Hi, we've successfully configured GlusterFS mirroring across two identical nodes [1]. We're running the file share under a Virtual IP address using UCarp. We have different clients connected using NFS, CIFS and GlusterFS. When we simulate a node failure, by unplugging it, it takes about 5 seconds for the CIFS and GlusterFS clients to refresh the connection and continue operation. The NFS client goes stale. Any connection (like an ongoing copy) with any client breaks and have to be started over. The connection break behavior is to be expected - the TCP connection doesn't handle the switch of host. I didn't expect the NFS client to go stale. Will this behavior be any different with four nodes [2]? Or is there something else I should consider to help the situation? [1] http://www.howtoforge.com/high-availability-storage-with-glusterfs-3.2.x-on-ubuntu-12.04-automatic-file-replication-across-two-storage-servers [2] http://www.misdivision.com/blog/setting-up-a-highly-available-storage-cluster-using-glusterfs-and-ucarp Thank you for your help. -- Best Regards Runar Ingebrigtsen -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://supercolony.gluster.org/pipermail/gluster-users/attachments/20121018/6b0a9cb2/attachment.html>
On Thu, Oct 18, 2012 at 06:48:42PM +0200, Runar Ingebrigtsen wrote:> The connection break behavior is to be expected - the TCP connection > doesn't handle the switch of host. I didn't expect the NFS client to go > stale.I can't answer this directly but I did notice something in the 3.3.1 change log which might be relevant: <http://www.gluster.org/community/documentation/index.php/GlusterFS_3.3.1> "nfs: make NFS filehandles deterministic The NFS3 file handles now includes just the bare minimum of (ident, exportid, gfid) and removes legacy 'variable' members which are unnecessary since the introduction of GFID backend" If you're not on 3.3.1 already, it could be worth trying. Regards, Brian.