Has anyone done an analysis of how having a smaller MTU affects slow-start? It seems to me that a TCP session would come up to speed about 3 times faster with a mtu of ~500 vs a mtu of ~1500.

It doesn't, because the smaller MTU doesn't mean you get more acks per unit time, just that you get more acks overall. So you still get the same slow start increase pattern, only your segments are 1/3 the size they could be, so your data rate is 1/3 what it could otherwise be (until cwnd is big enough to consume the available bandwidth). The improved-performance-with-low-MTU is a puzzle I'd certainly like to understand better. It seems step 1 is to eliminate the possibility that the 1500 MTU is leading to fragmentation. Vern