Results after running outside of a VM:

Looked into this quite a bit and tried a few alternative implementations (none of which had a definitive advantage over the current approach). At the end of the day, I'm not confident in the way that we are measuring performance (even with the same code, I've often seen benchmark differences of 10% and more). I am working on a more comprehensive strategy to determine what our overall performance goals are.

I re-ran some of the benchmarks outside of a VM context with slightly less worrying results (see screenshot below). Since we don't have a comprehensive plan of what performance levels are acceptable, I'm closing this for now (with a new issue to comes soon to discuss strategies for determining what our performance goals are: #214).