Parallel programming is difficult. Switching from sequential to parallel programming introduces entire new classes of errors for the programmer to make, such as deadlock and race conditions, which are difficult to debug and complicate testing and correctness proofs. Yet there are entire classes of programs with computational demands so great that sequential solutions are infeasible. We do parallel programming because we care about performance.

How do we know if we are getting good performance? We must observe the execution of our programs to determine if they are making good use of the resources available to them. Once we have made observations, how can we use those observations to improve performance? We can use autotuning to identify variants and parameters that give better performance than others, but this process itself is slow, so we can attempt to synthesize performance data into empirical models to guide the process. Alternately, we may ask not simply for better performance, but for an explanation of performance: automated performance diagnosis. These techniques are all well-developed for the current high-performance computing environment, but the advent of exascale computers will be a disruptive change which will require new parallel programming models, languages, and runtimes, which will in turn require new techniques for performance monitoring and analysis.