[tavert]

All of the MPI communication was happening in C code (the built-in distributed code in Julia exists, and sometimes works, but that's the nicest thing you can say about it - it has not been leaned on heavily at all, and is often awkward to use). Lots of low level bit twiddling was involved in Celeste. Without Keno that application would have been lucky to top 100-200 teraflops.