Setting up a cluster filesystem

Revision as of 08:52, 31 December 2016 by MartinSchwenke (talk | contribs) (Inconsistent device numbers)

Goal

Set up a clustered file system to be used with CTDB for providing clustered file services.

In addition,

  • How to test if posix locking is supported on the file system?
  • Limitations when using clustered file system

Setting up clustered file system has nothing to do with CTDB. This information is provided for completeness. Users should be aware of any limitations of particular clustered file system.

Cluster file systems

Components

Any cluster file system will have some or all of following components:

  • Shared or distributed storage
  • Kernel or user space file system driver
  • User space file system daemon(s)
  • User space distributed lock manager
  • User space tools for management

Limitations

Every clustered file system has its quirks and limitations. Some of the file system limitations will affect the configuration of file services (Samba or NFS).

  • Does file system provide a consistent view across all the nodes (for example - uniform device and inode numbering) ?
  • Does file system provide posix locking semantics (cluster-aware locking)?
  • Does file system have specific quorum requirements?

Implementation

Each clustered file system example will describe how to set up a clustered file system for 3 node cluster. The implementation can be scaled down to 2 nodes or scaled up to more nodes.

GPFS

GPFS is a proprietary cluster file system from IBM.


GFS2

GFS2 is a clustered file system supported by Red Hat.


Lustre

Lustre file system is an open-source, parallel file system that supports many requirements of leadership class HPC simulation environments.


GlusterFS

GlusterFS is a scalable network file system.


OCFS2

OCFS2 is a general-purpose shared-disk cluster file system for Linux capable of providing both high performance and high availability.

Other cluster filesystems

If you can't find documentation about your choice of cluster filesystem and clustered Samba then you might need to work around some limitations.

Inconsistent device numbers

Note: This section probably wants to be in a future page about cluster filesystems and Samba configuration. It can be moved later...

Locking will not work if a cluster filesystem does not provide unique device numbers across nodes.

Consider the following example:

# onnode all stat /clusterfs/testfile

>> NODE: 10.1.1.1 <<
  File: `/clusterfs/testfile'
  Size: 1286700       Blocks: 2514       IO Block: 65536  regular file
Device: 29h/41d    Inode: 35820037    Links: 1
Access: (0774/-rwxrwxr--)  Uid: ( 3535/     foo)   Gid: (  513/Domain Users)
Access: 2016-11-03 19:51:46.000000000 +0000
Modify: 2016-11-01 13:06:04.000000000 +0000
Change: 2016-11-01 13:06:04.000000000 +0000

>> NODE: 10.1.1.2 <<
  File: `/clusterfs/testfile'
  Size: 1286700       Blocks: 2514       IO Block: 65536  regular file
Device: 29h/41d    Inode: 35820037    Links: 1
Access: (0774/-rwxrwxr--)  Uid: ( 3535/     foo)   Gid: (  513/Domain Users)
Access: 2016-11-03 19:51:46.000000000 +0000
Modify: 2016-11-01 13:06:04.000000000 +0000
Change: 2016-11-01 13:06:04.000000000 +0000

>> NODE: 10.1.1.3 <<
  File: `/clusterfs/testfile'
  Size: 1286700       Blocks: 2514       IO Block: 65536  regular file
Device: 26h/38d    Inode: 35820037    Links: 1
Access: (0774/-rwxrwxr--)  Uid: ( 3535/     foo)   Gid: (  513/Domain Users)
Access: 2016-11-03 19:51:46.000000000 +0000
Modify: 2016-11-01 13:06:04.000000000 +0000
Change: 2016-11-01 13:06:04.000000000 +0000

Note that the device numbers are not consistent across nodes. Locks set for the file on the first 2 nodes will not affect the 3rd node.

To work around this, the following settings should be in the global section of the Samba configuration:

vfs objects = fileid
fileid:mapping = fsname

See vfs_fileid(8) for more information.