Performance analysis compares the cost of the same simulation under different execution strategies. The repository's plotting script derives three metrics from the recorded timings:

throughput = data["width"] * data["nsteps"] / data["seconds"] / 1e6
speedup = np.ones(len(data))
 
for width in np.unique(data["width"]):
    serial = data[(data["accel"] == "none") & (data["width"] == width)]
    parallel = (data["accel"] == "openmp") & (data["width"] == width)
    speedup[parallel] = np.median(serial["seconds"]) / data["seconds"][parallel]
 
efficiency = speedup / data["threads"]

Throughput reports work per second. Speedup uses the median serial time for the same grid width, while efficiency divides speedup by the number of threads. These calculations assume a consistent numerical workload, including step count.

Run the benchmark across widths and thread counts. Small grids expose synchronization overhead; large working sets can pressure the memory hierarchy. The curves identify changes in performance, while additional profiling is needed to distinguish their causes.

Source: plot_benchmark.py.