Heterogeneous Computing Issues
Experience developing math software using clusters of workstations (PVM & MPI).
Some users would like/want/expect/demand the same results they obtained yesterday the next time they run the application.
How can we guarantee reproducibility?
Issues to be considered:
- Communication of floating point values between processors
- Machine precision and other machine parameters
- Checking global floating-point arguments
- Algorithmic Integrity
- Deadlock
-