diff --git a/Exec/Make.Castro b/Exec/Make.Castro
index 5b1575820b..106ef51cc0 100644
--- a/Exec/Make.Castro
+++ b/Exec/Make.Castro
@@ -55,6 +55,9 @@ endif
 # Require C++17
 CXXSTD := c++17
 
+# Use Lazy QueueReduction for the timing outputs
+LAZY := TRUE
+
 # default integrator
 INTEGRATOR_DIR ?= VODE
 
diff --git a/Source/diffusion/Castro_diffusion.cpp b/Source/diffusion/Castro_diffusion.cpp
index c02fc3f611..0a2a98fbce 100644
--- a/Source/diffusion/Castro_diffusion.cpp
+++ b/Source/diffusion/Castro_diffusion.cpp
@@ -22,15 +22,16 @@ Castro::construct_old_diff_source(MultiFab& source, MultiFab& state_in, Real tim
 
     if (verbose > 1)
     {
-        const int IOProc   = ParallelDescriptor::IOProcessorNumber();
-        Real      run_time = ParallelDescriptor::second() - strt_time;
-
+        const int IOProc = ParallelDescriptor::IOProcessorNumber();
+        amrex::Real run_time = ParallelDescriptor::second() - strt_time;
+        amrex::Real llevel = level;
 #ifdef BL_LAZY
         Lazy::QueueReduction( [=] () mutable {
 #endif
         ParallelDescriptor::ReduceRealMax(run_time,IOProc);
 
-        amrex::Print() << "Castro::construct_old_diff_source() time = " << run_time << " on level " << level << "\n" << "\n";
+        amrex::Print() << "Castro::construct_old_diff_source() time = " << run_time
+                       << " on level " << llevel << "\n" << "\n";
 #ifdef BL_LAZY
         });
 #endif
@@ -59,15 +60,17 @@ Castro::construct_new_diff_source(MultiFab& source, MultiFab& state_old, MultiFa
 
     if (verbose > 1)
     {
-        const int IOProc   = ParallelDescriptor::IOProcessorNumber();
-        Real      run_time = ParallelDescriptor::second() - strt_time;
+        const int IOProc = ParallelDescriptor::IOProcessorNumber();
+        amrex::Real run_time = ParallelDescriptor::second() - strt_time;
+        amrex::Real llevel = level;
 
 #ifdef BL_LAZY
         Lazy::QueueReduction( [=] () mutable {
 #endif
         ParallelDescriptor::ReduceRealMax(run_time,IOProc);
 
-        amrex::Print() << "Castro::construct_new_diff_source() time = " << run_time << " on level " << level << "\n" << "\n";
+        amrex::Print() << "Castro::construct_new_diff_source() time = " << run_time
+                       << " on level " << llevel << "\n" << "\n";
 #ifdef BL_LAZY
         });
 #endif
diff --git a/Source/driver/Castro.cpp b/Source/driver/Castro.cpp
index 3108b7794f..283cdf1e11 100644
--- a/Source/driver/Castro.cpp
+++ b/Source/driver/Castro.cpp
@@ -3048,14 +3048,16 @@ Castro::reflux (int crse_level, int fine_level, bool in_post_timestep)
     if (verbose)
     {
         const int IOProc = ParallelDescriptor::IOProcessorNumber();
-        Real      end    = ParallelDescriptor::second() - strt;
+        amrex::Real end = ParallelDescriptor::second() - strt;
+        amrex::Real llevel = level;
 
 #ifdef BL_LAZY
         Lazy::QueueReduction( [=] () mutable {
 #endif
         ParallelDescriptor::ReduceRealMax(end,IOProc);
         if (ParallelDescriptor::IOProcessor()) {
-          std::cout << "Castro::reflux() at level " << level << " : time = " << end << std::endl;
+            std::cout << "Castro::reflux() at level " << llevel
+                      << " : time = " << end << std::endl;
         }
 #ifdef BL_LAZY
         });
diff --git a/Source/gravity/Castro_gravity.cpp b/Source/gravity/Castro_gravity.cpp
index bf9dc4f6b7..395e0e675e 100644
--- a/Source/gravity/Castro_gravity.cpp
+++ b/Source/gravity/Castro_gravity.cpp
@@ -122,15 +122,17 @@ Castro::construct_old_gravity (Real time)
 
     if (verbose > 0)
     {
-        const int IOProc   = ParallelDescriptor::IOProcessorNumber();
-        Real      run_time = ParallelDescriptor::second() - strt_time;
+        const int IOProc = ParallelDescriptor::IOProcessorNumber();
+        amrex::Real run_time = ParallelDescriptor::second() - strt_time;
+        amrex::Real llevel = level;
 
