Software: Computing in a Parallel Universe

Computing in a Parallel Universe
American Scientist (12/07) Vol. 95, No. 6, P. 476; Hayes, Brian
Multicore chips that facilitate parallel processing will require a major rethinking in program design, writes Brian Hayes. Software for parallel processors is vulnerable to subtle bugs that cannot manifest themselves in strictly sequential programs. Running correct concurrent programs is possible, but a key challenge is that running the same set of programs on the same set of inputs can entail different results depending on the precise timing of events. One concept for addressing this problem is to have the operating system manage the allocation of tasks to processors and balance the workload, which is currently the chief strategy with time-sliced multiprocessing and dual-core chips. Another tactic is to assign this responsibility to the compiler, which, like the earlier strategy, would be the job of expert programmers. But making the most of parallel computing requires all programmers to deal with the problems of creating programs that run efficiently and properly on multicore systems. "We have a historic opportunity to clean out the closet of computer science, to throw away all those dusty old sorting algorithms and the design patterns that no longer fit," Hayes concludes. "We get to make a fresh start."
Click Here to View Full Article