Runtime Performance - nasa/gunns GitHub Wiki

Here is some benchmark data for the performance of GUNNS networks on various CPU types. This is only for CPU computing, and doesn't include GPU options. See ROSES for a discussion of GPU options and performance.

The data is recorded in this spreadsheet: documents/CPU_Performance_Benchmark_Data.xlsx

The data is output by the sims/SIM_timing_basic Trick simulation, compiled with -g, -O2 options. The simulation runs a basic GUNNS network of typical sparsity (2 to 4 links incident on each node) of various node sizes, and records the average wall clock time to step the network. We typically see the time increase as roughly the 2.8 power of the number of nodes in the network.

There are other factors that affect the timing. Fluid aspect networks are slower than this benchmark data because of the extra fluid properties and transport math it has to compute. Electrical aspect networks are slower because they typically have to update the system multiple times to convergence, per network step. See ROSES#run-time-instrumentation for more information about seeing these trends for your networks.

From the CPU_Performance_Benchmark_Data.xlsx sheet: