Results

The benchmark which was published in the supercell paper (PbSnTe₂ 2x2x2) is obsolete. The current version of supercell program can process the configuration in a few seconds. More importantly, this test does nothing useful. Therefore, below I use a more advanced test case with 10¹¹ total combinations, which does a “wet-run” and generates output structures with random sampling and the lowest electrostatic energy. This is a real practical case of using the supercell program. The input file can be downloaded from supercell distribution CaAl6Te10.cif.

Test command

supercell -i CaAl6Te10.cif -m -q -n r10 -n l10 -v 2

The tests were run once for each of the configurations below:

• Intel® Xeon® Gold 6226R 2.90GHz CPU (32 cores).
• Ubuntu 16.04 LTS
• Binary files from supercell site.

During this run 10¹¹ structures were processed in 117 and 6 minutes on 1 and 32 cores respectively. The maximum performance is around 300 mln structures per second.

Scalability

Some tests were performed to check the scalability of the program (see figure below). The y-axis shows performance per core normalized to v2.1 performance on one core. The extrapolation shows that even at 64 cores adding more cores still decreases processing time, because per-core performance is still above the saturation limit.

Optimal performance recommendations

• Use Linux. Binaries for other platforms are not as well optimized as the Linux binary.
• Optimal compiling of supercell program requires advanced programming skills, so binaries from the site are OK for most users.
• Supercell performance correlates very well with benchmark scores from cpubenchmark.net. You can use that data to find an optimal solution for running the program.