Splitting the DRL would probably make things worse.
Have you checked whether the Vm has enought memory available during that process?
By the way, how did you end up with 11K DRL? ;-)
If you remove the salience that should save you a few msec on parsing, should be marginal.
Are you using Java or MVEL dialect on the RHS of the rules? Switching to MVEL might save you some time. There also was a mention of excessive classloading attempts in 4.07. This should be fixed in Drools 5.01. Please see the drools-user mailing list for details.
The Drools engine will deal well with a few thousand rules. If we are talking about several tens of thousands, then I would recommend splitting the rulebase, either vertically (by subject matter) or horizontally (by recurrent snapshots and old rules removal).