On Riser A, put the 2 GB modules in the front slots (1 and 2) and move the 1GB ones to the rear, and see if that improves things.
For performance reasons alone, you always want to put your highest-capacity RAM up front, otherwise you incur a small latency cost when the smaller modules up front "fill up" and have to mediate access to the rear modules.
But also for reliability. For some occult reason, I've found that running with smaller modules in front always seems to generate more ECC errors. When I first ran my Netlist 1GB modules behind Apple 512MB ones, I'd get cascading ECC errors that disappeared once I put the higher capacity RAM up front.