How much overhead does it take to record disk latency data in realtime? Does it slow down the disks? I read on the guy's blog that ZFS's write caching meant that end users didn't experience increased latency when the guy yelled at the disks, but surely there must be some performance cost to all this instrumentation.