I was running some simple experiments with sysbench on EC2 images. I wanted to see how fast the new EBS 1000 IOPS instance was vs. the ephemeral storage (SSDs) on a High I/O instance.
The process was simple: Fire up Ubuntu 12, install sysbench, mkfs.ext3 on the volume, run sysbench prepare, and then fire off a series of rndrw tests with sysbench with # threads from 2^1 to 2^(whenever it broke). Here’s the interesting graph. The yellow line is the IO performance of /dev/xvdb on a h1.4xlarge and the blue line is IO performance (rndrw) for a high performance EBS volume @ 1000 IOPS.
I’m sure there’s some magic tweak I could have used in /etc/fstab to improve performance, but from what I can see the answer to this question is that the highIO instance buys you about an order of magnitude improvement.