I fired up ten 2x(dc2.large) clusters, one immediately after the other, benchmarking each cluster (which means benchmarking the nodes in each cluster), and then

I fired up ten 2x(dc2.large) clusters, one immediately after the other, benchmarking each cluster (which means benchmarking the nodes in each cluster), and then produced the benchmark results, max/avg/min, for each benchmark type. There's a lot of variation. There's an intriguing hint that this really is on a per-node basis, with nodes usually being normal, but there being bad or particular bad nodes. This all matters a lot, because queries execute as slowly as the slowest node in a cluster.