On 27/09/2016 4:13 AM, mabi wrote:> I would also say do not forget to set "sync=disabled".I wouldn't be doing that - very high risk of gluster corruption in the event of power loss or server crash. Up to 5 seconds of writes could be lost that way. If writes aren't fast enough I'd add a SSD partition for slog. Preferably a data center quality one. -- Lindsay Mathieson
2016-09-26 23:08 GMT+02:00 Lindsay Mathieson <lindsay.mathieson at gmail.com>:> I wouldn't be doing that - very high risk of gluster corruption in the event > of power loss or server crash. Up to 5 seconds of writes could be lost that > way. > > If writes aren't fast enough I'd add a SSD partition for slog. Preferably a > data center quality one.You are right, I don't want to loose some data by using 'sync=disabled' Also, I would like to try to force the use of SLOG to an Enterprise SSD (also used for L2ARC) by setting 'sync=always' In this way all writes are synchronous and made against the SSD. This should also "hide" the overhead of RAIZ2 or RAIDZ3 as writes are on SLOG and not directly on the vpool So: 2x Enterprise SSD (Intel 3600) 1 partition on both is mirrored and used as SLOG The other 2 partitions (one on each SSD) are used as L2ARC and not mirrored (is not possible and has nonsense) Everything configure with 'sync=always' RAIDZ-3 made with 12x 4TB disks (I don't know the resilvering time for a huge vpool like this) In this way, an SSD failure doesn't cause data loss (SLOG is mirrored) and L2ARC cache is automatically disabled by ZFS At the same time, there isn't any write penalty for the RAIZ3 because all writes are made to the SLOG and only after that wrote on the pool What do you think?
That's not correct. There is no risk of corruption using "sync=disabled". In the worst case you just end up with old data but no corruption. See the following comment from a master of ZFS (Aaron Toponce): https://pthree.org/2013/01/25/glusterfs-linked-list-topology/#comment-227906 Btw: I have enterprise SSD for my ZFS SLOG but in the case of GlusterFS I see not much improvement. The real performance improvement comes by disabling ZFS synchronous writes. I do that for all my ZFS pools/partitions which have GlutserFS on top. -------- Original Message -------- Subject: Re: [Gluster-users] Production cluster planning Local Time: September 26, 2016 11:08 PM UTC Time: September 26, 2016 9:08 PM From: lindsay.mathieson at gmail.com To: gluster-users at gluster.org On 27/09/2016 4:13 AM, mabi wrote:> I would also say do not forget to set "sync=disabled".I wouldn't be doing that - very high risk of gluster corruption in the event of power loss or server crash. Up to 5 seconds of writes could be lost that way. If writes aren't fast enough I'd add a SSD partition for slog. Preferably a data center quality one. -- Lindsay Mathieson _______________________________________________ Gluster-users mailing list Gluster-users at gluster.org http://www.gluster.org/mailman/listinfo/gluster-users -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://www.gluster.org/pipermail/gluster-users/attachments/20160928/7baeae2c/attachment.html>