We need to call into MPI every once in a while as otherwise polling on a local variable may not trigger progress if that is needed.

Codecov Report

Merging #712 into development will increase coverage by 1.12%. The diff coverage is n/a.

@@               Coverage Diff               @@
##           development     #712      +/-   ##
===============================================
+ Coverage        84.06%   85.19%   +1.12%     
===============================================
  Files              336      336              
  Lines            24954    24945       -9     
  Branches         11349    11540     +191     
===============================================
+ Hits             20977    21251     +274     
  Misses            3693     3693              
+ Partials           284        1     -283     

Impacted Files	Coverage Δ
dart-impl/mpi/src/dart_communication.c	69.44% <ø> (-0.33%)	:arrow_down:
dash/include/dash/pattern/BlockPattern1D.h	99.49% <0.00%> (+0.50%)	:arrow_up:
dash/include/dash/pattern/ShiftTilePattern.h	93.06% <0.00%> (+0.66%)	:arrow_up:
dash/include/dash/memory/GlobHeapMem.h	97.18% <0.00%> (+0.70%)	:arrow_up:
dash/include/dash/GlobRef.h	92.03% <0.00%> (+0.88%)	:arrow_up:
.../include/dash/matrix/internal/LocalMatrixRef-inl.h	95.19% <0.00%> (+0.96%)	:arrow_up:
dash/include/dash/matrix/internal/Matrix-inl.h	98.17% <0.00%> (+1.21%)	:arrow_up:
dash/include/dash/memory/GlobStaticMem.h	90.90% <0.00%> (+1.29%)	:arrow_up:
dash/include/dash/pattern/TilePattern.h	95.53% <0.00%> (+1.31%)	:arrow_up:
dash/include/dash/pattern/BlockPattern.h	99.15% <0.00%> (+1.41%)	:arrow_up:
... and 32 more

After having given this some more thought I modified the PR to always call into MPI for local memory accesses. If fast local access without progress is required the application should just dereference the native pointer provided by DASH. Once we are in DART we have lost the latency race anyway. Everything else just adds complexity.

+1