Monday Nov 25, 2013

In-Session Parallelism in ODI12c

contributed by Ayush Ganeriwal

In this post we shall discuss the new in-session parallelism introduced in the ODI12c release. ODI12c now comes with the intelligence to concurrently execute part of the mappings that are independent of each other. For example the data load into C$ tables from two disparate data source is done in parallel as these operations are independent of each other and can be done simultaneously. Similarly, such parallelism can be achieved for flow control and target loads too in different use cases. ODI12c automatically identifies the session tasks that can be executed concurrently and generates code for their parallel execution. Users can visually see what part of the mapping would be executed in parallel and if needed this behavior can be changed to execute tasks sequentially if the business logic demands so.

Parallelism Display in Deployment Plan

A typical deployment plan is shown above where the data from a file and relations data store are first loaded into C$ tables in staging area and then joined and loaded into target table. Here different source stores and their corresponding C$ tables, join components and, the target data store are organized in separate blue and yellow boxes which are called execution units and execution unit groups respectively. Execution unit group contains one or more execution units which are independent of each other and can be executed concurrently. Execution unit contains the set of operations that need to be executed serially. In above example the SOURCE_GROUP contains two execution units which are independent of each other and are executed in parallel. The TARGET_GROUP contains only one execution unit indicating absence of parallelism.

How is parallelism achieved within a session?

The session tasks within a step are now generated in a hierarchical manner similar to a load plan steps hierarchy. There are a couple of new type of tasks introduced namely Serial task and Parallel task. These are container tasks which can have one or more child tasks under them. As the name suggests the child tasks under a serial container task would be executed serially whereas the children tasks of parallel container task would be executed in parallel. These container tasks can be nested within each other resulting in multiple levels in the task hierarchy. Depending upon the deployment plan, ODI12c generates the hierarchy of these serial and parallel tasks to achieve in session parallelism. Here is example of a serial and parallel task hierarchy.

Mapping segments candidate for Parallelism

Following below are some of the common mapping parts under which tasks are usually independent of each other and can be performed simultaneously.

1. Loading source data into staging area

2. Loading data into multiple target tables

3. Performing flow control on I$ table

These are some of the common candidate area for parallelism and depending upon the mapping flow there may be other areas that could be executed in parallel. Let’s look at an example of one of the above common parallelism candidates and see how the behavior can be altered by tweaking the deployment plan.

Loading source data into staging area

In this mapping a file and DB table are joined and then loaded in the target table. ODI generates following deployment plan for it so that the sources can be loaded in collector tables in staging are concurrently and then join and target load related tasks performed sequentially.

On executing the mapping with this deployment plan the tasks hierarchy is generated as follows. We can notice here that following set of tasks are organized under parallel container task

a. LKM tasks for dropping and recreating C$ tables for each data sources

      b. LKM task to load data into each of the C$ tables

Suppose we do not want these C$ tables to be loaded in parallelly and rather want them to load sequentially then we can achieve that by simply modifying deployment plan by dragging and dropping one of the execution unit outside of the execution unit group. This would create a separate execution unit group for each of the source dataset forcing all the execution units to run serially.

On executing the task hierarchy of it is generated as follows showing serial execution of all tasks.

Threads configuration for parallel tasks

ODI12c runtime spawns a separate thread for each of the parallel tasks in a session. This effectively means that execution of a highly parallelize mapping may take high number of threads exhausting all system resources. ODI provide two level of configuration to control the number of such parallel threads in the physical agent configuration. One to control the maximum number of threads in the agent and other is to control max thread count within a session. A parallel task would be started only if a new thread can be spawned according to these thread configurations. Therefore, these should be kept at sufficiently large levels.

Connection management for parallel task

Since each of the parallel tasks is executed by a separate thread, they should not work on the common connection for any non-transactional behavior. So for performing the non-transactional operations each of the parallel tasks acquires its own parallel connection from the connection pool. Therefore, for highly parallel mappings the connection pool size should be configured appropriately according to the level of parallelism.

One other point to keep in mind here is that since each parallel task acquires a new auto-commit connection closes it on task completion thus the OnConnect/OnDisconnect tasks would be executed while creating and closing of such connection. Thus you may see multiple OnConnect/OnDisconnect tasks entries for a mapping having parallel tasks.

Another implication of such parallelism is that any parallel tasks cannot rely on any alteration in database connection done by the earlier tasks because each of the parallel tasks would get a new connection which would be unaware of any operations performed by earlier tasks.

For operations performed on a transaction connection the parallel tasks do not have much flexibility and has to perform operations one by one. For such transactional connections synchronization is maintained so that only one parallel task can perform operation on the connection. Thus having parallel tasks participating in such transactions may actually degrade performance due to the extra locking/unlocking for such synchronization.

The take-away

ODI12c allows parallelizing parts of a mapping to achieve extreme performance. It has the intelligence to automatically identify parallelizable parts within a mapping and also provide flexibility to achieve serial behavior for special business needs. Such parallelism would result in an improved performance and for even higher levels of parallelism the connections pool and thread count can be further fine tuned.

For more uptodate Oracle Data Integration news also follow the twitter feeds from @ORCLGoldenGate and @nmadhu2k3.


Learn the latest trends, use cases, product updates, and customer success examples for Oracle's data integration products-- including Oracle Data Integrator, Oracle GoldenGate and Oracle Enterprise Data Quality


« November 2013 »