You can use profiling to determine the stacking number
for your dynamic cluster. The stacking number is the number of cluster
instances that are needed to use all the power of a node.
About this task
The
Allow more than one instance to start on the same
node setting in the administrative console enables vertical stacking
on the node. When you select this option, you must provide the number
of instances to start on the same node. This value is the stacking
number.
You can determine the stacking number by profiling of your
application. If your dynamic cluster is homogeneous, and all of the
nodes for the dynamic cluster have similar power, you
need to profile and determine the stacking number on one node only.
If the member nodes for the dynamic cluster are heterogeneous and
vary in power, you must profile each node and determine the stacking
number for each node.
Procedure
- Start one server instance on the test node.
- View CPU utilization and average
throughput curves for the node. To view the charts, click .
- Increase the load on the node, with the goal of reaching
90-100% node utilization.
- If you cannot reach 90-100% node utilization, add and start
another server instance on the test node. Continue adding server instances
until the node utilization reaches 90-100%.
- The current number of server instances on the test node
is the stacking number. Remember this number so that you can configure
the stacking number in the administrative console.
- If your dynamic cluster has member nodes that are heterogeneous,
repeat these steps for each node and record a stacking number for
each node.
What to do next
Enable vertical stacking for your cluster. With vertical
stacking enabled, the autonomic managers can limit the processor
percentage that is used by each stacked server instance. The general
formula to determine the amount of processor that can be used by a
single instance is: 100%/maximum_number_of_stacked_instances.
For example, if you have three stacked server instances on a node,
the workload is throttled to prevent any single server instance form
using more than 33 percent of the processor.