Hi Steffen, Replies inline. On Mon, May 26, 2008 at 8:01 AM, Steffen Grunewald < steffen.grunewald@xxxxxxxxxx> wrote: > I'm looking into GlusterFS to store data on a cluster (amd64, Debian Etch) > in a distributed way with redundancy (kind of RAID-1, two nodes mirroring > each other). Each node has a dedicated harddisk which can be used for data > storage, and I'm still free to decide which FS to choose. Sure, we advice you to feel comfortable before choosing anything. > > For easier recovery I'd prefer not to split files across nodes (typical > size ~ a few MB). > Yes, we recommend that for filesizes below few hundred MBs. > > Would GlusterFS be suited for such a task? Yes. > > Would it scale to a couple of hundreds of disk "pairs"? Yes. > > Which underlying filesystem should I choose? you have lot of options. ext3, reiserfs, xfs, ZFS (if you use Solaris, on BSD extended attribute support for ZFS is not yet there, hence you may not get RAID-1 functionality of GlusterFS) > > > Since I'm using my own "homegrown" kernel, which modules would I have to > build - is it mandatory to use the patched version of fuse? No, its not mandatory to use GlusterFS patched fuse. Its just recommended because, it has higher I/O performance. But yes, if you want to use GlusterFS, you need FUSE module. If you have Infiniband setup, few IB modules (mainly ib_uverbs) are required to get ''ib-verbs'' tranport to work. GlusterFS has native support for Infiniband userspace verbs (ie, RDMA support) protocol, using which you can gain a lot in performance, for both i/o intensive and low latency applications. > Any suggestions for the stack of translators (and their order therein)? > How to best organize redundancy? (since I don't have it in hardware, and > I can get hold of "missing" files *if* I know their names, this should > be not too hard?) > Unify (afr1 (disk1a, disk2a.. diskNa), afr2 (disk 1b, disk2b...diskNb)... afrN(...)) [1] Hope the above description is understood as a tree graph. Yes, your file resides on the hardware (ie, on backend filesystem) as is. So, you will get the file back, even if you take the disk out. > Any alternatives? (as fas as I know, e.g. Lustre doesn't have redundancy > features...) > No idea. Regards, Amar [1] disks can be mapped to remote servers too. -- Amar Tumballi Gluster/GlusterFS Hacker [bulde on #gluster/irc.gnu.org] http://www.zresearch.com - Commoditizing Super Storage!
