LinuxConf Europe 2007

Photos and reports

Timetable

Programme

Registration

LPI Exams

Conference Dinner (Sunday)

Duxford Excursion (Monday)

Exhibitors and Sponsors

Accommodation

Venue

Travel

About Cambridge

Kernel Summit 2007

Other GUUG events

Other UKUUG events

Lars Ellenberg - LINBIT

DRBD 8: shared disk semantics on a shared nothing clusterBofB

So, you have an HA-Cluster. Fine. What about when your storage goes down?

Regardless of how you deploy your cluster, and regardless of what services you provide, if the cluster manager decides that some cluster node shall take over the services formerly provided by some other cluster node, it needs access to the data corresponding to the service, including most recent changes.

The Distributed Replicated Block Device (DRBD), as developed mainly by Lars Ellenberg and Philipp Reisner and their team at LINBIT, is a way to have your data highly available, even if a storage node goes down completely.

I could probably come up with some lame excuses as to why DRBD is still an out-of-tree kernel module, but at the time of writing (July 2007) we are starting a major clean-up to prepare for inclusion in mainline.

I'll outline the problems we try to solve; how we deal with it in DRBD; why we deal with it the way we do; explain the design ideas of some algorithms we have developed for DRBD that I think may be useful for software raid or snapshots as well (dm-mirror and friends) and in cluster file systems and distributed databases. Or have we just been ignorant and reinvented the wheel?

I'll provide an overview of typical and not-so-typical usage scenarios, the current limitations and some hints about how you can (ab)use DRBD to your advantage, e.g. when upgrading your hardware.

Finally, I'll mention benchmarks, and upcoming features, such as

• cache warming on the receiving side;
• improved handling of cluster-partition/rejoin
scenarios with multiple-primaries (interesting for cluster file systems);
• the as-yet-only-conceptual possibility of scaling out to many (simultaneously-active) nodes (N > 2).

Submitted paper

and Paper (tgz) .