eclipse-qvtd-bot
commented
3 weeks ago By Ed Willink on Dec 12, 2016 05:06
The current processing to identify the schedule is indexing/call stack then connections. In the same way that focusing on connections makes incremental execution tractable, focusing on connections could make the processing less local and reduce many of the impediments to a parallel schedule.
The current schedule is often largely flat, one connection after another loped over by the root (in the right order). The current quadratic/worse efforts at finding hierarachy could be abandoned. Instead.
Create the connections first linked to their producers and consumers; the dependencies establish a partial order.
Group compatible consumers into clusters so that the late consumer merger can merge them. Compatible: same inputs, no partial order violations. Better: perform the late consumer merge here, so that compatible excludes incompatible failure predicates. Better still perform the alternator here too so that switch cases do not inhibit clustering.
Conflict: clustering A-consumers may inhibit clustering B-consumers and vice-versa. Which is better? Potentially NP-complete. Simple heuristic, biggest cluster is best. Better heuristic, determine a profit/loss metric for each clustering choice.
Optimization: single-source connections can be migrated from the root to the source. This may be beneficial if it partitions the connection (e.g. rather than all UML Classes at root, loop over child Classes at each Package.) Alternatively it may be beneficial if the partial order permits the consumers to be called from the producer perhaps facilitating a late producer merge.
Multiple-source connections can also be migrated if there is some common hierarchy.
Overall, it would seem that discovering hierarchy in a flat schedule should be more efficient and more global and more parallel than trying to create it directly.
| --- | --- | | Bugzilla Link | 508669 | | Status | RESOLVED FIXED | | Importance | P3 normal | | Reported | Dec 05, 2016 08:22 EDT | | Modified | Oct 22, 2018 18:30 EDT | | Depends on | 510344 | | Reporter | Ed Willink |
Description
The ScheduleIndexer makes some attempts to make sensible mappings consecutive which helps the late merger. It could probably do much better. The ScheduleIndexer is quadratic. Can this be improved?