In my current contract we have a nice database server with 4 datacentre quality SSD’s arranged in raid5, a battery-backed (Dell PERC / megaraid) controller and very poor disk performance. As I was doing a dump of the database (with parallel process, obviously), iostat was showing 100% disk utilization at 200-300 iops and 40-60mb/sec. Really not what it should be. After poking the raid card a lot with the MegaCli tool, I discovered that by changing the Linux IO scheduler to noop or deadline we suddenly got great performance – 50% disk utilization at 1300 iops and 200mb/sec read. So, an issue with the CFQ scheduler then as everywhere seems to suggest? In fact not – it turns out the root of the problem is that because of the raid controller the kernel did not detect that the underlying disks were SSD rather than HDD. Just by tweaking this setting CFQ provides the same performance as noop or deadline schedulers (in fact better for our postgres workload). So, here’s the command we used to get a 1000% improvement in database performance:

echo 0 > /sys/block/sdb/queue/rotational