Agreed… It took me almost 2 years of tweaking and testing to get the performance I wanted.
Different workloads require different configurations. Test different configurations and find what works best for you! > On Jul 4, 2016, at 2:15 PM, t...@encoding.com wrote: > > I would highly stress, regardless of whatever solution you choose - make sure > you test actual workload performance before going all-in. > > In my testing, performance (esp. iops and latency) decreased as I added > bricks and additional nodes. Since you have many spindles now, I would > encourage you to test your workload up to and including the total brick count > you ultimately expect. RAID level and whether it’s md, zfs, or hardware > isn’t likely to make as significant of a performance impact as Gluster and > its various clients will. Test failure scenarios and performance > characteristics during impairment events thoroughly. Make sure heals happen > as you expect, including final contents of files modified during an > impairment. If you have many small files or directories that will be > accessed concurrently, make sure to stress that behavior in your testing. > > Gluster can be great for targeting availability and distribution at low > software cost, and I would say as of today at the expense of performance, but > as with any scale-out NAS there are limitations and some surprises along the > path. > > Good hunting, > -t > >> On Jul 4, 2016, at 10:44 AM, Gandalf Corvotempesta >> <gandalf.corvotempe...@gmail.com> wrote: >> >> 2016-07-04 19:35 GMT+02:00 Russell Purinton <russell.purin...@gmail.com>: >>> For 3 servers with 12 disks each, I would do Hardware RAID0 (or madam if >>> you don’t have a RAID card) of 3 disks. So four 3-disk RAID0’s per server. >> >> 3 servers is just to start. We plan to use 5 server in shorter time >> and up to 15 on production. >> >>> I would set them up as Replica 3 Arbiter 1 >>> >>> server1:/brickA server2:/brickC server3:/brickA >>> server1:/brickB server2:/brickD server3:/brickB >>> server2:/brickA server3:/brickC server1:/brickA >>> server2:/brickB server3:/brickD server1:/brickB >>> server3:/brickA server1:/brickC server2:/brickA >>> server3:/brickB server1:/brickD server2:/brickB >>> >>> The benefit of this is that you can lose an entire server node (12 disks) >>> and all of your data is still accessible. And you get the same space as >>> if they were all in a RAID10. >>> >>> If you lose any disk, the entire 3 disk brick will need to be healed from >>> the replica. I have 20GbE on each server so it doesn’t take long. It >>> copied 20TB in about 18 hours once. >> >> So, any disk failure would me at least 6TB to be recovered via >> network. This mean an high network utilization and as long gluster >> doesn't have a dedicated network for replica, >> this can slow down client access. >> _______________________________________________ >> Gluster-users mailing list >> Gluster-users@gluster.org >> http://www.gluster.org/mailman/listinfo/gluster-users > > _______________________________________________ > Gluster-users mailing list > Gluster-users@gluster.org > http://www.gluster.org/mailman/listinfo/gluster-users _______________________________________________ Gluster-users mailing list Gluster-users@gluster.org http://www.gluster.org/mailman/listinfo/gluster-users