concept hardware core in category parallel computing

appears as: hardware core, hardware cores, hardware cores
Parallel and High Performance Computing

This is an excerpt from Manning's book Parallel and High Performance Computing.

For our first try to run our application, we simply ask mpirun to launch 32 processes.

The last piece to control is affinity itself. Affinity is the option of how tightly to bind the process to the hardware resource. The form of the option is similar to the previous options

--bind-to [slot | hwthread | core | socket | numa | node]

The default setting of core is sufficient for most MPI applications (without the --bind-to option the default is socket for greater than two processes as mentioned in section 14.4.1). But there are cases where that affinity setting causes problems. As we saw in the example for figure 14.9, the affinity is set to the two hyperthreads on the hardware core. We might want to try --map-to core --bind-to hwthread to distribute the processes across the cores, but to bind each process more tightly to a single hyperthread. The performance difference from such fine-tuning is probably very small. The greater problem comes when we try to implement a hybrid MPI and OpenMP application. It is important to realize that child processes inherit the affinity settings of their parent. If we use the options of --npersocket 4 --bind-to core and then launch two threads, we have two locations for the threads to run (two hyperthreads per core), so we are okay. If we launch four threads, they will share only two logical processor locations and performance will be limited.