Hi Derek,
that graph is indeed obsolete, sorry about that.
From a principal perspective, I would run the scaling tests on the systems that are of interest to you! Even if ATK scales fantastically for carbon nanotubes, it doesn't really matter, if you are going to study metallic interfaces...
Generally speaking, ATK scales better in parallel for two-probes with many k-points in the transverse plane, or medium-sized bulk systems with many k-points. The scaling is also better for "long" systems compared to compact ones, since the long system can be partitioned in diagonal blocks more efficiently.
And yet, to be a bit more constructive than that...! I was able to dig up the original system used in the scaling test reported in the aforementioned figure. The systems are attached in a zip file as VNL files. If you give a day or two, I'll convert them to full, ready-to-run scripts with good parameters for convergence.
The systems are (I may skip some later, but here they all are, for now):
- A molecular cluster with 30 water atoms
- A C60 buckyball
- A bulk cluster with 63 Si atoms and one B doping atom
- A small bulk perovskite
- A (12,0) CNT (bulk)
- A (5,5) CNT (bulk)
- Li-DTB-Li two-probe
- A two-probe with a (5,5) nanotube with an N-O adsorbant
In addition, please find attached also a recent scaling test performed in Japan, on the Tsubame supercluster. For this quite typical two-probe system, quite nice scaling is observed even up to 32 cores for the SCF part, and all the way up to 128 cores for "analysis" (transmission spectrum). Note how going from 64 nodes to 128 cuts the "physical properties" calculation almost straight in half, and it's 100 times faster than for 1 node. It's nice if you can get the results in 4-5 hours instead of 4-5 days!!!
Thanks for the bash path hint! There are of course many such little tricks that can make life a lot easier. Just remember to update the path statement when you install a new version of ATK!
