 |
» |
|
|
|
The two-data-center architecture is based on a standard Serviceguard configuration
with half of the nodes in one data center, and the other half in
another data center. Nodes can be located in separate data centers
in the same building, or even separate buildings within the limits
of FibreChannel technology. Configurations with two data centers
have the following requirements: There must
be an equal number of nodes (1 or 2) in each data center. In order to maintain cluster
quorum after the loss of an entire data center, you must configure
dual cluster lock disks (one in each data center). Since cluster
lock disks are only supported for up to 4 nodes, the cluster can
contain only 2 or 4 nodes. The Serviceguard Quorum Server cannot
be used in place of dual cluster disks, as the Quorum Server must
reside in a third data center. Therefore, a three data center cluster
is a preferable solution, if dual cluster lock disks cannot be used,
or if the cluster must have more than 4 nodes. To protect against the possibility
of a split cluster inherent when using dual cluster locks, at least
two (three preferred) independent paths between the two data centers
must be used for heartbeat and cluster lock I/O. Specifically, the
path from the first data center to the cluster lock at the second
data center must be different than the path from the second data
center to the cluster lock at the first data center. Preferably,
at least one of the paths for heartbeat traffic should be different
from each of the paths for cluster lock I/O. There can be separate networking
and FibreChannel links between the two data centers, or both networking
and Fibre Channel can go over DWDM links between the two data centers.
See the section below “Network and Data Replication Links
Between the Data Centers” for more detail. FibreChannel Direct Fabric
Attach (DFA) is recommended over FibreChannel Arbitrated loop configurations,
due to the superior performance of DFA, especially as the distance
increases. Therefore Fibre Channel switches are preferred over
Fibre Channel hubs. Any combination of the following
FibreChannel capable disk arrays may be used: HP StorageWorks Virtual
Arrays, HP StorageWorks Disk Array XP, Enterprise Virtual Arrays
(EVA) or EMC Symmetrix Disk Arrays. Refer to the HP 9000
Servers Configuration Guide (available through your
HP representative) for a list of supported FibreChannel hardware. Application data must be
mirrored between the primary data centers. If MirrorDisk/UX is used,
Mirror Write Cache (MWC) must be the Consistency Recovery policy
defined for all mirrored logical volumes. This will allow for resynchronization
of stale extents after a node crash, rather than requiring a full
resynchronization. For SLVM (concurrently activated) volume groups,
Mirror Write Cache must not be defined as the Consistency Recovery
policy for mirrored logical volumes (I.E. NOMWC must be
used). This means that a full resynchronization may be
required for shared volume group mirrors after a node crash, which
can have a significant impact on recovery time. You must ensure
that the mirror copies reside in different data centers, so it is
recommended to configure physical volume groups for the disk devices
in each data center, and to use Group Allocation Policy for all
mirrored logical volumes. Due to the maximum of 3 images
(1 original image plus two mirror copies) allowed in MirrorDisk/UX,
if JBODs are used for application data, only one data center can
contain JBODs while the other data center must contain disk arrays
with hardware mirroring. Note that having three mirro copies will
affect performance on disk writes. VxVM and CVM mirroring does not
have a limit on the number of mirror copies, so it is possible to
have JBODS in both data centers, however increasing the number of
mirror copies may adversely affect performance on disk writes. No routing is allowed for
the networks between data centers. VERITAS Volume Manager (VxVM)
mirroring is supported for distances of up to 100 kilometers for
clusters of 16 nodes. However, on HP-UX 11i v2, VxVM supports up
to 10 kilometers for clusters of 16 nodes. You must ensure that
the mirror copies reside in different data centers and the DRL (Dirty
Region Logging) feature must be used. Raid 5 mirrors are not supported.
It is important to note that the data replication links between
the data centers VxVM can only perform a full resynchronization
(i.e. it cannot perform an incremental synchronization) when recovering
from the failure of a mirror copy or loss of connectivity to a data
center. This can have a significant impact on perfomance and availability
of the cluster if the disk groups are large. VERITAS Cluster Volume Manager
(CVM) mirroring is supported for Serviceguard, Serviceguard OPS
Edition, or Serviceguard Extension for RAC clusters for distances
up to 10 kilometers for 2 or 4 node clusters, and up to 100 kilometers
for 2 node clusters. Since CVM allows only one heartbeat network
to be defined for the cluster, you must make the heartbeat network
highly available, using a standby LAN to provide redundancy for
the heartbeat network. The heartbeat subnet should be a dedicated
network, to ensure that other network traffic will not saturate
the heartbeat network. The CVM Mirror Detachment Policy must be
set to “Global”. For clusters using VERITAS
CVM, only a single heartbeat subnet is supported, so you must have
both Primary and Standby LANs configured for the heartbeat subnet
on all nodes. For SGeRAC clusters, it is recommended to have an
additional network for Oracle RAC cache fusion traffic. It is acceptable
to use a single Standby network to provide backup for both the heartbeat
network and the RAC cache fusion network, however it can only provide
failover capability for one of these networks at a time.
Two
Data Center FibreChannel Implementations | |
In a two data center configuration, shown in Figure 2-1 “Two
Data Centers with FibreChannel Hubs”, you must use a cluster lock disk,
which is only supported for up to 4 nodes. This configuration can
be implemented using any HP-supported FibreChannel devices. Disks
must be available from all nodes using redundant links. Not all
links are shown in Figure 2-1 “Two
Data Centers with FibreChannel Hubs”. The two cluster lock disks should be located on separate FibreChannel loops
to guard against single point of failure. The lock
disks can also be used as data disks. They must be connected to
all nodes using redundant links (not all links are shown in Figure 2-1 “Two
Data Centers with FibreChannel Hubs”). Nodes can connect to disks in the same data center using short
wave ports, and hubs can connect between data centers using long-wave
ports. This gives you a maximum distance of 10 kilometers between
data centers, making it possible to locate data centers in different
buildings. Advantages
and Disadvantages of a Two-Data-Center Architecture | |
The advantages of a two-data-center architecture are: Only two data centers are needed, meaning less space
and less coordination between operations staff. No arbitrator nodes are needed. All systems are connected to both copies of data,
so that if a primary disk fails but the primary system stays up,
there is greater availabilty because there is no package failover.
The disadvantages of a two-data-center architecture are: There is a slight chance of
split brain syndrome. Because there are two cluster lock disks,
you would get split brain syndrome if the following occurred simultaneously: The chances are slight, however these events happening at
the same time would result in split brain syndrome and probable
data inconsistency. Planning different physical routes for both
network and data connections or adequately protecting the physical
routes greatly reduces the possibility of split brain syndrome. Software mirroring increases CPU overhead. The cluster must be either two or four nodes with
cluster lock disks. Larger clusters are not supported due to cluster
lock requirements. Although it is a low cost solution, it does require
some additional cost: FibreChannel links are required for
both local and remote connectivity. All systems must be connected to multiple copies
of the data and to both cluster lock disks.
|
Printable version