 #ifdef BL_LAZY
         Lazy::QueueReduction( [=] () mutable {
 #endif
         ParallelDescriptor::ReduceRealMax(run_time,IOProc);
 
-        amrex::Print() << "Castro::construct_old_gravity() time = " << run_time << " on level " << level << "\n" << "\n";
+        amrex::Print() << "Castro::construct_old_gravity() time = " << run_time << " on level "
+                       << llevel << "\n" << "\n";
 #ifdef BL_LAZY
         });
 #endif
@@ -278,15 +280,17 @@ Castro::construct_new_gravity (Real time)
 
     if (verbose > 0)
     {
-        const int IOProc   = ParallelDescriptor::IOProcessorNumber();
-        Real      run_time = ParallelDescriptor::second() - strt_time;
+        const int IOProc = ParallelDescriptor::IOProcessorNumber();
+        amrex::Real run_time = ParallelDescriptor::second() - strt_time;
+        amrex::Real llevel = level;
 
 #ifdef BL_LAZY
         Lazy::QueueReduction( [=] () mutable {
 #endif
         ParallelDescriptor::ReduceRealMax(run_time,IOProc);
 
-        amrex::Print() << "Castro::construct_new_gravity() time = " << run_time << " on level " << level << "\n" << "\n";
+        amrex::Print() << "Castro::construct_new_gravity() time = " << run_time
+                       << " on level " << llevel << "\n" << "\n";
 #ifdef BL_LAZY
         });
 #endif
@@ -431,14 +435,16 @@ void Castro::construct_old_gravity_source(MultiFab& source, MultiFab& state_in,
     if (castro::verbose > 1)
     {
         const int IOProc   = ParallelDescriptor::IOProcessorNumber();
-        Real      run_time = ParallelDescriptor::second() - strt_time;
+        amrex::Real run_time = ParallelDescriptor::second() - strt_time;
+        amrex::Real llevel = level;
 
 #ifdef BL_LAZY
         Lazy::QueueReduction( [=] () mutable {
 #endif
         ParallelDescriptor::ReduceRealMax(run_time,IOProc);
 
-        amrex::Print() << "Castro::construct_old_gravity_source() time = " << run_time << " on level " << level << "\n" << "\n";
+        amrex::Print() << "Castro::construct_old_gravity_source() time = " << run_time
+                       << " on level " << llevel << "\n" << "\n";
 #ifdef BL_LAZY
         });
 #endif
@@ -667,15 +673,17 @@ void Castro::construct_new_gravity_source(MultiFab& source, MultiFab& state_old,
 
     if (castro::verbose > 1)
     {
-        const int IOProc   = ParallelDescriptor::IOProcessorNumber();
-        Real      run_time = ParallelDescriptor::second() - strt_time;
+        const int IOProc = ParallelDescriptor::IOProcessorNumber();
+        amrex::Real run_time = ParallelDescriptor::second() - strt_time;
+        amrex::Real llevel = level;
 
 #ifdef BL_LAZY
         Lazy::QueueReduction( [=] () mutable {
 #endif
         ParallelDescriptor::ReduceRealMax(run_time,IOProc);
 
-        amrex::Print() << "Castro::construct_new_gravity_source() time = " << run_time << " on level " << level << "\n" << "\n";
+        amrex::Print() << "Castro::construct_new_gravity_source() time = " << run_time
+                       << " on level " << llevel << "\n" << "\n";
 #ifdef BL_LAZY
         });
 #endif
diff --git a/Source/gravity/Gravity.cpp b/Source/gravity/Gravity.cpp
index 72a2565935..18eebefbfc 100644
--- a/Source/gravity/Gravity.cpp
+++ b/Source/gravity/Gravity.cpp
@@ -457,13 +457,15 @@ Gravity::solve_for_phi (int               level,
     if (gravity::verbose)
     {
         const int IOProc = ParallelDescriptor::IOProcessorNumber();
-        Real      end    = ParallelDescriptor::second() - strt;
+        amrex::Real end = ParallelDescriptor::second() - strt;
+        amrex::Real llevel = level;
 
