Very interesting use case!
The correlation seems to be doing what it is doing because its a single thread that is executing the correlation request and therefore its just 1 instances that does the correlation!
Assuming there is not a api i am missing that would allow you do do some fancy job processing,
You could do the following:
Create a message query to find all process instances that match your variable query. This will create a array of process instances that you want to correlate against.
Then you can pass that array/collection into a Parallel-Multi-Instance Send Task, and have the multi-instance set as Async Before. This will generate a single job for each of the process instances that you want to correlate against.
Then the job executors from the cluster will pick up on those jobs and start processing in parallel and across all of the nodes.
edit: You might also want to batch the jobs together in say groups of 100, or whatever. This way you dont have overhead of creating Jobs in the DB for just a single process correlation.
Would be very interested to see some performance graphs on this!