Figure 4.

Benchmark results for the use of CSA in NEST through PyNN, comparing the two CSAConnector implementations explained in section 5 and two of the CSA implementations mentioned in section 3. Color and dash codes are given in the legends. Slope is the ratio of logarithms of the last and first data point shown. Pale lines denote the expected scaling. (A) Shows the run time for connecting a network using CSAs random mask with a probability of 0.1 for different numbers of neurons. This connector creates O(n²) connections for n neurons. The expected slope is thus 2. (B) Shows the same as in (A), but using CSAs oneToOne mask, which creates O(n) connections for n neurons and has an expected slope of 1. (C) Shows a strong scaling experiment, wiring a network of 48,000 neurons using CSAs random mask with a probability of 0.1 and varying the number of MPI processes from 1 to 48. The expected slope is −1, meaning that the run time drops linearly with the number of processes. (D) Shows the results of a weak scaling experiment, increasing the number of connections by approx. 4.8 · 10⁶ per additional MPI process for 1 to 48 processes. The expected slope is 0, as the load increases linearly with the number of processes.