Speaker Andrew Tridgell & Ronnie Sahlberg, Samba team
Time 2008-01-31 13:30
Conference LCA2008

Last year: Do no use this, only works in demonstrations.

Now: in use right now, television stations using clustered samba for serving TV shows.

Fastest NAS box around. Multi-gigabytes per second.

Use it in production!

All the bugs are Ronnie’s fault.

Andrew Bartlett.

Clustering backported to Samba 3.0 (stable version) from Samba 4.0 tree (bleeding edge).

CTDB

  • lightweight clustered database
  • people clustered Samba for years, badly, negative scaling.

1 node cluster 100Mbs
2 node cluster 2Mbs

Now gets faster as you add extra nodes.

Demo, how to install and use.

  • 4 linux blades
  • GPFS filesystem, 12 raid arrays.
  • Cluster is in German, not very portable.

Will demonstrate:

  • fast IP failover
  • snapshot exposure to windows clients
  • offline file handling
  • online software upgrade - live upgrade while people using it
  • crash resiliance - reset nodes that are in use

Multipath daemon, have had problems. Will redirect traffic to inactive node.

Windows XP box, you can boo and hiss now.

Single name on network, large files on file system.

Very large filesystems supported.

One name for server, not the individual node names.

Working NAS box, user connected to node 2. What happens if we kill node?

“File creation error” on user.

Hmmm. Lets try that again…

It paused a long time.

Users should just transfer to working node. Application needs to behave itself. Application has to be aware of fail over, and need to try and reconnect.

Under Vista, SMB2 level, fail over will be done by the kernel. Will be able to do full implementation, with documentation from Microsoft. Looking forward to a full implementation of the protocol.

Will we be feeding errors in the documentation back to Microsoft? Yes, part of the agreement is that Microsoft has a team of people to answer questions.

90s Microsoft would answer questions, etc. Then the lawyers got in the way. Now we are talking to the engineers again, and they like talking.

NFS has had this for some time.

SMB does not have automatic reconnect.

Linux implementation of CIFS in kernel does have automatic fail over.

Not a limitation of the protocol, although SMB2 has features that make is easier.

File system advanced file system. Snapshots. Make it easy for users to get back there old files. shadow copy 2 module. Samba knows how to export these so windows will see them as different versions. Part of GPFS features.

offline (tape) vs online (disk) storage. clock icon on windows, means file is offline. Via Tivoli storage manager. Server can still do stuff simultanously.

Hard shut down. TCP problem with fail over.

Hard reset of CPUS, while copying files, use following command:

echo b > /proc/sysrq-trigger

Pause, because client has to be sure server is dead, and other nodes need to be sure client is really dead.

telnet to a port on a server, and that server goes offline (somebody trips over power cord). Assume client trying to read at the time. It will exit immediately on a write though. What happens to client waiting for time out? It times out. Default is 2 hours. Fail over needs to be fast.

We need to tell client to reconnect, and do it very very quickly.

Telnet to ssh port on node, and samba port 445.

ssh port hangs, samba port 445 did get a disconnect.

Tickle act. We distribute across all nodes details of all sockets on all node. When another node takes over, it sends an ACK with an invalid sequence number to the client. The client will send an ACK when it receives an out of window ACK. That ACK hits the new node, which sends a reset with the real sequence number.

Client continues even though power was lost to node.

Scaling NAS

  • 30,000 NAS users
  • 100 NAS servers + 100 backup servers
  • every day, one of them runs out of disk space.

Solution? Create a really big NAS box, with backups, no space management in dividing space up between NAS boxes required.

Usual situation, primary server and hot spare. Hot spare turned off or disabled, and comes up if the primary goes down. The problem is if the backup doesn’t take over as required.

Active/passive failover, hope the passive system will start up, when it was powered of for several months.

Don’t find out failover has failed until after it has failed, and people are screaming.

Other problem, what if the passive system wakes up when the active system is still alive. Or if the active system wakes up after the passive system is alive.

With shared read/write partition, data will be lost within seconds.

Clustered samba doesn’t failover, alternative node is already up and running, and known to be good. Allows you to sleep much better at night.

All active system. Better performance.

All-active NAS.

Protocol hooks.

Event hooks.

Rewrite heartbeat; heartbeat is more orientated with active-passive. Didn’t work well with fast fail over. Can turn off HA features in heartbeat and use database features. Heartbeat 2 might be more friendly. We use single file and integrated. Merging with heartbeat might be a possibility in the future.

1 nodes 109 Mbytes/second
2 nodes 201 Mbytes/second
3 nodes 278 Mbytes/second
4 nodes 308 Mbytes/second

Want to try it?

Requirements:

  • Linux cluster
  • Set of really fast disks on a SAN
  • Clustered filesystem, most testing on GPFS. GFS and GFS2 and Lustre work.
  • Samba from http://ctdb.samba.org/, aim is for cluster support out of the box.

IBM supports it as SOFS.

Minimum config:

clustering = yes idmap backend = tdb2 (more efficent)

/etc/ctdb/nodes
/etc/ctdb/public_addresses (if not using failover stuff)
/etc/sysconfig/ctdb (optional additional parameters)

filesystem specific options, file id mapping.

Show your managers. Flash movies: http://samba.org/~tridge/ctdb_movies