A1393
Title: Task-parallel model selection and numerical optimization in R
Authors: Panagiotis Hadjidoukas - University of Patras (Greece) [presenting]
Eugenia-Maria Kontopoulou - University of Patras (Greece)
Efstratios Gallopoulos - University of Patras (Greece)
Abstract: Task-parallel execution can improve the scalability of computational statistics workflows in R while preserving a familiar programming environment for data analysts and statisticians. TorcR interfaces R with the TORC task runtime system, exposing asynchronous task submission, future-style result handling and dynamic scheduling on multi-core and distributed-memory systems. Its programming model follows the widely used futures abstraction, allowing R users to express parallel computations without explicitly managing processes, communication or low-level synchronization. The approach is evaluated on workloads involving repeated cross-validation, Monte Carlo simulation, numerical optimization and parallel data preprocessing. Derivative-free optimization is used as a representative workload because it can optimize arbitrary objective functions expressed in R and exposes natural parallelism across candidate evaluations or independent runs. Preprocessing tasks include feature-wise summaries and standard scaling over large tabular datasets. These workflows generate independent or semi-independent tasks with variable execution times due to model-fitting costs, convergence behavior, data imbalance or objective-function complexity. The evaluation compares TorcR with sequential R and static parallel execution. The results show that TorcR makes parallel hardware more accessible for scalable statistical validation, optimization and data-analysis pipelines in R.
