One of the major features of the batch framework is the ability to support multi-threading. The multi-threading support allows a site to increase throughput on an individual batch job by splitting the total workload across multiple individual threads. This means each thread has fine level control over a segment of the total data volume at any time.
The idea behind the threading is based upon the notion that "many hands make light work". Each thread takes a segment of data in parallel and operates on that smaller set. The object identifier allocation algorithm built into the product randomly assigns keys to help ensure an even distribution of the numbers of records across the threads and to minimize resource and lock contention.
The best way to visualize the concept of threading is to use a "pie" analogy. Imagine the total workset for a batch job is a "pie". If you split that pie into equal sized segments, each segment would represent an individual thread.
The concept of threading has advantages and disadvantages:
- Smaller elapsed runtimes - Jobs that are multi-threaded finish earlier than jobs that are single threaded. With smaller amounts of work to do, jobs with threading will finish earlier.
Note: The elapsed runtime of the threads is rarely proportional to the number of threads executed. Even though contention is minimized, some contention does exist for resources which can adversely affect runtime.
- Threads can be managed individually – Each thread can be started individually and can also be restarted individually in case of failure. If you need to rerun thread X then that is the only thread that needs to be resubmitted.
- Threading can be somewhat dynamic – The number of threads that are run on any instance can be varied as the thread number and thread limit are parameters passed to the job at runtime. They can also be configured using the configuration files outlined in this document and the relevant manuals.
Note: Threading is not dynamic after the job has been submitted
- Failure risk due to data issues with threading is reduced – As mentioned earlier individual threads can be restarted in case of failure. This limits the risk to the total job if there is a data issue with a particular thread or a group of threads.
- Number of threads is not infinite – As with any resource there is a theoretical limit. While the thread limit can be up to 1000 threads, the number of threads you can physically execute will be limited by the CPU and IO resources available to the job at execution time.
Theoretically with the objects identifiers evenly spread across the threads the elapsed runtime for the threads should all be the same. In other words, when executing in multiple threads theoretically all the threads should finish at the same time. Whilst this is possible, it is also possible that individual threads may take longer than other threads for the following reasons:
- Workloads within the threads are not always the same - Whilst each thread is operating on the roughly the same amounts of objects, the amount of processing for each object is not always the same. For example, an account may have a more complex rate which requires more processing or a meter has a complex amount of configuration to process. If a thread has a higher proportion of objects with complex processing it will take longer than a thread with simple processing. The amount of processing is dependent on the configuration of the individual data for the job.
- Data may be skewed – Even though the object identifier generation algorithm attempts to spread the object identifiers across threads there are some jobs that use additional factors to select records for processing. If any of those factors exhibit any data skew then certain threads may finish later. For example, if more accounts are allocated to a particular part of a schedule then threads in that schedule may finish later than other threads executed.
Threading is important to the success of individual jobs. For more guidelines and techniques for optimizing threading refer to Multi-Threading Guidelines in the Batch Best Practices for Oracle Utilities Application Framework based products (Doc Id: 836362.1) whitepaper available from My Oracle Support