Marion Sudvarg
Current Institution: Washington University in St. Louis
Email: msudvarg@wustl.edu
Bio: Marion Sudvarg is a PhD student studying computer science at Washington University in St. Louis (WUSTL). He earned bachelor’s degrees in math and physics, and a master’s degree in computer science with an emphasis on data mining and machine learning, from WUSTL. His research interests are in developing robust, adaptable real-time computing systems, and he works with the ADAPT collaboration to develop real-time data analysis algorithms and systems for multi-messenger astronomy. His work has been published in the Real-Time Systems journal and the Proceedings of Science, presented at the ICRC and SC conferences and the UrgentHPC workshop, and will be presented at the upcoming RTNS conference. He is additionally involved in teaching, having instructed WUSTL’s upper-level undergraduate Operating Systems Organization and graduate-level Advanced Operating Systems courses. He led a significant restructuring of content for both courses, and developed the majority of material and assignments now used in Advanced Operating Systems.
Abstract: Generalized Elastic Scheduling for Nonlinear Cost Functions and Parallel Tasks
My research focuses on design and implementation of real-time systems that are adaptable to dynamic workloads. The elastic task model provides a framework for handling overload for weighted linear compression of individual task utilizations, for which I demonstrated a quasilinear-time solution and a linear admission control algorithm. I am seeking to apply its generalized quadratic optimization problem to the broader space of constrained resource allocation (e.g. CPU cores, memory), and to parallel tasks, for which elasticity of individual subtasks might be considered separately. A target application is APT, a planned satellite that will detect and localize gamma-ray bursts in real-time; the burst localization software’s cost function (expected angular error) is nonlinear over the utilizations of its constituent tasks. We are exploring new notions of non-linear elasticity and formulating more complex optimization problems to define how task utilizations can be adjusted to maintain real-time guarantees, while minimizing the localization error.