SC17 Denver, CO

Contemporary Design of Supercomputer Experiments

Authors: Dmitry Duplyakin (University of Utah)

Abstract: This BoF facilitates community discussion on the topics in contemporary design of supercomputer experiments. Design of Experiments (DOE) is a critical component of the knowledge discovery process in science and engineering and has impact in a wide variety of fields, including supercomputing. The scale of contemporary supercomputer simulations and the growth of computational capabilities toward exascale dictate the need for practical DOE that incorporate adaptive, multi-objective, and cost-efficient strategies. In this BoF, we propose to discuss promising state-of-the-art techniques and facilitate the community building around the research and development focused on innovative supercomputing DOE.

Long Description: This BoF aims to bring together researchers from many disciplines, including (but not limited to) Scientific Computing, Machine Learning, Performance Modeling and Benchmarking, Design of Experiments (DOE), Uncertainty Quantification (UQ), and broader statistical areas, with the goal of fostering conversations on the following topics:

- How do contemporary DOE methods in computing differ from classical methods?

- Active Learning (a.k.a. optimal or sequential experimental design): state-of-the-art techniques and main challenges in computing applications

- Dimensionality reduction: how and when can Principal Component Analysis (PCA), Active Subspaces, and sensitivity analysis improve DOE?

- How can we minimize the cost of the experiments needed to create reliable empirical performance models?

- How can we characterize the reproducibility and variability of performance predictions, both on a single machine and across machines of similar architecture?

- When performing experiments, what machine state information should be captured to facilitate reproducibility and uncertainty quantification?

- How to integrate domain constraints and enable expert knowledge in DOE? How to deal with model bias in DOE?

These questions are discussed in many recent studies. To the best of our knowledge, however, there is a shortage of venues for discussing ongoing and early work in the related areas that provide opportunities to state pending research questions, share successful and unsuccessful experiences, and gather constructive feedback. To bridge this gap, we propose this BoF on Contemporary Design of Supercomputer Experiments and plan to structure it as follows. We will begin with lightning talks on the aforementioned topics followed by short Q&A sessions. We will continue with a panel where the organizers and invited speakers can comment and question the presented ideas and technologies. We will conclude with an open forum and discussion between the audience and the speakers. In the course of this discussion, we will gather comments from the audience that demonstrate the interest in organization of a future meeting or a workshop with a similar scope. If the attendees express such interest, we will discuss tentative formats, timelines, and specific areas of focus.

Conference Presentation: pdf

Birds of a Feather Index