This is to introduce the related components in SUSE Linux Enterprise High Availability Extension product to build High Available Storage (ha-lvm/drbd/iscsi/nfs, clvm, ocfs2, cluster-raid1).

1. High Availability Storage
SUSE's Building Blocks
Roger Zhou
Sr. Eng Manager
zzhou@suse.com
Guoqing Jiang
HA Specialist
gqjiang@suse.com

2. 2
SLE-HA Introduction
The SUSE Linux Enterprise High Availability Extension is powered by Pacemaker & Corosync
to eliminate SPOF for critical data, applications, and services, to implement HA clusters.
TUT88811 - Practical High Availability
Web Site A
Web Site B
Web Server 1
Web Site C
Web Site D
Web Site E
Web Site F
Web Server 2 Web Server 3
Shared StorageFibre Channel Switch
Web Site A Web Site B
Web Server 1
Web Site C
Web Site D
Web Site E
Web Site F
Web Server 2 Web Server 3
Shared StorageFibre Channel Switch

3. 3
Storage
Software Elements
– Block Device: hdX, sdX, vdX => vgX, dmX, mdX
– Filesystem: ext3/4, xfs, btrfs => ocfs2, gfs2
Enterprise Requirements ( must-have )
– High Available (Active-Passive, Active-Active)
– Data Protection (Data Replication)

4. 4
Small Business

5. 5
High Availability Block Device
- active/passive

6. 6
High Availability – DRBD
• DRBD – Data Replication Block Device
• A special master/slave resources is
managed by pacemaker, corosync
software stack
• SLE HA stack manages the service
ordering, dependency, and failover
• Active-Passive approach
Host1 Host2
Virtual IP
Apache/ext4
DRBD Master
Kernel
Virtual IP
Apache/ext4
DRBD Slave
Pacemaker + Corosync
Kernel
Failover

7. 7
High Availability LVM
• HA-LVM resources managed by
pacemaker, corosync software stack
• lvmconf --enable-halvm
• Active-Passive approach
Host1 Host2
Virtual IP
LVM
DRBD Master
Kernel
Virtual IP
LVM
DRBD Slave
Pacemaker + Corosync
Kernel
Failover

8. 8
High Availability iSCSI Server
• iSCSI provides the block devices
over TCP/IP
• Active-Passive approach
Host1 Host2
Virtual IP
ISCSILogicalUnit
iSCSITarget
Kernel
Virtual IP
iSCSILogicalUnit
iSCSITarget
Pacemaker + Corosync
Kernel
Failover
DRBD
Master/LVM
DRBD
Slave/LVM

9. 9
High Availability Filesystem

10. 10
High Availability NFS (HA-NAS)
• NFS server on top of ext3/ext4
• Same applies to:
– xfs
– cifs samba
– etc.
• Active-Passive approach
Host1 Host2
Virtual IP
NFS Exports
Filesystem(ext4)
Kernel
Virtual IP
NFS Exports
Filesystem(ext4)
Pacemaker + Corosync
Kernel
Failover
DRBD
Master/LVM
DRBD
Slave/LVM

11. 11
High Availability NFS (HA-NAS) with Shared Disk
• Active-Passive approach
NOTE: multipathing is “must-have”
for Enterprise
Host1 Host2
Virtual IP
NFS Exports
Kernel
Virtual IP
NFS Exports
Pacemaker + Corosync
Kernel
Failover
Shared
Disk
Filesystem(ext4) Filesystem(ext4)

12. 12
Motivation for Your Storage Growth
Scalability —► extendable storage
Easy Management —► cluster aware components
Cost Effective —► shared resources
Performance —► active – active

13. 13
Concurrent Sharing – you need a lock

14. 14
Clustered LVM2 (cLVM2)
• cLVM2 enables multiple nodes to use
LVM2 on the shared disk
• cLVM2 coordinates LVM2 metadata,
does not coordinate access to the
shared data
• That said, multiple nodes access data
of different dedicated VG are safe
• lvmconf --enable-cluster
• Active-Active approach
Host1 Host2
Pacemaker + Corosync +DLM
LVM2
Metadata
clvmd
Shared LUN 2
Shared LUN 1
LVM2
Metadata
clvmd

15. 15
Clustered FS – OCFS2 on shared disk
• OCFS2 resources is managed by
using pacemaker, corosync, dlm
software stack
• Multiple host can modify the same
file with performance penalty
• DLM lock protects the data access
• Active-Active approach
Host1 Host2
Clustered File System
Kernel
Pacemaker + Corosync + DLM
Kernel
OCFS2 OCFS2
Shared
Disk

16. 16
OCFS2 + cLVM2 (both volumes and data are safe)
• cLVM provides clustered VG. Metadata
is protected
• OCFS2 protect the data access inside
the volume
• Safe to enlarge the filesystem size
• Active-Active approach
Host1 Host2
Clustered File System
Kernel
Pacemaker + Corosync + DLM
Kernel
cLVM2 cLVM2
Shared
Disk
OCFS2 OCFS2

17. 17
Is Data Safe Enough?

