_patch> *sPp;
// Pre-build grid segment lists once for this level's patches.
// These are reused across predictor + 3 corrector SyncBegin calls,
// avoiding O(cpusize * blocks^2) rebuild each time.
Parallel::SyncPlan *sync_plan = Parallel::SyncPreparePlan(GH->PatL[lev], Symmetry);
// Predictor
MyList<Patch> *Pp = GH->PatL[lev];
while (Pp)
@@ -3160,7 +3166,7 @@ void bssn_class::Step(int lev, int YN)
}
// Start async ghost zone exchange - overlaps with error check and Shell computation
Parallel::SyncHandle *sync_pre = Parallel::SyncBegin(GH->PatL[lev], SynchList_pre, Symmetry);
Parallel::SyncHandle *sync_pre = Parallel::SyncBeginWithPlan(sync_plan, SynchList_pre);
// check error information (overlaps with MPI transfer)
{
@@ -3536,7 +3542,7 @@ void bssn_class::Step(int lev, int YN)
}
// Start async ghost zone exchange - overlaps with error check and Shell computation
Parallel::SyncHandle *sync_cor = Parallel::SyncBegin(GH->PatL[lev], SynchList_cor, Symmetry);
Parallel::SyncHandle *sync_cor = Parallel::SyncBeginWithPlan(sync_plan, SynchList_cor);
// check error information (overlaps with MPI transfer)
{
@@ -3908,6 +3914,8 @@ void bssn_class::Step(int lev, int YN)
Porg0[ithBH][2] = Porg1[ithBH][2];
}
}
Parallel::SyncFreePlan(sync_plan);
}
//================================================================================================
@@ -4830,6 +4838,12 @@ void bssn_class::Step(int lev, int YN)
int ERROR = 0;
MyList<ss_patch> *sPp;
// Pre-build grid segment lists once for this level's patches.
// These are reused across predictor + 3 corrector SyncBegin calls,
// avoiding O(cpusize * blocks^2) rebuild each time.
Parallel::SyncPlan *sync_plan = Parallel::SyncPreparePlan(GH->PatL[lev], Symmetry);
// Predictor
MyList<Patch> *Pp = GH->PatL[lev];
while (Pp)
@@ -4957,7 +4971,7 @@ void bssn_class::Step(int lev, int YN)
// misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"after Predictor rhs calculation");
// Start async ghost zone exchange - overlaps with error check and BH position
Parallel::SyncHandle *sync_pre = Parallel::SyncBegin(GH->PatL[lev], SynchList_pre, Symmetry);
Parallel::SyncHandle *sync_pre = Parallel::SyncBeginWithPlan(sync_plan, SynchList_pre);
// check error information (overlaps with MPI transfer)
{
@@ -5159,7 +5173,7 @@ void bssn_class::Step(int lev, int YN)
// misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"before Corrector error check");
// Start async ghost zone exchange - overlaps with error check and BH position
Parallel::SyncHandle *sync_cor = Parallel::SyncBegin(GH->PatL[lev], SynchList_cor, Symmetry);
Parallel::SyncHandle *sync_cor = Parallel::SyncBeginWithPlan(sync_plan, SynchList_cor);
// check error information (overlaps with MPI transfer)
{
@@ -5299,6 +5313,8 @@ void bssn_class::Step(int lev, int YN)
// if(myrank==GH->start_rank[lev]) cout<<GH->mylev<<endl;
// misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"complet GH Step");
Parallel::SyncFreePlan(sync_plan);
}
//================================================================================================