Some of my calculations are taking a long time to run on a single compute node, and I would like to speed them up as much as possible. I am also willing to use semi-empirical methods such as DFT, or even a fragmentation method, if they scale better for my system.
Does GAMESS benefit from running calculations on multiple compute nodes? If so, what methods scale the best across many compute nodes on a cluster?
Scalability and parallelization in GAMESS varies greatly depending on the QM method being used.
For example, the distributed memory MP2 algorithm scales extremely well across many nodes, up to thousands of cores.
On the other hand, most semi-empirical methods in GAMESS are serial algorithms that do not benefit from more than one core.
However, from a speed perspective the serial semi-empirical methods are going to be significantly faster than running the same calculation with MP2.
But both of these methods are applicable to very different types of chemical problems. You really need to give more specifics about the chemistry you are investigating.
Before you start your project, you need to identify the correct level of theory to use based on the chemistry, and then start performing some test calculations and investigating parallelization strategies.
Adding to the spruitt’s answer above:
- Hartree-Fock, DFT, MP2 all will scale quite well
- MCSCF: it depends on orbital optimizer and CI method
- Full CI (ALDET) method does NOT scale beyond one process
GAMESS official reference on parallelization, etc – it digs very deep:
Some notes on GAMESS parallelization with benchmarks (unfortunately graphs are too small)