If CloudFlare is using Kernel version 3.3 or higher, they should look into using fq_codel as their scheduler instead of pfifo_fast to decrease latency under load. I suspect the 16MB buffer in the NIC isn't doing them any favors.
"We are indeed looking at Codel. We were actually working on backporting BQL+Codel to the 2.6.x kernel but the Google guys finally got the network stack under control enough for us to deploy >3.3. The 16MB of buffers hasn't hurt us much yet, and may in the long run save us from switches that have too shallow a buffer for the high contention ratios we run on the switch." -LinuXY