Tuesday, January 10, 2012

Parallel execution in SSIS

Parallel execution in SSIS improves performance on computers that have multiple physical or logical processors. To support parallel execution of different tasks in a package, SSIS uses two properties: MaxConcurrentExecutables and EngineThreads.

If SSIS runs on a dedicated server and you have a lot of operations that run in parallel, you will likely want to increase this setting if some of the operations do a lot of waiting for external systems to reply. On the other hand, if you do not have a dedicated SSIS machine and your data integration application runs alongside several other applications, you may need to reduce this setting to avoid resource conflicts.

The MaxConcurrentExecutables property is a property of the package. This property defines how many tasks can run simultaneously by specifying the maximum number of executables that can execute in parallel per package. The default value is -1, which equates to the number of physical or logical processors plus 2.

Please note that if your box has hyper threading turned on, it is the logical processor rather than the physically present processor that is counted.

The EngineThreads property is a property of each Data Flow task. This property defines how many threads the data flow engine can create and run in parallel. The EngineThreads property applies equally to both the source threads that the data flow engine creates for sources and the worker threads that the engine creates for transformations and destinations. Therefore, setting EngineThreads to 10 means that the engine can create up to ten source threads and up to ten worker threads. The default is 5 in SQL Server 2005 and 10 in SQL Server 2008, with a minimum value of 2.

One thing we want to be clear about EngineThreads is that it governs both source threads (for source components) and work threads (for transformation and destination components). Source threads and work threads are both engine threads created by the Data Flow’s scheduler.

One other thing to consider: If you are using the Execute Package Task, the child package to be executed can be run in-process or out-of-process by use of the ExecuteOutOfProcess property. If a child package is executed out-of-process, you will see another dtshost.exe process start. These processes will remain “live”, using up resources, for quite a while after execution is complete.

If executing in-process, a bug in a task of the child package will cause the master package to fail. Not so if executing out-of-process. On 32-bit systems a process is able to consume up to 2GB of virtual memory. Executing out-of-process means each process can claim its own 2GB portion of virtual memory. Therefore if you are simply using many packages to structure your solution in a more modular fashion, executing in-process is probably the way to go because you don’t have the overhead of launching more processes.

A thread will process one buffer at a time, executing it against all transforms in the execution tree before working on the next buffer in the flow, at which point it would pass the current buffer to another thread executing another execution tree and it would pull a new data buffer from its buffer list which was queued from an upstream component (either a data source or the last asynchronous transform before this execution tree started).

However, the general rule is to not run more threads in parallel than the number of available processors. Running more threads than the number of available processors can hinder performance because of the frequent context-switching between threads.


·         This is a property on the ForEachLoop which says how many instances of the loop contents can be run in parallel.


Suppose we have a package with 3 Data Flow Tasks. Each task has 10 flows in the form of “OLE DB Source -> SQL Server Destination”.

Set MaxConcurrentExecutables to 3, then all 3 Data Flow Tasks will run simultaneously.

Now whether all 10 flows in each individual Data Flow Task get started concurrently is a different story. This is controlled by the second property: EngineThreads.

The EngineThreads is a property of the Data Flow Task that defines how many work threads the scheduler will create and run in parallel. Its default value is 5.  

If we set EngineThreads to 10 on all 3 Data Flow Tasks, then all the 30 flows will start off at once.

1 comment:

  1. This is excellent explanation of the use of Parallel Processing. I am excited about how taking the time to think through and design a project can really improve performance over time. I fight that battle as I see myself building just one Project at a time, I think, this detail is not absolutely necessary. But I know a year down the road and a few dozen projects and I will wish I had thought through this a little more. Thanks again for sharing. Great work.