18. 18
Data Replication
Data Replication in general
– File Replication (Not the focus of this talk.)
– Block Level Replication
• A great new feature “Clustered MD RAID 1” now is available in in SLE12 HA SP2
– Object Storage Replication
• TUT89016 - SUSE Enterprise Storage: Disaster Recovery with Block Mirroring
Requirements in HA context
– Multiple nodes
– Active-active

19. 19
Data Replication – DRBD
DRBD can be thought as a networked RAID1.
DRBD allows you to create a mirror of two block devices that
are located at two different sites across the network.
It mirrors data in real-time, so its replication occurs
continuously, and works well for long distance replication.
SLE 12 HA SP2 now supports DRBD 9.

20. 20
Data Replication – clvm/cmirrord
● There are different types of LV in CLVM: Snapshot, Striped and
Mirrored etc.
● CLVM has been extended from LVM to support transparent
management of volume groups across the whole cluster.
● For CLVM, we can also created mirrored lv to achieve data
replication, and cmirrord is used to tracks mirror log info
in a cluster.

21. 21
Data Replication – clustered md/raid1
The cluster multi-device (Cluster MD) is a software based RAID
storage solution for a cluster.
The biggest motivation for Cluster MD is that CLVM/cmirrord has
severe performance issue and people are not happy about it.
Cluster MD provides the redundancy of RAID1
mirroring to the cluster. SUSE considers to
support other RAID levels with further evaluation.

22. 22
Data Replication – clustered md/raid1 (cont.)
Internals:
– Cluster MD keeps write-intent-bitmap for each cluster node.
– During "normal" I/O access, we assume the clustered filesystem ensures that only one
node writes to any given block at a time.
– With each node have it's own bitmap, there would be no locking
and no need to keep sync array during normal operation.
– Cluster MD would only handle the bitmaps when
resync/recovery etc happened.

23. 23
Data Replication – clustered md/raid1 (cont.)
It coordinates RAID1 metadata, not coordinate access to the
shared data, and it’s performance is close to native RAID1.
CLVM must send a message to user space, that must be
passed to all nodes. Acknowledgments must return to the
originating node, then to the kernel module.
Some related links:
https://lwn.net/Articles/674085/
http://www.spinics.net/lists/raid/msg47863.html

24. 24
Data Replication – Comparison
Active/Active mode Suitable for Geo Shared Storage
DRBD
Supported
(limited to two nodes)
Yes
No,
storage is dedicated
to each node.
CLVM
(cmirrord)
Supported, the node number
is limited by pacemaker and
corosync
No Yes
Clustered
md/raid1
Supported, the node number
is limited by pacemaker and
corosync
No Yes

25. 25
Data Replication – Performance Comparison
FIO test with sync engine
Read Write
4k
16k
0 500 1000 1500 2000 2500 3000 3500
RawDisk
NativeRaid
Clustermd
Cmirror
Average iops
Blocksize
4k
16k
0 1000 2000 3000 4000 5000 6000 7000 8000
RawDisk
NativeRaid
Clustermd
Cmirror
Average iops
Blocksize

26. 26
Data Replication – Performance Comparison
FIO test with libaio engine
Read Write
4k
16k
0 5000 10000 15000 20000 25000 30000 35000 40000
RawDisk
NativeRaid
Clustermd
Cmirror
Average iops
Blocksize
4k
16k
0 5000 10000 15000 20000 25000
RawDisk
NativeRaid
Clustermd
Cmirror
Average iopsBlocksize

27. 27
Recap

28. 28
Recap: HA Storage Building Blocks
Block-level
HA DRBD
HA LVM2
HA iSCSI
cLVM2
Clustered MD RAID1
Filesystem-level
HA NFS / CIFS
HA EXT3/EXT4/XFS
OCFS2 (GFS2)

29. 29
Question & Answer