Andrew Savchenko <bircoph <at> gmail.com> writes:
> Ceph is optimized for btrfs by design, it has no configure options > to enable or disable btrfs-related stuff: > https://github.com/ceph/ceph/blob/master/configure.ac > No configure option => no use flag. Good to know; nice script. > Just use the latest (0.80.7 ATM). You may just nerame and rehash > 0.80.5 ebuild (usually this works fine). Or you may stay with > 0.80.5, but with fewer bug fixes. So just download from ceph.com, put it in distfiles and copy-edit ceph-0.80.7 in my /usr/local/portage, or is there an overlay somewhere I missed? > If raid is supposed to be read more frequently than written to, > then my favourite solution is raid-10-f2 (2 far copies, perfectly > fine for 2 disks). This will give you read performance of raid-0 and > robustness of raid-1. Though write i/o will be somewhat slower due > to more seeks. Also it depends on workload: if you'll have a lot of > independent read requests, raid-1 will be fine too. But for large read > i/o from a single or few clients raid-10-f2 is the best imo. Interesting. For now I'm going to stay with simple mirroring. After some time I might migrate to a more agressive FS arrangement, once I have a better idea of the i/o needs. With spark(RDD) on top of mesos, I shooting for mostly "in-memory" usage so i/o is not very heavily used. We'll just have to see how things work out. Last point. I'm using openrc and not systemd, at this time; any ceph issues with openrc, as I do see systemd related items with ceph. > Andrew Savchenko Very good advice. Thanks, James