 #ifdef BL_LAZY
         Lazy::QueueReduction( [=] () mutable {
 #endif
         ParallelDescriptor::ReduceRealMax(end,IOProc);
-        amrex::Print() << "Gravity::solve_for_phi() time = " << end << " on level " << level << std::endl << std::endl;
+        amrex::Print() << "Gravity::solve_for_phi() time = " << end << " on level "
+                       << llevel << std::endl << std::endl;
 #ifdef BL_LAZY
         });
 #endif
diff --git a/Source/hydro/Castro_ctu_hydro.cpp b/Source/hydro/Castro_ctu_hydro.cpp
index 68c93bf857..d443bbe060 100644
--- a/Source/hydro/Castro_ctu_hydro.cpp
+++ b/Source/hydro/Castro_ctu_hydro.cpp
@@ -1419,12 +1419,13 @@ Castro::construct_ctu_hydro_source(Real time, Real dt)  // NOLINT(readability-co
 
 #ifdef RADIATION
   if (radiation->verbose>=1) {
+      amrex::Real llevel = level;
 #ifdef BL_LAZY
     Lazy::QueueReduction( [=] () mutable {
 #endif
        ParallelDescriptor::ReduceIntMax(nstep_fsp, ParallelDescriptor::IOProcessorNumber());
        if (ParallelDescriptor::IOProcessor() && nstep_fsp > 0) {
-         std::cout << "Radiation f-space advection on level " << level
+         std::cout << "Radiation f-space advection on level " << llevel
                    << " takes as many as " << nstep_fsp;
          if (nstep_fsp == 1) {
            std::cout<< " substep.\n";
@@ -1482,15 +1483,17 @@ Castro::construct_ctu_hydro_source(Real time, Real dt)  // NOLINT(readability-co
 
   if (verbose > 0)
     {
-      const int IOProc   = ParallelDescriptor::IOProcessorNumber();
-      Real      run_time = ParallelDescriptor::second() - strt_time;
+      const int IOProc = ParallelDescriptor::IOProcessorNumber();
+      amrex::Real run_time = ParallelDescriptor::second() - strt_time;
+      amrex::Real llevel = level;
 
 #ifdef BL_LAZY
       Lazy::QueueReduction( [=] () mutable {
 #endif
         ParallelDescriptor::ReduceRealMax(run_time,IOProc);
 
-        amrex::Print() << "Castro::construct_ctu_hydro_source() time = " << run_time << " on level " << level << "\n" << "\n";
+        amrex::Print() << "Castro::construct_ctu_hydro_source() time = " << run_time
+                       << " on level " << llevel << "\n" << "\n";
 #ifdef BL_LAZY
         });
 #endif
diff --git a/Source/hydro/Castro_hybrid.cpp b/Source/hydro/Castro_hybrid.cpp
index b1bdabc4f9..0191587b6f 100644
--- a/Source/hydro/Castro_hybrid.cpp
+++ b/Source/hydro/Castro_hybrid.cpp
@@ -22,8 +22,8 @@ Castro::construct_old_hybrid_source(MultiFab& source, MultiFab& state_old, Real
 
     if (verbose > 1)
     {
-        const int IOProc   = ParallelDescriptor::IOProcessorNumber();
-        Real      run_time = ParallelDescriptor::second() - strt_time;
+        const int IOProc = ParallelDescriptor::IOProcessorNumber();
+        amrex::Real run_time = ParallelDescriptor::second() - strt_time;
 
 #ifdef BL_LAZY
         Lazy::QueueReduction( [=] () mutable {
@@ -65,8 +65,8 @@ Castro::construct_new_hybrid_source(MultiFab& source, MultiFab& state_old, Multi
 
     if (verbose > 1)
     {
-        const int IOProc   = ParallelDescriptor::IOProcessorNumber();
-        Real      run_time = ParallelDescriptor::second() - strt_time;
+        const int IOProc = ParallelDescriptor::IOProcessorNumber();
+        amrex::Real run_time = ParallelDescriptor::second() - strt_time;
 
