The GAMPL Procedure

Computational Method: Multithreading

Threading is the organization of computational work into multiple tasks (processing units that can be scheduled by the operating system). Each task is associated with a thread. Multithreading is the concurrent execution of threads. When multithreading is possible, you can realize substantial performance gains compared to the performance you get from sequential (single-threaded) execution.

The number of threads that the GAMPL procedure spawns is determined by the number of CPUs on a machine and can be controlled in the following ways:

  • You can specify the number of CPUs in the CPUCOUNT= SAS system option. For example, if you specify the following statement, the GAMPL procedure determines threading as if it were executing on a system that had four CPUs, regardless of the actual CPU count: 

    options cpucount=4;

  • You can specify the NTHREADS= option in the PERFORMANCE statement to control the number of threads. This specification overrides the CPUCOUNT= system option. Specify NTHREADS=1 to force single-threaded execution.

The GAMPL procedure allocates one thread per CPU by default.

The tasks that are multithreaded by the GAMPL procedure are primarily defined by dividing the data that are processed on a single machine among the threads—that is, the GAMPL procedure implements multithreading through a data-parallel model. For example, if the input data set has 1,000 observations and PROC GAMPL is running with four threads, then 250 observations are associated with each thread. All operations that require access to the data are then multithreaded. These operations include the following:

  • variable levelization

  • effect levelization

  • formation of the initial crossproducts matrix

  • truncated eigendecomposition

  • formation of spline basis expansions

  • objective function calculation

  • gradient calculation

  • Hessian calculation

  • scoring of observations

  • computing predictions for smoothing component plots

In addition, operations on matrices such as sweeps can be multithreaded if the matrices are of sufficient size to realize performance benefits from managing multiple threads for the particular matrix operation.