I am benchmarking Drools 4.0 and received excellent results in rule execution.  With our non-optimized test rule set and data, Drools out performed ILOG/JRules by quite a bit.  Congratulations to Drools' team for a job well done.
 
I encountered a strange behavior when trying to test the performance of rule registration, meaning to add Package's to a new instance of RuleBase.  I ran the same test 5 times in a sequence, and the RuleBase instance was discarded right away.  The test ran with different different number of packages, 20, 50, 100, and 200.  The results are as the followings:
 
Added 20 packages to rulebase in 0:00:0.15
Added 20 packages to rulebase in 0:00:0.21
Added 20 packages to rulebase in 0:00:0.21
Added 20 packages to rulebase in 0:00:0.301
Added 20 packages to rulebase in 0:00:0.18
Added 20 packages to rulebase 5 times in 0:00:1.051
 
Added 50 packages to rulebase in 0:00:0.17
Added 50 packages to rulebase in 0:00:1.422
Added 50 packages to rulebase in 0:00:1.252
Added 50 packages to rulebase in 0:00:1.322
Added 50 packages to rulebase in 0:00:1.331
Added 50 packages to rulebase 5 times in 0:00:5.497
 
Added 100 packages to rulebase in 0:00:0.19
Added 100 packages to rulebase in 0:00:5.278
Added 100 packages to rulebase in 0:00:5.297
Added 100 packages to rulebase in 0:00:5.228
Added 100 packages to rulebase in 0:00:5.327
Added 100 packages to rulebase 5 times in 0:00:21.32
 
Added 200 packages to rulebase in 0:00:0.22
Added 200 packages to rulebase in 0:00:21.691
Added 200 packages to rulebase in 0:00:21.481
Added 200 packages to rulebase in 0:00:21.642
Added 200 packages to rulebase in 0:00:21.341
Added 200 packages to rulebase 5 times in 0:01:26.385
 
As you can see, the timing of creating first RuleBase instance is always good, while the subsequent one's were very bad.  Is there an attribute I could adjust in configuration to improve the performance for this scenario?  Please advise if there is other alternatives.  I would be glad to provide details if needed.
 
Your help is greatly appreciated!
-Ming