- Initial fields allocated after creating a region
- Track values escaping from futures
- More flexible selection of checkpoint location (directory prefix, file patterns, etc.)
- Double allocate regions to maintain consistent tree IDs on replay?
- Needed because otherwise I don't think subtask allocation of partitions can be safe---if we no-op out the tasks, we only have the Future, which inherently capture handle IDs
- (In other words, we need stable handle IDs to make nested partition creation work without instrumenting nested tasks)
- Bonus is that it avoids runtime creation of new region trees
- Nested partitions
- Distributed region checkpoints
- Add timing code on checkpoints so I can start tracking that
- Do we handle recusive destroy semantics correctly?