Stars Application Scenarios

MPI applications use non-blocking collective communication APIs to offload communication tasks to the Stars hardware for execution, freeing up CPU resources for handling other computing tasks. This feature is ideal for large-packet communication scenarios. For Kunpeng 920 72F, it is recommended that Hyper MPI and Stars be used when the number of processes in a single communication domain is fewer than or equal to 1024. The following is an example of the programming paradigm:

MPI_Ibcast(); // Call the asynchronous collective communication API to offload communication tasks to the Stars hardware for execution.
Compute();   // Insert computing tasks into the MPI application and execute the communication tasks in the background.
MPI_Wait();   // Wait until the communication is complete.

When using Stars, specify the Stars collective communication algorithm in the boot options of the mpirun command, for example, -x UCG_PLANC_STARS_IBCAST_ATTR=I:{Algorithm ID}S. For details, see userg_huaweimpi_0031.html#EN-US_TOPIC_0000002533333077__li12956143113403.