Designing Disaster Tolerant High Availability Clusters, 10th Edition, March 2003 (B7660-90013)

Building a Metropolitan Cluster Using MetroCluster/CA
Preparing an MC/ServiceGuard Cluster for MetroCluster /CA
Chapter 390
Although the risk of this sequence of events taking place is extremely
low, if your business cannot afford even this quite small risk, then you
must enable Fence level = DATA to ensure that the data at the SVOL side
are always consistent.
The disadvantage of enabling Fence level = DATA is that when the CA
link fails, or if the entire remote (SVOL) data center fails, all I/Os will be
refused (to those devices) until the CA link is restored, or manual
intervention is undertaken to split the PVOL from the SVOL.
Applications may fail or may continuously retry the I/Os (depending on
the application) if Fence level = DATA is enabled and the CA link fails.
Fence Level of ASYNC
Fence level = ASYNC is recommended to improve performance in data
replication between the primary and the remote site.
NOTE If Data currency is required on all sides, Fence level = DATA should be
used.
The XP disk array supports asynchronous mode with guaranteed
ordering. When the host does a write I/O to the XP disk array, as soon as
the data is written to cache, the array sends a reply to the host. A copy of
the data with a sequence number is saved in an internal buffer, known
as the side file, for later transmission to the remote XP disk array.
When synchronous replication is used, the primary system cannot
complete a transaction until a message is received acknowledging that
data has been written to the remote site. With asynchronous replication,
the transaction is completed once the data is written to the side file on
the primary system, which allows I/O activity to continue even if the CA
link is temporarily unavailable.
The side file is 30% to 70% of cache (default 50%) that is assigned
through the XP system’s Service Processor (SVP). The high water
mark (HWM) is 30% of the cache; if the quantity of data in the side file
exceeds this value, the write I/O will be delayed to the side file starting
from .5 seconds and increasing to 4 seconds maximum with every 5%
increase over HWM in 500 ms increments.
If the HWM continues to grow, it will eventually hit the side file
threshold (30 to 70% of cache). When this limit has been reached, the
XP on the primary site cannot write to the XP on the secondary site until