Scheduler performance testing/improvement with many small algorithms
Motivated by an ATLAS discussion about a possible refactoring of code, I intend to test scheduler performance when there are many algorithms and few dependencies w.r.t. the thread pool size.
Possible areas of investigation:
-
Develop/examine existing stress-test JO(s) for scheduler performance.
-
I have observed better scheduler performance with more nesting of CF structures (assumed due to status caching effect of CF nodes). Test, verify, potentially use to improve performance elsewhere.
-
In the past it was suggested that DATAREADY algorithms should be queued, rather than directly scheduled, to avoid inefficiency when the thread pool is full. Prototype, quantify improvement.
-
Potential to improve scheduling performance by separating EventView/subSlot graph traversal from regular system. Started in !735 (closed), need to complete and quantify improvement.