- BYOM predict function run time:
- Linearly correlated to the size of the data set being scored.
- Linearly correlated to the size of the model.
- Inversely correlated to the system CPU and memory availability.
Teradata recommends pre-testing to estimate the run time before executing a very large data set batch scoring job.
- Model Size Expansion in Memory
Model size (PMML or MOJO file size) increases by a factor that depends on the model when loaded in memory (in cache or outside cache). The maximum factor is 2.7.
For large models, you may need to tune JVM heap space and cache size (see Configuration and Tuning Settings).
For function calls with a high concurrency level—especially for single-row queries—you may need to update Cufconfig parameters related to Java Servers (see Recommended Java Server Settings for Concurrent BYOM Sessions).
If you change any Cufconfig parameters, you must restart the system using tpareset. For more information, see Changing CufConfig Parameter Values.