Benchmarks

Unstructured benchmarks

Receive type comparison

Times and bandwidth for the different receive types which are available in the unstructured case:

• UNORDERED: non-contiguous receive halos + buffered receive;
• ORDERED: contiguous receive halos + buffered receive;
• IN-PLACE RECEIVE: contiguous receive halos + in-place receive.

Benchmark details - CPU

• Triangular mesh (average node connectivity = 6), 8*10^6 vertices;
• 100 vertical layers;
• 1 field exchanged, of type int_64;
• 50 warm-up exchanges, 50 measured exchanges, results are averaged over measured repetitions, threads and MPI ranks; final results are then averaged over 3 independent simulations;
• system: daint-gpu;
• configuration: CPU-only, 1 rank per node, 12 threads per rank, 1 local domain per thread;
• transport layer: MPI;
• thread type: std::thread.

Benchmark details - GPU

• Triangular mesh (average node connectivity = 6), 5*10^7 vertices;
• 100 vertical layers;
• 1 field exchanged, of type int_64;
• 50 warm-up exchanges, 50 measured exchanges, results are averaged over measured repetitions and MPI ranks; final results are then averaged over 30 independent simulations;
• system: daint-gpu;
• configuration: 1 rank per node, 1 thread per rank, 1 local domain per rank;
• transport layer: MPI.

Weak scaling using Atlas meshes and fields

Benchmark details - GPU

• Weak scaling starting from 2 nodes (taken as the baseline);
• Atlas Octahedral Gaussian Grid with two different baseline sizes: either 80 or 160 parallels between the Pole and the Equator;
• 100 vertical layers;
• two different exchange setups: either 1 int field, halo depth 1 or 8 int fields, halo depth 2;
• 1 warm-up exchange, 50 measured exchanges, results are averaged over measured repetitions and MPI ranks; final results are then averaged over 30 independent simulations;
• system: daint-gpu;
• configuration: 1 rank per node, 1 thread per rank, 1 local domain per rank;
• transport layer: MPI.