Monday, August 15, 2005

Xen process migration, filesystem options

Last weekend I was at UKUUG's Linux 2005 event and, amongst other things, learned a bit about Xen. Most impressive is its process migration (need to replace the hardware that a Quake3 server is running on? while players are using it? no problem; in the tested system, the service interruption was just 60ms, which client prediction can ordinarily be expected to cope with anyway).

A pre-condition for migration is access to the same filesystem(s). Amongst the options are SAN devices, NFS/CIFS (but remember to include /!) and some interesting filesystems/drivers that I'd not encountered before:
  • DRBD, which appears to perform RAID over a network, and can handle master/copy switchovers on the fly.
  • Lustre, which provides a DFS for giant clusters (up to tens of thousands of nodes) where not only is user/authentication seperated from file storage, but so is file metadata storage, the latter being on a seperate cluster built just for this purpose. The data itself is on whatever devices are in use for storage, but those devices don't mediate sessions, handle coherence, etc., they're "just" dumb storage, all the smarts are in the metadata cluster.