Hi Frank, Thanks for your answer. On the ZFS side I've got everything pretty much sorted out (which RAIDZ to use, ZIL or not, compression, etc). To answer your questions I will start small with only two nodes in replicated mode and add more nodes as soon as needed using distributed-replicated mode then. I plan to use a 10 Gbit/s fiber dedicated private network just between the nodes for glusterfs.
So now what I am really wondering is: should I have two bricks per node or just one know that per node I only have one HBA controller. So if I have two bricks on one node they still would be located on exactly the same controller but on two different ZFS volumes. On Monday, February 23, 2015 2:14 PM, Frank Rothenstein <f.rothenst...@bodden-kliniken.de> wrote: Hi, I read your mails but I'm not sure if I fully understand your planned setup. How many Gluster-Nodes have you planned? How fast will be your interconnects between them? I think at first you should test one node itself with all possible options that ZFS has to offer (like compression or even deduplication and also the RAIDZ1/2/3-option). Then of course consider your needs - more capacity vs higher throughput (meaning one or more bricks per node). And also don't forget the LARC2- and log-options offered by ZFS to speed up the backend. And then test the GlusterFS. Adding another dataset/brick is always possible, so you can easily benchmark your system. If you stay at the same replica count there should be no loss in size whether you set up 2 or 4 or 8 bricks (for rep 2), afaik. I think there is no real advise for your setup, ZFS on Linux is not that common, to "gluster" it even less... I hope my english is good enough to let you understand, what i mean ;) Greetings, Frank Am Montag, den 23.02.2015, 08:47 +0000 schrieb ML mail: > Just saw that my post below never got replied and would be very glad if > someone, maybe Niels?, could comment on this. Cheers! > > > > > On Saturday, February 7, 2015 10:13 PM, ML mail <mlnos...@yahoo.com> wrote: > Thank you Niels for your input, that definitely makes me more curious... Now > let me tell you a bit more about my intended setup. First of all my major > difference is that I will not be using XFS but ZFS. Then second major > difference I will not be using any hardware RAID card but one single HBA (LSI > 3008 chip). My inteded ZFS setup would consist of one ZFS pool per node. This > pool will have 3 virtual devices of 12 disks each (6 TB per disk) each using > RAIDZ-2 (equivalent to RAID 6) for integrity. This gives me a total of 36 > disks for a total of 180 TB of raw capacity. > > > > I will then create one big 180 TB ZFS data set (virtual device, file system > or whatever you want to call it) for my GlusterFS brick. Now as mentioned I > could also create have two bricks by creating two ZFS data sets of around 90 > TB each. But as everything is behind the same HBA and same ZFS pool there > will not be any gain in performance nor availability from the ZFS side. > > On the other hand, you mention in your mail that having two bricks per node > means having two glusterfsd processes running and allows me to handle more > clients. Can you tell me more about that? Will I also see any genernal > performance gain? For example in terms of MB/s throughput? Also are there > maybe any disadvantages of running two bricks on the same node, especially in > my case? > > > > > > On Saturday, February 7, 2015 10:24 AM, Niels de Vos <nde...@redhat.com> > wrote: > On Fri, Feb 06, 2015 at 05:06:38PM +0000, ML mail wrote: > > > Hello, > > > > I read in the Gluster Getting Started leaflet > > (https://lists.gnu.org/archive/html/gluster-devel/2014-01/pdf3IS0tQgBE0.pdf) > > that the max recommended brick size should be 100 TB. > > > > Once my storage server nodes filled up with disks they will have in > > total 192 TB of storage space, does this mean I should create two > > bricks per storage server node? > > > > Note here that these two bricks would still be on the same controller > > so I don't really see the point or advantage of having two 100 TB > > bricks instead of one single brick of 200 TB per node. But maybe > > someone can explain the rational here? > > This is based on the recommendation that RHEL has for maximum size of > XFS filesystems. They might have adjusted the size with more recent > releases, though. > > However, having multiple bricks per server can help with other things > too. Multiple processes (one per brick) could handle more clients at the > same time. Depending on how you configure your RAID for the bricks, you > could possibly reduce the performance loss while a RAID-set gets rebuild > after a disk loss. > > Best practise seems to be to use 12 disks per RAID-set, mostly RAID10 or > RAID6 is advised. > > HTH, > Niels > _______________________________________________ > Gluster-users mailing list > Gluster-users@gluster.org > http://www.gluster.org/mailman/listinfo/gluster-users > ______________________________________________________________________________ BODDEN-KLINIKEN Ribnitz-Damgarten GmbH Sandhufe 2 18311 Ribnitz-Damgarten Telefon: 03821-700-0 Fax: 03821-700-240 E-Mail: i...@bodden-kliniken.de Internet: http://www.bodden-kliniken.de Sitz: Ribnitz-Damgarten, Amtsgericht: Stralsund, HRB 2919, Steuer-Nr.: 081/126/00028 Aufsichtsratsvorsitzende: Carmen Schröter, Geschäftsführer: Dr. Falko Milski Der Inhalt dieser E-Mail ist ausschließlich für den bezeichneten Adressaten bestimmt. Wenn Sie nicht der vorge- sehene Adressat dieser E-Mail oder dessen Vertreter sein sollten, beachten Sie bitte, dass jede Form der Veröf- fentlichung, Vervielfältigung oder Weitergabe des Inhalts dieser E-Mail unzulässig ist. Wir bitten Sie, sofort den Absender zu informieren und die E-Mail zu löschen. Bodden-Kliniken Ribnitz-Damgarten GmbH 2014 *** Virenfrei durch Kerio Mail Server und Sophos Antivirus *** _______________________________________________ Gluster-users mailing list Gluster-users@gluster.org http://www.gluster.org/mailman/listinfo/gluster-users