-
Enhancement
-
Resolution: Done
-
Critical
-
5.5.0.Final, 6.0.1.Final
-
None
For most use cases (including all examples in 6.0) the number of entities is stable during planning. For those use cases, the entity selectors should not default to cacheType STEP, but PHASE instead.
For cases with a large dataset (at least 10k entities), for fast-stepping algorithms (LA, SA), this can heavily affect the average calculate count per second.
benchmark machineReassignment stepLimit dataset: model_b_8.txt (50K entities) logLevel: info LA maximumStepCount 100000: old way (STEP cache) 2014-02-12 11:56:21,169 [pool-2-thread-1] INFO Solving ended: time spend (242314), best score (0hard/-12997504249soft), average calculate count per second (2845). LA maximumStepCount 100000: with PHASE cache 2014-02-12 11:51:41,911 [pool-2-thread-1] INFO Solving ended: time spend (33062), best score (0hard/-12997504249soft), average calculate count per second (20855). => 633% faster!!! TS maximumStepCount 10000: old way (STEP cache) 2014-02-12 13:23:28,226 [pool-2-thread-1] INFO Solving ended: time spend (361940), best score (0hard/-1214638137soft), average calculate count per second (55271). TS maximumStepCount 10000: with PHASE cache 2014-02-12 13:29:50,181 [pool-2-thread-1] INFO Solving ended: time spend (334319), best score (0hard/-1214638137soft), average calculate count per second (59837). => 8% faster
The workaround is simple, just force cacheType PHASE manually:
<localSearch>
<unionMoveSelector>
<changeMoveSelector>
<entitySelector>
<cacheType>PHASE</cacheType>
</entitySelector>
</changeMoveSelector>
<swapMoveSelector>
<entitySelector>
<cacheType>PHASE</cacheType>
</entitySelector>
</swapMoveSelector>
</unionMoveSelector>
...
</localSearch>
- relates to
-
-
- Open
-