Many Resident Task Computing in Support of Dynamic Ensemble Computations

Modern agent-based models (ABMs) that can simulate large populations are increasingly used to answer important questions in a variety of topic areas. For these models to serve as useful electronic laboratories, they require a battery of experimental analyses, including calibration and validation, sensitivity analysis, optimization, data assimilation, and multimodel integration. These analyses are typically run as workflows that coordinate ensembles of simulations on open science highperformance computing (HPC) systems, enabling large numbers of concurrent simulations. The execution model of these workflows is driven by the algorithmic logic, which can involve arbitrary loops and recursive behavior in pursuit of model convergence. In this paper, we describe a many resident task computing framework that coordinates ABM ensembles on open science HPC systems, driven by a pluggable, stateful optimization engine. While this model challenges conventional notions about workflows consisting of many run-to-completion tasks, we show how it enables rapid prototyping of many parameter search and optimization strategies. Our focus here is on ABMs and optimization, but these techniques are more widely applicable to any black-box scientific code and adaptive parameter space characterization. Ultimately, the goal is to democratize the use of HPC resources by enabling non-expert researchers to take advantage of the extreme scale systems that will be available in the next few years.