SIwave-PSI Best Practices: Number of Cores

Minimum: 8

Recommended: 12-16

All of the computationally intensive parts of the solver are shared memory multi-threaded, and performance is strongly dependent on the number of cores. Best performance is seen with 12 to 16 shared memory cores per node. Using fewer than 8 cores impacts performance negatively.

Distribution of frequency points over a large distributed compute cluster can drastically reduce overall simulation time since each frequency point can be solved independently of another. Frequency point solve time scales almost linearly; in other words, scaling up to 512, 1024, … cores is efficient as long as discrete frequency points are defined in the frequency sweep setup and more frequency points exist than cores.