Alex Gorrod asked me to look into the large performance drop in the medium-lsm and voxer-10k Jenkins plots. Initial concerns were that the LSM changes were the cause. I spent time investigating that yesterday and it is due to the changes in the wt_random code. The drops are mentioned in @keithbostic's reply in changeset 031e80a. I have some very interesting results from my Mac and I'm running on the AWS HD and SSD systems to compare.