Testing with the latest 4.0 snapshot has shown some good improvement on my RPI2 test system.
The start-up time is still twice as slow compared to Nashorn or Groovy, but the heap usage is now only about 90mb bigger. Also the actual heap usage is much better, filling up at a similar rate to Nashorn/Groovy and only triggering minor GCs. Previously it was constantly doing full GCs with noticeable pauses.