Generally, I'd spend some time writing a script to load data into it, start loading data, respond to it crashing a few hours later, increase the memory available to the process, start up again, and respond to it crashing a few hours later. I was never able to get any reasonably-sized graph[0] working reliably well without using an egregious amount of memory, and knowing that I would continue to face memory issues, I gave up on Neo4j and found another way to solve my problem.
It may be that I simply was not competent at setting it up properly, but no other data store I've worked with has been as hard to get stable over a moderately sized data set. I spoke with some other people who had worked with Neo4j at the time, and they expressed the same issues - they couldn't make it work for any reasonably-sized dataset and had to find another solution.
[0] Not big, mind you, just reasonably-sized. E.g. 4 million nodes, with each node having an average of 5 edges and 2-4 properties.
We found that we could do everything we needed with secondary processing against our document store at runtime for so much less without adding another layer of complexity to the architecture.
Edit: forgot to mention - no we weren't use batch-insertion in all cases, IIRC, we had issues with duplication and had to do check-if-exists -> create-if-not as we were reading from raw data sources that were heavy with duplicates.
I would love to learn about your Neo4j setup, and the issues in detail, I want to make it easier for people in your circumstances in the future to get quickly up and running with Neo4j in a reliable manner. If you're willing to help out, please drop me an email at michael at neotechnology dot com.
http://www.neo4j.org/