 #ifdef BL_LAZY
         Lazy::QueueReduction( [=] () mutable {
diff --git a/Source/hydro/Castro_mol_hydro.cpp b/Source/hydro/Castro_mol_hydro.cpp
index 6eb352ac14..9fee21334d 100644
--- a/Source/hydro/Castro_mol_hydro.cpp
+++ b/Source/hydro/Castro_mol_hydro.cpp
@@ -743,8 +743,8 @@ Castro::construct_mol_hydro_source(Real time, Real dt, MultiFab& A_update)
 
     if (verbose > 0)
     {
-        const int IOProc   = ParallelDescriptor::IOProcessorNumber();
-        Real      run_time = ParallelDescriptor::second() - strt_time;
+        const int IOProc = ParallelDescriptor::IOProcessorNumber();
+        amrex::Real run_time = ParallelDescriptor::second() - strt_time;
 
 #ifdef BL_LAZY
         Lazy::QueueReduction( [=] () mutable {
diff --git a/Source/problems/Castro_problem_source.cpp b/Source/problems/Castro_problem_source.cpp
index dc4d08f575..68a48c64df 100644
--- a/Source/problems/Castro_problem_source.cpp
+++ b/Source/problems/Castro_problem_source.cpp
@@ -24,8 +24,8 @@ Castro::construct_old_ext_source(MultiFab& source, MultiFab& state_in, Real time
 
     if (verbose > 1)
     {
-        const int IOProc   = ParallelDescriptor::IOProcessorNumber();
-        Real      run_time = ParallelDescriptor::second() - strt_time;
+        const int IOProc = ParallelDescriptor::IOProcessorNumber();
+        amrex::Real run_time = ParallelDescriptor::second() - strt_time;
 
 #ifdef BL_LAZY
         Lazy::QueueReduction( [=] () mutable {
@@ -95,8 +95,8 @@ Castro::construct_new_ext_source(MultiFab& source, MultiFab& state_old, MultiFab
 
     if (verbose > 1)
     {
-        const int IOProc   = ParallelDescriptor::IOProcessorNumber();
-        Real      run_time = ParallelDescriptor::second() - strt_time;
+        const int IOProc = ParallelDescriptor::IOProcessorNumber();
+        amrex::Real run_time = ParallelDescriptor::second() - strt_time;
 
 #ifdef BL_LAZY
         Lazy::QueueReduction( [=] () mutable {
diff --git a/Source/reactions/Castro_react.cpp b/Source/reactions/Castro_react.cpp
index ee9807dfb9..b7a422f1dd 100644
--- a/Source/reactions/Castro_react.cpp
+++ b/Source/reactions/Castro_react.cpp
@@ -449,14 +449,16 @@ Castro::react_state(MultiFab& s, MultiFab& r, Real time, Real dt, const int stra
     if (verbose > 0)
     {
         const int IOProc   = ParallelDescriptor::IOProcessorNumber();
-        Real      run_time = ParallelDescriptor::second() - strt_time;
+        amrex::Real run_time = ParallelDescriptor::second() - strt_time;
+        amrex::Real llevel = level;
 
 #ifdef BL_LAZY
         Lazy::QueueReduction( [=] () mutable {
 #endif
         ParallelDescriptor::ReduceRealMax(run_time,IOProc);
 
-        amrex::Print() << "Castro::react_state() time = " << run_time << " on level " << level << "\n" << "\n";
+        amrex::Print() << "Castro::react_state() time = " << run_time
+                       << " on level " << llevel << "\n" << "\n";
 #ifdef BL_LAZY
         });
 #endif
@@ -852,15 +854,17 @@ Castro::react_state(Real time, Real dt)
 
         amrex::Print() << "... Leaving burner on level " << level << " after completing full timestep of burning." << std::endl << std::endl;
 
-        const int IOProc   = ParallelDescriptor::IOProcessorNumber();
-        Real      run_time = ParallelDescriptor::second() - strt_time;
+        const int IOProc = ParallelDescriptor::IOProcessorNumber();
+        amrex::Real run_time = ParallelDescriptor::second() - strt_time;
+        amrex::Real llevel = level;
 
 #ifdef BL_LAZY
         Lazy::QueueReduction( [=] () mutable {
 #endif
         ParallelDescriptor::ReduceRealMax(run_time, IOProc);
 
