fix memory allocation bug: WABBIT now is more certain to allocate exa…

…ctly what --memory tells it to

LIB/MODULE/module_globals.f90

@@ -32,7 +32,10 @@ module module_globals
   ! this parameter is a hack. in most parts of the code, a block has n_eqn component entries.
   ! universality dictates that we can also use a different number of components, for example
   ! when syn'ing the mask function (which in many cases has six entries.)
-  integer, public  :: N_MAX_COMPONENTS = 6
+  ! New in 06/2021: the hack continues. We now set this parameter at different places
+  ! to save on memory. That can be params%n_eqn (default in simulations), 6 (if mask is synced). The new default is 3,
+  ! for postprocessing.
+  integer, public  :: N_MAX_COMPONENTS = 3
 
   !subroutines of this module
   interface abort

LIB/MPI/module_mpi.f90

@@ -198,8 +198,9 @@ subroutine init_ghost_nodes( params )
             allocate( tmp_block( Bs(1)+2*g, Bs(2)+2*g, 1, Neqn) )
         end if
 
-        ! size of ghost nodes buffer. Note this contains only the ghost nodes layer
-        ! for all my blocks. previous versions allocated one of those per "friend"
+        ! size of ghost nodes buffer. The worst case is that we have to send ALL my ghost node
+        ! points to another rank, but this happens only if ALL my blocks have ONLY neighbors
+        ! on other MPIRANKS.
         if ( params%dim==3 ) then
           buffer_N = number_blocks * Neqn * ( (Bs(1)+2*g)*(Bs(2)+2*g)*(Bs(3)+2*g) - (Bs(1)*Bs(2)*Bs(3)) )
         else
@@ -213,20 +214,14 @@ subroutine init_ghost_nodes( params )
         ! allocate synch buffer
         if (rank==0) then
             write(*,'("GHOSTS-INIT: Attempting to allocate the ghost-sync-buffer.")')
-
-            write(*,'("GHOSTS-INIT: buffer_N_int=",i12," buffer_N=",i12," Nstages=",i1)') &
-            buffer_N_int, buffer_N, Nstages
-
-            write(*,'("GHOSTS-INIT: On each MPIRANK, Int  buffer:", f9.4, "GB")') &
-                2.0*dble(buffer_N_int)*dble(Nstages)*8e-9
-
-            write(*,'("GHOSTS-INIT: On each MPIRANK, Real buffer:", f9.4, "GB")') &
-                2.0*dble(buffer_N)*dble(Nstages)*8e-9
+            write(*,'("GHOSTS-INIT: buffer_N_int=",i12," buffer_N=",i12," Nstages=",i1)') buffer_N_int, buffer_N, Nstages
+            write(*,'("GHOSTS-INIT: Int  buffer:", f9.4, " GB per rank")') 2.0*dble(buffer_N_int)*dble(Nstages)*8e-9
+            write(*,'("GHOSTS-INIT: Real buffer:", f9.4, " GB per rank")') 2.0*dble(buffer_N)*dble(Nstages)*8e-9
             write(*,'("---------------- allocating now ----------------")')
         endif
 
         ! wait now so that if allocation fails, we get at least the above info
-        call MPI_barrier( WABBIT_COMM, status(1))
+        call MPI_barrier( WABBIT_COMM, status(1) )
 
         allocate( int_send_buffer( 1:buffer_N_int, 1:Nstages), stat=status(1) )
         allocate( int_recv_buffer( 1:buffer_N_int, 1:Nstages), stat=status(2) )
@@ -236,7 +231,6 @@ subroutine init_ghost_nodes( params )
         if (maxval(status) /= 0) call abort(999999, "Buffer allocation failed. Not enough memory?")
 
         if (rank==0) then
-
             write(*,'("GHOSTS-INIT: on each mpirank, Allocated ",A25," SHAPE=",7(i9,1x))') &
              "new_send_buffer", shape(new_send_buffer)
 

LIB/MPI/synchronize_ghosts_generic.f90

@@ -31,6 +31,7 @@ subroutine synchronize_ghosts_generic_sequence( params, lgt_block, hvy_block, hv
 
     integer(kind=ik) :: ijk(2,3)
     integer(kind=ik) :: bounds_type, istage, istage_buffer(1:4), rounds(1:4), inverse
+    logical, save :: informed = .false.
 
 
     if (.not. ghost_nodes_module_ready) then
@@ -42,8 +43,23 @@ subroutine synchronize_ghosts_generic_sequence( params, lgt_block, hvy_block, hv
     ! if this mpirank has no active blocks, it has nothing to do here.
     if (hvy_n == 0) return
 
-    if (size(hvy_block,4)>N_max_components) then
-        call abort(160720191,"You try to ghost-sync a vector with too many components.")
+    if (size(hvy_block,4)>N_MAX_COMPONENTS .and. .not. informed) then
+        if (params%rank ==0) then
+            write(*,*) "-------------------------------------------------------------------------"
+            write(*,*) "---warning---warning---warning---warning---warning---warning---warning---"
+            write(*,*) "---warning---warning---warning---warning---warning---warning---warning---"
+            write(*,*) "-------------------------------------------------------------------------"
+            write(*,*) " A warning from the ghost nodes module: we have allocated a buffer with an estimation for"
+            write(*,*) " neqn=", N_MAX_COMPONENTS, " components of a vector, but you try to sync"
+            write(*,*) " neqn=", size(hvy_block,4), " This may work just fine: but in some (rare) cases, "
+            write(*,*) " we will see a buffer overflow. The code will then abort with an error, and you have"
+            write(*,*) " to restart this simulation with more memory."
+            write(*,*) "-------------------------------------------------------------------------"
+            write(*,*) "---warning---warning---warning---warning---warning---warning---warning---"
+            write(*,*) "---warning---warning---warning---warning---warning---warning---warning---"
+            write(*,*) "-------------------------------------------------------------------------"
+        endif
+        informed = .true.
     endif
 
