1: \begin{abstract}

2: {\em Extreme dynamic heterogeneity} in high performance computing systems

3: and the convergence of traditional HPC

4: with new simulation, analysis, and data science approaches

5: impose increasingly more complex requirements

6: on resource and job management software (RJMS).

7: However, there is a paucity of RJMS techniques

8: that can solve key technical challenges

9: associated with those new requirements, particularly 

10: when they are coupled. In this paper, we propose a novel

11: dynamic and multi-level resource model

12: approach to address three key well-known challenges 

13: individually and in combination: i.e.,

14: 1) RJMS dynamism to facilitate job and workflow

15: adaptability, 2) integration of specialized external resources

16: (e.g. user-centric cloud bursting), and

17: 3) scheduling cloud orchestration framework tasks.

18: The core idea is to combine a dynamic directed graph

19: resource model with fully hierarchical

20: scheduling to provide a unified solution to all three key challenges.

21: Our empirical and analytical evaluations of the solution using our

22: prototype extension to Fluxion, a production

23: hierarchical graph-based scheduler, suggest

24: that our unified solution can significantly

25: improve flexibility, performance

26: and scalability across all three problems 

27: in comparison to limited traditional approaches.

28: \end{abstract}

29: