tag:blogger.com,1999:blog-37750309.post776212761762832063..comments2024-02-16T09:39:29.877+00:00Comments on GridPP storage news: ZFS vs Hardware RaidGreighttp://www.blogger.com/profile/10952696274183590650noreply@blogger.comBlogger1125tag:blogger.com,1999:blog-37750309.post-84788712787583069122017-02-02T14:56:42.006+00:002017-02-02T14:56:42.006+00:00You really should run tests with files much bigger...You really should run tests with files much bigger than RAM size, else caching will get in the way and make the results irrelevant. You have 24G of RAM, you should run your tests with 48G fileS.<br /><br />You may also run "echo 3 > /proc/sys/vm/drop_cache" to empty the file cache between runs for more consistent results.<br /><br />Another point to take into account is the IO scheduler. Most distributions use cfq (Completely Fair Scheduler) as a default, unfortunately it's most of the time a poor choice for a server, particularly when using hardware RAID. Use "noop" scheduler for perfectly fair tests: run "echo noop > /sys/block//queue/scheduler" for all drives.<br /><br />Last you may need to adjust IO queue length and read ahead. Default values are quite correct for old ATA drives with small cache, but very suboptimal for RAID arrays. Most RAID controller needs much longer queues than the default 128 (512 or 1024): echo 1024 > /sys/block//queue/nr_requests <br />And most hardware RAID controller have large caches that give better results with sequential IO with big read-ahead values: echo 8192 > /sys/block:/queue/read_ahead_kb<br /><br />wazooxhttps://www.blogger.com/profile/15736555357082535545noreply@blogger.com