DYNAMICLINEAR Procedure

When you work with SGDLMs, it is crucial to understand and plan for memory requirements, because these requirements scale with the size of the problem. For example, consider modeling a system that contains 400 variables. The adjacency matrix and the precision matrix , which are pivotal for importance sampling and prediction generation, would each have dimensions of , totaling 160,000 elements per matrix. Storing each element in double precision, which is standard practice and requires 8 bytes per element, means that a single matrix consumes approximately 1.28 MB of memory, calculated by 160,000 bytes.

In Monte Carlo simulations that encompass 10,000 iterations, storing all adjacency and precision matrices would require approximately 25.6 GB of memory, which is twice the product of 1.28 MB per matrix and 10,000 iterations, or 12.8 GB multiplied by 2. This figure does not account for the memory consumption during other steps. Adequate memory resources are paramount. If you have insufficient memory, PROC DYNAMICLINEAR cannot run.

If memory constraints present an issue, you might need to reduce the number of variables being modeled, shorten the length of time series, or decrease the number of simulations. Each of these adjustments comes with trade-offs: reducing the number of variables and shortening series length might oversimplify the model, whereas fewer simulations could compromise the statistical robustness of the forecast results. Therefore, it is essential to find an optimal point that meets the memory limitations without significantly compromising the model’s integrity and results.

Generating detailed output from the filtering and forecasting process demands significant memory resources, because these data must be retained in memory prior to being recorded in tables upon completion of the program’s execution. Exercise caution when enabling such output options. For more information, see the document of the FILTER statement and the FORECAST statement.

Currently, an SGDLM that contains approximately 350 variables is estimated to require about 50 GB of memory for processing on CPUs. Through careful design and optimized implementation, the GPU workflow of PROC DYNAMICLINEAR uses memory more efficiently than the CPU workflow. Using single-precision floating-point arithmetic (FP32) for GPU computations further reduces memory consumption while maintaining sufficient accuracy for the SGDLM. For instance, Comparing CPU and GPU Workflows in Simultaneous Graphical Dynamic Linear Models has 300 time series () and 300 observations (), where each variable has 10 elements in its parental set. In that case, FP32 precision requires approximately 7.6 GB of GPU memory, whereas FP64 precision requires approximately 15.04 GB.