I've finally managed to get this solved using GlusterFS in both boxes. Some things learned in the process:
- First I've tried a generic RAID 1 setup. The main problem with this is that the client always use tcp to contact both servers, even when one of them is in the same machine. So I've to change client configurations to replace the tpc 'local' volume with a direct access (storage/posix) volume
- To avoid stressing the network link, every client read use the local storage with directive
option read-subvolume. Off course to keep the RAID1 integrity GlusterFS always check other volumes as well, but the actual file is retrieved directly from disk
- Performance is good, but client process seems like memory hug. I think is related to quickread volume, I need to investigate further
Modified client configuration:
# Server1 configuration (RAID 1)
option transport-type tcp
option remote-host server2
option transport.socket.nodelay on
option transport.remote-port 6996
option remote-subvolume brick1
option directory /shared
option thread-count 8
option read-subvolume brick-local
subvolumes brick-local server2-tcp
Answering my both questions:
It'll duplicate the files on the same box?
No, the fs is mounted using FUSE. Current /etc/fstab line:
/etc/glusterfs/client.vol /mnt/shared glusterfs defaults 0 0
Instead, should I use the filesystem
directly, locally writing on /shared?
Does the replication work in this way
without mountin a client?
No, always use mounted volumes to make read/writes, using directly the filesystem may lead to inconsistencies.