Anyone using either Glusterfs or DRBD in their mail setup? How is performance, manageability? Problems? Tips?
Ed W
I tried to use Glusterfs (1.3.7) on mail cluster (qmail + dovecot), but for me it was terrifying mistake. The cluster is serving about 1,5k mail accounts. It was slow, unreliable and unstable. In bonnie++ tests it has nice values, about 65MB/s read and 30MB/s write (storage was ext3 on raid5 on SATA2 drives on Intel E6750), but random seeks was slow (I think) 80 seeks/s. This values are for single client (old one), and storage server could easily handle 3 simultaneous bonnie tests. But when it comes to serve many small files it caused big load on cluster nodes, especially on older ones (P4 Northwood 3,0) causing poor performance. It showed up unstability getting segfaults in io-cache brick or even in client brick causing disconnecting the gluster mount. I was using glusterfs modified fuse module (2.7.2glfs8). With fuse module shipped with 2.6.24 kernel I have some filesystem permissions issues causing reading wrong permissions of files. DRBD should be right choice if you only need failover function. I was using very minimal config and maybe I have misconfigured something. Pawel