What Do Programmers of Parallel Machines Need? A Survey

We performed semistructured, open-ended interviews with 11 professional developers of parallel, scientific applications to determine how their programming time is spent and where tools could improve productivity. The subjects were selected from a variety of research laboratories, both industrial and governmental. The major findings were that programmers would prefer a global over a per-processor view of data structures, struggle with load balancing and optimizations, and need interactive tools for observing the behavior of parallel programs. Furthermore, handling and processing massive amounts of data in parallel is emerging as a new challenge.