On computer systems that have more than one central processing unit, multi-threadingcan be used to increase the speed of numeric calculations by using multiple system processors to simultaneously carry out different parts of the computation. In a multithreaded environment, each thread handles a portion of the overall task; if several threads can run in parallel, the computation can often be completed more quickly than if the different portions of the task ran in series.

IDL’s thread pool, a pool of computation threads that are used as helpers to accelerate numerical computations, allows for multithreading when multiple CPUs are present. IDL automatically evaluates all computations performed by routines that may benefit from multithreading to determine whether or not to use the thread pool in the current computation. This decision is based on attributes such as the number of data elements involved, the availability of multiple CPUs, and the availability of a multithreaded implementation of the algorithm in use. You can alter the parameters used by IDL to make this decision, either on a global basis for the duration of a single IDL session, or for an individual computation.

Note: Multithreading does not offer the possibility of increased execution speed for all IDL routines. For a list of the routines that have been implemented to use multithreading when possible, see Routines that Use the Thread Pool.

Benefits of the IDL Thread Pool

The IDL thread pool will increase processing performance on certain computations. When not involved in a calculation, the threads in the thread pool are inactive and consume little in the way of system resources. When IDL encounters a computation that can use the thread pool and which would benefit from parallel execution, it divides the task into sub-parts for each thread, enables the thread pool to do the computation, waits until the thread pool completes, and then continues. Other than the improved performance, the end result is virtually indistinguishable when compared to the same computation performed in the standard single-threaded manner.

Possible Drawbacks to Using the IDL Thread Pool

There are instances when allowing IDL to use its default thread pool settings can lead to undesired results. In some instances, a multithreaded implementation using the thread pool may actually take longer to complete a given job than a single-threaded implementation. If a computation uses the thread pool in an inappropriate situation, there may be other undesirable effects. The following are some situations in which the default thread pool settings may provide less than optimal results.

Computation of a Relatively Small Number of Data Elements

Use of the IDL thread pool requires a small fixed overhead when compared to a non-threaded version of the same computation. Normally, computational speed increases when multiple CPUs work in parallel, and the speed-up is much larger than the loss due to thread pool overhead. However, if the computation does not include a large enough number of data elements (each element being a data value of a particular data type), the loss due to thread pool overhead can exceed the benefit and the overall computation speed can be slower.

To prevent using the thread pool for computations that involve too few data elements, IDL supports a minimum threshold value for thread pool computations. The minimum threshold value is contained in the TPOOL_MIN_ELTS field of the !CPU system variable. See the following sections for details on modifying this value.

Large Computation that Requires Virtual Memory Use

If a computation is too large to fit into physical memory, the threads in the thread pool may cause page faults that will activate the virtual memory system. If more than one thread encounters this situation simultaneously, the threads will compete with each other for access to memory and performance will fall below that of a single-threaded approach to the computation.

To prevent using the thread pool for computations that involve too many data elements, IDL supports a maximum threshold value for thread pool computations. The maximum threshold value is contained in the TPOOL_MAX_ELTS field of the !CPU system variable. See the following sections for details on modifying this value.

Multiple Users Competing for CPU Resources

On a large multi-user system, an IDL application that uses the thread pool may consume all available CPUs, thus affecting other users of the system by reducing overall performance.

To prevent using all system processors by routines that use the thread pool, IDL allows you to specify explicitly the number of CPUs that should be used in calculations that involve the thread pool. The number of processors to be used for thread pool operations is contained in the TPOOL_NTHREADS field of the !CPU system variable. See the following sections for details on modifying this value.

Note: To change the default number of threads used by IDL, set the IDL_CPU_TPOOL_NTHREADS preference. For more information, see !CPU Settings Preferences.

Sensitivity to Numerical Precision

Algorithms that are sensitive to the order of operations may produce different results when performed by the thread pool. Such results are due to the use of finite precision floating point types, and are equally correct within the precision of the data type.

See Also

Controlling the IDL Thread Pool, Routines that Use the Thread PoolThread Pool Keywords