Are there any implementations of a parallel version of the Rete algorithm?

Question

I have done some research on the Rete algorithm and found several papers on making it parallel (both matching and rule firing) the earliest of which are from the mid '80s. However I can't find any information about any parallel implementation in any of the rule engines in the market today. I understand that parallel rule firing is problematic, but there are several proposed solutions. Does any rule engine implement any parallel version of Rete? If so, which? If not, why not? Is it too hard to achieve any significant performance gain or are there other reasons? Any links to more information is appreciated.

Answer 1

I believe ReteNT does.

We are working on something this summer for the Drools project, I hope we have something prototypable before end of this year. We are takling it from two angles. 1) Merge Rete and Leaps algorithm, that provides lazy beta propagations and evaluations are done on a rule cursor basis (allows for easier parallel rule evaluation). 2) Move from tuple based propagations collection propagations. This will allow nodes to be schedueld as tasks. Further nodes with large number of objects can be locally "map reduced" to further parallelise the creation of the collection that will be propagated from the node.

If you want to chat more, join us on irc :) http://www.jboss.org/drools/irc