     Bs    = params%Bs
@@ -1053,6 +1069,9 @@ subroutine AppendLineToBuffer( int_send_buffer, new_send_buffer, buffer_size, ne
 
     ! real data
     if (buffer_size>0) then
+        if (i0+buffer_size-1 >= size(new_send_buffer,1)) then
+            call abort(202106049, "Internal bug: we ran out of space for the ghost nodes. Restart simulation with more memory.")
+        endif
         new_send_buffer( i0:i0+buffer_size-1, istage  ) = line_buffer(1:buffer_size)
     endif
 

LIB/PARAMS/ini_file_to_params.f90

@@ -104,6 +104,15 @@ subroutine ini_file_to_params( params, filename )
     call read_param_mpi(FILE, 'VPM', 'mask_time_independent_part', params%mask_time_independent_part, .true.)
     call read_param_mpi(FILE, 'VPM', 'dont_use_pruned_tree_mask', params%dont_use_pruned_tree_mask, .false.)
 
+    if (params%physics_type == "ACM-new")  then
+        if (params%penalization) then
+            if ((.not.params%dont_use_pruned_tree_mask).and.(params%mask_time_independent_part)) then
+                ! we sync the mask array in this case, which has 6 components
+                N_MAX_COMPONENTS = max(6, params%n_eqn)
+            endif
+        endif
+    endif
+
     ! decide if we use hartens point value multiresolution transform, which uses a coarsening operator
     ! that just takes every 2nd grid point or biorthogonal wavlets, which apply a smoothing filter (lowpass)
     ! prior to downsampling.
@@ -270,6 +279,10 @@ subroutine ini_blocks(params, FILE )
         call abort(170619,"Error: Max treelevel cannot be larger 18 (64bit long integer problem) ")
     end if
 
+    ! the default case is that we synchronize (ghosts) with n-eqn compontents in the vector
+    ! may be overwritten if pruned tree mask is used (by six)
+    N_MAX_COMPONENTS = params%n_eqn
+
     ! read switch to turn on|off mesh refinement
     call read_param_mpi(FILE, 'Blocks', 'adapt_mesh', params%adapt_mesh, .true. )
     call read_param_mpi(FILE, 'Blocks', 'adapt_inicond', params%adapt_inicond, params%adapt_mesh )

LIB/POSTPROCESSING/module_MOR.f90

@@ -460,6 +460,8 @@ subroutine post_POD(params)
     call get_cmd_arg( "--start_from_eigenbasis", eigenbasis_files)
     call get_cmd_arg( "--components", n_components, default=1_ik)
 
+    N_MAX_COMPONENTS = n_components ! used for ghost node sync'ing (buffer allocation)
+
     !-------------------------------
     ! Set some wabbit specific params
     !-------------------------------
@@ -751,6 +753,7 @@ subroutine post_PODerror(params)
     call get_cmd_arg( "--components", n_components, default=1_ik)
     call get_cmd_arg( "--iteration", iteration, default=1_ik)
 
+    N_MAX_COMPONENTS = n_components ! used for ghost node sync'ing (buffer allocation)
 
     if ( iteration>0 ) then
       if ( params%rank == 0 ) write(*,*) "Iteration reconstructed: " ,iteration
@@ -1320,6 +1323,7 @@ subroutine post_reconstruct(params)
     call get_cmd_arg( "--iteration", iteration, default=-1_ik)
     call get_cmd_arg( "--nmodes", N_modes_used, default=1_ik)
 
+    N_MAX_COMPONENTS = n_components ! used for ghost node sync'ing (buffer allocation)
 
     if ( iteration>0 ) then
       save_all = .False.
@@ -1729,6 +1733,8 @@ subroutine post_timecoef_POD(params)
     call get_cmd_arg( "--components", n_components, default=1_ik)
     call get_cmd_arg( "--iteration", iteration, default=1)
 
+    N_MAX_COMPONENTS = n_components ! used for ghost node sync'ing (buffer allocation)
+
     if ( iteration>0 ) then
       if ( params%rank == 0 ) write(*,*) "Iteration reconstructed: " ,iteration
     endif

LIB/POSTPROCESSING/post_generate_forest.f90

@@ -62,6 +62,8 @@ subroutine post_generate_forest(params)
     params%block_distribution = "sfc_hilbert"
     params%time_step_method = 'none'
 
+    N_MAX_COMPONENTS = params%n_eqn ! used for ghost node sync'ing (buffer allocation)
+
 
     ! we have to allocate grid if this routine is called for the first time
     call allocate_forest(params, lgt_block, hvy_block, hvy_neighbor, lgt_active, &

TESTING/runtests.sh

@@ -95,7 +95,10 @@ do
         echo "Writing output to: " ${logfile}
 
         # run the actual test
+        T2="$(date +%s)"
         ./${ts} > $logfile
+        T2="$(($(date +%s)-T2))"
+        echo "Time used in test: ${T2} seconds"
 
         if [ $? == 0 ]; then
             printf "%s \n" "${pass_color} pass ${end_color}"