Determines the number of logical CPU cores for WFO training (Zorro S
required). Either a positive value for the number of cores to use, or a negative value for the number of cores not to use, f.i. -1 for using all available cores but one. Default = 0 for no multi-core training. The WFO cycles are evenly distributed among parallel Zorro processes assigned to different CPU cores.
The more cores, the faster is the training run; f.i. assigning 5 cores of a 6-core CPU will reduce the training time by 80%.
The current core in a WFO multicore training run, or 0 when no multicore process is running (read/only).
- NumCores must be set in the INITRUN.
- Multicore training uses the command line for passing parameters and script name to the separate training processes. The script name must be command line compliant, i.e. no blanks and no path,
and the script must be located in the Strategy folder. No
other parameters can be passed to the training processes via command line.
- The TESTNOW flag is not compatible to multicore
- Multicore training assumes that sliders or other user interface controls are in the default positions.
- It is possible, but not recommended to assign more cores than CPU threads (logical processors) are available. The training process will then consume most CPU resources and can render the PC unresponsive during the training. The number of physical and logical cores is displayed in the Windows Task Manager.
- Training R-based algorithms works also in
multi-core mode. Several R instances will then run in parallel. The
used R libraries must support parallel processes. That's not the case when they connect to a local or remote server (such as H2O),
or when they use the GPU or other hardware resources (such as Keras, TensorFlow,
or MxNet in GPU mode).
- The processes run as minimized Zorro instances. The message windows display only the cycles trained by the dedicated process.
- Parameter diagrams in multi-core mode are generated from the main process only, not from the parallel processes.
NumCores = -2; // use all logical processors but two
► latest version online