Скачать или смотреть Optimizing Parallel R Programs via Dynamic Scheduling Strategies

We present scheduling strategies for optimizing the overall runtime of parallel R programs. Our proposal improves upon the existing mclapply function of the parallel package, which already offers a load balancing option that dynamically allocates tasks to worker processes. However, this mechanism has shortcomings when used on heterogeneous hardware architectures, where different CPU cores might have vastly different performance characteristics. We thus propose to enhance mclapply with a new parameter that allows mapping tasks to specific CPUs. The new affinity.list parameter, already available on the R-devel branch, allows setting a so-called CPU affinity mask that specifies on which CPU a given task is allowed to run. We demonstrate the benefits of the new mclapply version by showing how it can speed up parallel applications like parameter tuning. In this case study, we develop a regression model that guides the scheduling by estimating the runtime of a task for each processor type based on previous executions. In a series of code examples, we explain how this approach can be generalized to develop efficient scheduling strategies for parallel R programs.