Next: Workstation Clusters
Up: Massively Parallel Processors
Previous: Massively Parallel Processors
Although MPPs are in the market already for a while, non-converging discussions about programming-paradigms kept commercial software-vendors from starting to port their codes.
Today we can distinct between three principal paradigms:
- Message Passing can be characterised by explicit communications and explicit data-distribution. Typical and portable implementations
are MPI and PVM.
- Data-/Worksharing programming-models started off with the SIMD-machines but have grown to really usable and powerful programming-environments. Examples
are HPF (High Performance Fortran) and CRAFT (Cray Adaptive Fortran). With the
latter it is even possible, to mix message-passing and distributed-shared-memory techniques together with data-/worksharing in a very flexible way.
Data-/Worksharaing-models are characterised by explicit data distribution, but implicit communications.
- Shared-Virtual-Memory is the (though not yet very successful) attempt to automize everything: communications as well as data-distribution. The vehicle for this is the paging-mechanism of the virtual-memory-system.
The problem is the coherency, which has to be kept over the whole system.
The situation today is still not that there is a convergence to one of these paradigms, but that there is convergence within these paradigms.
Initiatives like MPI and HPF at least give the software-developer a certain protection of his investments.
However, he still has to decide about the programming paradigm itself. Taking a model like HPF makes it easier than in message-passing, but the performance that can be squeezed out is still lower and the variety of hardware-platforms that can be used is smaller.
Vendor specific interfaces have the potential to bring highest performance, but naturally lack of portability.
top500@rz.uni-mannheim.de
Tue Nov 14 16:00:38 PST 1995