diff --git a/src/Algebra.jl b/src/Algebra.jl
index 45244ec..e211bb8 100644
--- a/src/Algebra.jl
+++ b/src/Algebra.jl
@@ -507,17 +507,17 @@ end
 
 function Algebra.create_from_nz(a::DistributedAllocationCOO{<:SubAssembledRows})
   f(x) = nothing
-  A, = _sa_create_from_nz_with_callback(f,f,a)
+  A, = _sa_create_from_nz_with_callback(f,f,a,nothing)
   return A
 end
 
 function Algebra.create_from_nz(a::ArrayBlock{<:DistributedAllocationCOO{<:SubAssembledRows}})
   f(x) = nothing
-  A, = _sa_create_from_nz_with_callback(f,f,a)
+  A, = _sa_create_from_nz_with_callback(f,f,a,nothing)
   return A
 end
 
-function _sa_create_from_nz_with_callback(callback,async_callback,a)
+function _sa_create_from_nz_with_callback(callback,async_callback,a,b)
   # Recover some data
   I,J,V = get_allocations(a)
   test_dofs_gids_prange = get_test_gids(a)
@@ -538,7 +538,10 @@ function _sa_create_from_nz_with_callback(callback,async_callback,a)
   # Here we can overlap computations
   # This is a good place to overlap since
   # sending the matrix rows is a lot of data
-  b = callback(rows)
+  if !isa(b,Nothing)
+    bprange=_setup_prange_from_pvector_allocation(b)
+    b = callback(bprange)
+  end
 
   # Wait the transfer to finish
   wait(t)
@@ -701,10 +704,10 @@ end
 
 function Algebra.create_from_nz(
   a::DistributedAllocationCOO{<:SubAssembledRows},
-  c_fespace::PVectorAllocationTrackOnlyValues{<:SubAssembledRows})
+  b::PVectorAllocationTrackTouchedAndValues)
 
   function callback(rows)
-    _rhs_callback(c_fespace,rows)
+    _rhs_callback(b,rows)
   end
 
   function async_callback(b)
@@ -712,7 +715,7 @@ function Algebra.create_from_nz(
     assemble!(b)
   end
 
-  A,b = _sa_create_from_nz_with_callback(callback,async_callback,a)
+  A,b = _sa_create_from_nz_with_callback(callback,async_callback,a,b)
   return A,b
 end
 
@@ -759,24 +762,31 @@ end
   nothing
 end
 
+
+function _setup_touched_and_allocations_arrays(values)
+  touched = map(values) do values
+    fill!(Vector{Bool}(undef,length(values)),false)
+  end
+  allocations = map(values,touched) do values,touched
+   ArrayAllocationTrackTouchedAndValues(touched,values)
+  end
+  touched, allocations
+end
+
 function Arrays.nz_allocation(a::DistributedCounterCOO{<:SubAssembledRows},
                               b::PVectorCounter{<:SubAssembledRows})
   A      = nz_allocation(a)
   dofs   = b.test_dofs_gids_prange
   values = map(nz_allocation,b.counters)
-  B = PVectorAllocationTrackOnlyValues(b.par_strategy,values,dofs)
+  touched,allocations=_setup_touched_and_allocations_arrays(values)
+  B = PVectorAllocationTrackTouchedAndValues(allocations,values,dofs)
   return A,B
 end
 
 function Arrays.nz_allocation(a::PVectorCounter{<:SubAssembledRows})
   dofs = a.test_dofs_gids_prange
   values = map(nz_allocation,a.counters)
-  touched = map(values) do values
-     fill!(Vector{Bool}(undef,length(values)),false)
-  end
-  allocations = map(values,touched) do values,touched
-    ArrayAllocationTrackTouchedAndValues(touched,values)
-  end
+  touched,allocations=_setup_touched_and_allocations_arrays(values)
   return PVectorAllocationTrackTouchedAndValues(allocations,values,dofs)
 end
 
@@ -784,7 +794,7 @@ function local_views(a::PVectorAllocationTrackTouchedAndValues)
   a.allocations
 end
 
-function Algebra.create_from_nz(a::PVectorAllocationTrackTouchedAndValues)
+function _setup_prange_from_pvector_allocation(a::PVectorAllocationTrackTouchedAndValues)
   test_dofs_prange = a.test_dofs_gids_prange # dof ids of the test space
   ngrdofs = length(test_dofs_prange)
   
@@ -810,10 +820,13 @@ function Algebra.create_from_nz(a::PVectorAllocationTrackTouchedAndValues)
   gids_ghost_to_global, gids_ghost_to_owner = map(
     find_gid_and_owner,I_ghost_lids_to_dofs_ghost_lids,indices) |> tuple_of_arrays
 
-  rows = _setup_prange_impl_(ngrdofs,indices,gids_ghost_to_global,gids_ghost_to_owner)
+  _setup_prange_impl_(ngrdofs,indices,gids_ghost_to_global,gids_ghost_to_owner)
+end
+
+function Algebra.create_from_nz(a::PVectorAllocationTrackTouchedAndValues)
+  rows = _setup_prange_from_pvector_allocation(a)
   b    = _rhs_callback(a,rows)
   t2   = assemble!(b)
-
   # Wait the transfer to finish
   if t2 !== nothing
     wait(t2)
@@ -821,9 +834,7 @@ function Algebra.create_from_nz(a::PVectorAllocationTrackTouchedAndValues)
   return b
 end
 
-
 # Common Assembly Utilities
-
 function first_gdof_from_ids(ids)
   if own_length(ids) == 0
     return 1
diff --git a/src/FESpaces.jl b/src/FESpaces.jl
index cd922dd..8972ff4 100644
--- a/src/FESpaces.jl
+++ b/src/FESpaces.jl
@@ -621,7 +621,10 @@ end
 function FESpaces.symbolic_loop_matrix_and_vector!(A,b,a::DistributedSparseMatrixAssembler,data)
   rows = get_rows(a)
   cols = get_cols(a)
-  map(symbolic_loop_matrix_and_vector!,local_views(A,rows,cols),local_views(b,rows),local_views(a),data)
+  Aviews=local_views(A,rows,cols)
+  bviews=local_views(b,rows)
+  aviews=local_views(a)
+  map(symbolic_loop_matrix_and_vector!,Aviews,bviews,aviews,data)  
 end
 
 function FESpaces.numeric_loop_matrix_and_vector!(A,b,a::DistributedSparseMatrixAssembler,data)