It is interesting that this paper seeks for opportunities to systematically address individual node memory pressure problem, which has been largely unexplored. I like figure 8 as it neatly explains how different components interact with each other. However, I have few concerns/questions. 1. I feel like the evaluation should not be drawn from stackoverflow(especially, I don't think we should use fix recommended on Stack Overflow as baselines). 2. In section 5.2, they should run some sensitivity analysis about M and N. 3. In section 5.4, when the authors talk about how to select tasks to interrupt, I think they should provided a more detailed algorithm, instead of just few heuristics. 4. my biggest concern is: will this approach introduce more stragglers? I didn't see any heuristic that prioritize tasks within the same stage.