Bulk synchronous parallel programming is a programming model that divides the execution flow into two alternating phases:
- local computation
- global communication and synchronisation
While MPI itself is rather hairy to program, BSP facilitates parallel programming a lot. It automatically avoids deadlocks as all processes wait for each other after the communication step. In that respect it looks very much like the map-reduce model that google uses, but seems to be a generalisation as it also allows both global and distributed arrays. Also a BSP program is a program for N processors while map-reduce is a model for N-1 slave processors and one additional master. In the weeks to come I will tell more about the experience I will make with that programming model. Stand by.
Keine Kommentare:
Kommentar veröffentlichen