5 Oracle Streams AQ Performance and Scalability

5.1.1 Oracle Streams AQ and Oracle Real Application Clusters

Real Application Clusters (RAC) can be used to ensure highly available access to queue data. The entry and exit points of a queue, commonly called its tail and head respectively, can be extreme hot spots. Because RAC may not scale well in the presence of hot spots, limit usual access to a queue from one instance only. If an instance failure occurs, then messages managed by the failed instance can be processed immediately by one of the surviving instances.

You can associate RAC instance affinities with 8.1-compatible queue tables. If you are using q1 and q2 in different instances, then you can use ALTER_QUEUE_TABLE or CREATE_QUEUE_TABLE on the queue table and set primary_instance to the appropriate instance_id.

5.1.2 Oracle Streams AQ in a Shared Server Environment

Queue operation scalability is similar to the underlying database operation scalability. If a dequeue operation with wait option is applied, then it does not return until it is successful or the wait period has expired. In a shared server environment, the shared server process is dedicated to the dequeue operation for the duration of the call, including the wait time. The presence of many such processes can cause severe performance and scalability problems and can result in deadlocking the shared server processes. For this reason, Oracle recommends that dequeue requests with wait option be applied using dedicated server processes. This restriction is not enforced.

5.2.1 Using Storage Parameters

Storage parameters can be specified when creating a queue table using the storage_clause parameter. Storage parameters are inherited by other IOTs and tables created with the queue table. The tablespace of the queue table should have sufficient space to accommodate data from all the objects associated with the queue table. With retention specified, the history table as well as the queue table can grow to be quite big.

Oracle recommends you use automatic segment-space management (ASSM). Otherwise initrans, freelists and freelist groups must be tuned for AQ performance under high concurrency.

Increasing PCTFREE will reduce the number of messages in a queue table/IOT block. This will reduce block level contention when there is concurrency.

Storage parameters specified at queue table creation are shared by the queue table, IOTs and indexes. These may be individually altered by an online redefinition using DBMS_REDEFINTION.

5.2.2 I/O Configuration

Because Oracle Streams AQ is very I/O intensive, you will usually need to tune I/O to remove any bottlenecks.

5.2.3 Running Enqueue and Dequeue Processes Concurrently in a Single Queue Table

Some environments must process messages in a constant flow, requiring that enqueue and dequeue processes run concurrently. If the message delivery system has only one queue table and one queue, then all processes must work on the same segment area at the same time. This precludes reasonable performance levels when delivering a high number of messages.

The best number for concurrent processes depends on available system resources. For example, on a four-CPU system, it is reasonable to start with two concurrent enqueue and two concurrent dequeue processes. If the system cannot deliver the wanted number of messages, then use several subscribers for load balancing rather than increasing the number of processes.

Tune the enqueue and dequeue rates on the queue so that in the common case the queue size remains small and bounded. A queue that grows and shrinks considerably will have indexes and IOTs that are out of balance, which will affect performance.

With multi-consumer queues, using several subscribers for load balancing rather than increasing the number of processes will reduce contention. Multiple queue tables may be used garnering horizontal scalability.

5.2.4 Running Enqueue and Dequeue Processes Serially in a Single Queue Table

When enqueue and dequeue processes are running serially, contention on the same data segment is lower than in the case of concurrent processes. The total time taken to deliver messages by the system, however, is longer than when they run concurrently. Increasing the number of processes helps both enqueuing and dequeuing. The message throughput rate is higher for enqueuers than for dequeuers when the number of processes is increased. Usually, the dequeue operations throughput is much less than the enqueue operation (INSERT) throughput, because dequeue operations perform SELECT, DELETE, and UPDATE.

5.2.5 Creating Indexes on a Queue Table

Creating an index on a queue table is useful if you:

• Dequeue using correlation ID

  An index created on the column corr_id of the underlying queue table AQ$_QueueTableName expedites dequeues.

• Dequeue using a condition

  This is like adding the condition to the where-clause for the SELECT on the underlying queue table. An index on QueueTableName expedites performance on this SELECT statement.

5.2.6 Other Tips

• Ensure that statistics are being gathered so that the optimal query plans for retrieving messages are being chosen. By default, queue tables are locked out from automatic gathering of statistics. The recommended use is to gather statistics with a representative queue message load and lock them.

• The queue table indexes and IOTs must be coalesced periodically. In 10.2 with automatic space segment management (ASSM), or an online shrink operation may be used for the same purpose. This reduces queue monitor CPU consumption and ensures optimal enqueue dequeue performance.

• Ensure that there are enough queue monitor processes running to perform the background tasks. The queue monitor must also be running for other crucial background activity. Multiple qmn processes share the load; make sure that there are enough of them. These are auto-tuned, but can be forced to a minimum number, if needed.

• It is recommended that dequeue with a wait time is only used with dedicated server processes. In a shared server environment, the shared server process is dedicated to the dequeue operation for the duration of the call, including the wait time. The presence of many such processes can cause severe performance and scalability problems and can result in deadlocking the shared server processes.

• Long running dequeue transactions worsen dequeue contention on the queue, and must be avoided.

• Dequeue operations are typically slower than enqueue. Overall dequeue/enqueue rates also depend on application design.

• Use NEXT as navigation mode, if not using message priorities. This offers the same semantics but improved performance.

• Use the REMOVE_NODATA dequeue mode if dequeuing in BROWSE mode followed by a REMOVE.