Special Configuration Instructions

This section describes special configuration instructions for the Cluster Edition.

Increasing Adaptive Server shared memory requirements

Adaptive Server requires an additional 32MB of shared memory than in earlier versions. Increase the value of the configuration parameter 'max memory' to accommodate the additional memory requirement.

Number of preallocated extents

In earlier versions of Adaptive Server, the maximum value allowed for the number of pre-allocated extents was 31. For the Cluster Edition, the maximum values has been increased to pre-allocated extents is 32.Using a value of 32 for the number of pre-allocated extents has a special significance for configuration, and impacts the space allocations Adaptive Server performs internally. If you set the number of pre-allocated extents to 32, Adaptive Server attempts to reserve an entire allocation unit of extents for utility operations that use a large-scale allocation scheme of space reservation, such as bcp in and select into. Using the maximum number of preallocated extents can greatly improve the performance of these utilities, particularly when you run them concurrently on multiple nodes. Using a value of 32 greatly increases the likelihood that each cluster node can work independently on its own allocation unit without interference from the other nodes.

Client failover due to host failure

When the node hosting an Adaptive Server instance fails, clients connected to that instance may not fail over immediately. The TCP driver does not tell the client that the connection has been lost.

The TCP driver uses a mechanism known as keepalive to detect when a connection has been lost. The keepalive configuration varies by operating system, but generally defaults to a a two-hour window during which the disconnect is detected. You may want to tune the TCP keepalive parameter to a smaller value on the client machine to speed the failure detection (see your operating system manual).

You can also avoid keepalive delays by coding client applications to use timeouts. For Open Client (ct-library) applications, this includes using the CS_TIMEOUT and CS_LOGIN_TIMEOUT parameters. See the “Troubleshooting” chapter in the Users Guide to Clusters.

Protect I/O from a nonmember instance in a cluster

Data corruption can happen if an instance leaves a cluster while in-flight buffers on the departing node are still transitioning toward shared disks. To protect against writing to a shared disk, enable I/O fencing by changing the “off” default in the server configuration file to “on”:

sp_configure "enable i/o fencing", 1

For more information about I/O fencing, see the installation guide.