Responsive Parallel Computation

As parallel (multicore) hardware proliferates, there is growing interest in developing languages and techniques for writing and reasoning about parallel programs. One important direction is to abstract away from the details of how parallelism is implemented by using implicitly parallel programming languages and reasoning about performance using abstract cost models based on the metrics of work and span, which can then be mapped to actual performance by taking advantage of classic results in the field, such as Brent’s Theorem. While very effective for compute-bound applications, the integration of these methods with effects such as input-output has not been well understood, even though many applications (e.g., games and servers) increasingly involve such interaction. In this paper, we propose responsive parallel computing to bring together performant implicit parallelism with responsive interaction. We first present a parallel language that provides separate mechanisms for interaction and parallelism, and for prioritization of computations. The language separates foreground (high priority) and background (low priority) computations using a type system based on linear temporal logic and comes with both an operational semantics and a cost model. The cost model is based on a refinement of the work-span model that introduces the notions of foreground work and span. As our key result, we prove a “Brent-type” theorem that establishes a bound on the responsiveness as well as the completion time of a computation. We present a small implementation and several examples that give some evidence of the practical implications of our results.