-        amrex::Print() << "Castro::react_state() time = " << run_time << " on level " << level << std::endl << std::endl;
+        amrex::Print() << "Castro::react_state() time = " << run_time
+                       << " on level " << llevel << std::endl << std::endl;
 #ifdef BL_LAZY
         });
 #endif
diff --git a/Source/rotation/Castro_rotation.cpp b/Source/rotation/Castro_rotation.cpp
index cfc6a2edc5..9c98f878ca 100644
--- a/Source/rotation/Castro_rotation.cpp
+++ b/Source/rotation/Castro_rotation.cpp
@@ -34,14 +34,16 @@ Castro::construct_old_rotation_source(MultiFab& source, MultiFab& state_in, Real
     if (verbose > 1)
     {
         const int IOProc   = ParallelDescriptor::IOProcessorNumber();
-        Real      run_time = ParallelDescriptor::second() - strt_time;
+        amrex::Real run_time = ParallelDescriptor::second() - strt_time;
+        amrex::Real llevel = level;
 
 #ifdef BL_LAZY
         Lazy::QueueReduction( [=] () mutable {
 #endif
         ParallelDescriptor::ReduceRealMax(run_time,IOProc);
 
-        amrex::Print() << "Castro::construct_old_rotation_source() time = " << run_time << " on level " << level << "\n" << "\n";
+        amrex::Print() << "Castro::construct_old_rotation_source() time = " << run_time
+                       << " on level " << llevel << "\n" << "\n";
 #ifdef BL_LAZY
         });
 #endif
diff --git a/Source/sources/Castro_geom.cpp b/Source/sources/Castro_geom.cpp
index 282c52b3b1..e3c92732e0 100644
--- a/Source/sources/Castro_geom.cpp
+++ b/Source/sources/Castro_geom.cpp
@@ -33,8 +33,8 @@ Castro::construct_old_geom_source(MultiFab& source, MultiFab& state_in, Real tim
 
   if (verbose > 1)
   {
-      const int IOProc   = ParallelDescriptor::IOProcessorNumber();
-      Real      run_time = ParallelDescriptor::second() - strt_time;
+      const int IOProc = ParallelDescriptor::IOProcessorNumber();
+      amrex::Real run_time = ParallelDescriptor::second() - strt_time;
 
 #ifdef BL_LAZY
       Lazy::QueueReduction( [=] () mutable {
diff --git a/Source/sources/Castro_sources.cpp b/Source/sources/Castro_sources.cpp
index 9a0be7b695..1c4981a7db 100644
--- a/Source/sources/Castro_sources.cpp
+++ b/Source/sources/Castro_sources.cpp
@@ -159,14 +159,16 @@ Castro::do_old_sources(
     if (verbose > 0)
     {
         const int IOProc   = ParallelDescriptor::IOProcessorNumber();
-        Real      run_time = ParallelDescriptor::second() - strt_time;
+        amrex::Real run_time = ParallelDescriptor::second() - strt_time;
+        amrex::Real llevel = level;
 
 #ifdef BL_LAZY
         Lazy::QueueReduction( [=] () mutable {
 #endif
         ParallelDescriptor::ReduceRealMax(run_time,IOProc);
 
-        amrex::Print() << "Castro::do_old_sources() time = " << run_time << " on level " << level << "\n" << "\n";
+        amrex::Print() << "Castro::do_old_sources() time = " << run_time
+                       << " on level " << llevel << "\n" << "\n";
 #ifdef BL_LAZY
         });
 #endif
@@ -242,15 +244,16 @@ Castro::do_new_sources(
 
     if (verbose > 0)
     {
-        const int IOProc   = ParallelDescriptor::IOProcessorNumber();
-        Real      run_time = ParallelDescriptor::second() - strt_time;
+        const int IOProc = ParallelDescriptor::IOProcessorNumber();
+        amrex::Real run_time = ParallelDescriptor::second() - strt_time;
+        amrex::Real llevel = level;
 
 #ifdef BL_LAZY
         Lazy::QueueReduction( [=] () mutable {
 #endif
         ParallelDescriptor::ReduceRealMax(run_time,IOProc);
 
-        amrex::Print() << "Castro::do_new_sources() time = " << run_time << " on level " << level << "\n" << "\n";
+        amrex::Print() << "Castro::do_new_sources() time = " << run_time << " on level " << llevel << "\n" << "\n";
 #ifdef BL_LAZY
         });
 #endif