Tuning the Cluster Interconnect - Expert Oracle RAC Performance Diagnostics and Tuning - page 442

Chapter 14

Tuning the Cluster Interconnect

In the previous chapter, we discussed that RAC was a configuration of multiple instances clustered together to provide
a scalable, high availability environment for information exchange. These servers communicate with each other using
a private network called the cluster interconnect.
The interconnect provides performance critical direct memory access and message passing between instances
on different servers for transparent consistent access to user data and metadata. The function of the interconnect
is to transport or transfer application (RDBMS and cluster) specific data between instances on different servers.
This suggestion also implies that the interconnect must have a specific configuration to perform its functions well.
An alternative formulation would be applications; middleware communicate with database instances via the
public interconnect, and instances in a tightly coupled RAC cluster communicate with each other via the private
interconnect.
Verification and monitoring of the performance of the interconnect is of primary importance. There are several
important factors that should be considered when configuring the interconnect, and such factors should be verified.
The first and foremost requirement is that it should be a dedicated network connection. The network should at a
minimum be a 1 GigE configuration. The network buffer sizes should be set to the maximum supported by the O/S.

■ Note The reader is requested to consult Chapter 13 before attempting to read this chapter.

Cluster Interconnect
The cluster interconnect must be a high-bandwidth, low-latency communication facility that connects each node
to other nodes in the cluster and routes messages among the nodes. In a RAC environment, the primary transport
mechanism to sharing data blocks between instances is the private interconnect. As illustrated in Figure 14-1 , if a user
on instance one makes a request for a data block, and the block is not available on the local cache of the instance,
then a request is sent to the master of the object. Based on the entries in the GRD, the master will send a request to the
current holder. The entire block is transferred to the requestor. However, if there are a few rows of the blocks that are
locked by the holder in exclusive mode because the user is making a DML operation, the user on the other instance
will see row locks. In certain situations, for example, when the buffer is pinned by the user, the block may not be sent
immediately. We illustrate this using a sequence diagram (see Figure 14-1 ).