Efficiency Of the Programmer Needs To Trump Efficiency Of the Machine For Parallel Programming To Go Mainstream

You can tell when technology has matured, in part by how much the user needs to be an expert in using it. A step in maturing for parallel programming will come as we see that the efficiency of a programmer is more important than the efficiency of the machine.

I was recently asked "isn't parallel programming a solved problem in the HPC world?" The answer is YES and NO. YES -- with expert programmers, and enough effort, amazing things happen. As parallelism is becoming more and more mainstream -- we need more solutions -- so I'd have to say "NO, it is not yet solved."

I have written and talked about key problems to solve before -- with correctness and scaling being the most critical categories, with general programmability and maintainability an important aspect for all programming, not just parallel programming.

But, the most critical change from HPC to mainstream use of parallelism will be the shift in greatly reducing the concern of machine efficiency in favor of programmer efficiency. It's not that HPC programmers do not like efficient programming -- it's really that they are experts willing to expend effort knowing how to get high efficiency -- they generally see that as part of their job and therefore deal with it when others would not. Other programmers, as we get more an more mainstream -- look at parallelism as something that they need to "just work" in general so their focus is on other concerns.

This points to a maturing that is coming: more and more developments to make parallel programming addressing programmer efficiency more than machine efficiency. And because parallelism is becoming mainstream -- there are many more opportunities to amortize the cost of such, or reap the benefits, because of a larger population of users!