Benchmarks for CCSDTQ

Just Got Here

3:04:55 PM PDT - Fri, Jul 6th 2012
Hello, I would like to run some high-accuracy calculations on small molecules (two or three atoms) with up to ~10 electrons. I am considering using the CCSDTQ code available in Nwchem (I am not currently a Nwchem user). I am looking for any time/memory benchmarks for the CCSDTQ code; at the moment I am using the MRCC code by M. Kallay but, 1), it is not parallel, 2), in many situations is very, very input/output time-limited and 3), the scaling with basis set size I empirically observe is higher than what it could be in optimal cases (e.g., time scales as N^10 for a 10-electron system, where N is the basis set size for a given molecule). Thanks!

Clicked A Few Times

12:08:05 PM PDT - Tue, Jul 10th 2012
We don't have the scaling of the CCSDTQ method as a function of the systems size. However, the scaling of the CCSDT approach was pretty consistent with its N^8 numerical complexity. Of course, the performance of the CCSDTQ implementation may be prone to the choice of the parameters which define the granularity of the code (the so-called "tilesize" parameter). This should be observed especially for the small systems. When running the CCSDTQ calculations one should rememebr about the local memory requirements defined by the (tilesize)^8. For example, when you define "tilesize 8" in your input, you will need to store 2 object of the tilesize^8 character which is equavalent to 250 GB. Best, Karol