Create and explore imaginative 3D worlds with our new product Dreamfield (opens in a new tab)

Effect of Greedy Policy Switching on Scheduling Outcomes

Determine whether greedy switching among scheduling policies improves HPC scheduling outcomes when driven by CQTwin’s online projection interface.

Background

CQTwin’s online twin can fork the current scheduler-visible state across multiple candidate policies and expose their projected outcomes to an external controller. The paper demonstrates this capability with a greedy controller that switches the committed policy 3,084 times during the Polaris-2023 trace.

The evaluation reports the computational overhead and confirms that projections do not alter the committed schedule, but it does not assess whether the greedy controller’s policy switching improves scheduling performance. The effectiveness of greedy switching therefore remains unresolved; the paper explicitly distinguishes this question from CQTwin’s primary purpose as an analytics and decision-support toolkit rather than a new scheduling method.

References

We deliberately do not evaluate whether greedy switching improves scheduling outcomes, as CQTwin's goal is operational analytics and decision support rather than developing a new scheduling method.

— CQTwin: A Digital-Twin Toolkit for Operational Analytics of HPC Schedulers  (2610.26609 - Xu et al., 19 Oct 2026) in Section 5.3, subsection “Online Overhead” (Section label: sec:onlineeval)