Binding Specified Threads to Cores

To improve CMF performance, bind specified threads to cores before customizing the L2 I-cache/D-cache separation settings.

When the Qwen 1.5B inference service is deployed, the server load is primarily constrained by the interrupt processing, main process, and Ascend Computing Language (AscendCL) thread modules. Given their varying characteristics, L2 I-cache/D-cache separation settings must be customized for them. Figure 1 shows the core binding policy.

Figure 1 Module-specific core binding of NPU services