Matrix Product Toolkit | Main / LargeScaleCalculations

My best guess as to how to do large calculations:

1. Use the trunk/ version

2. Set "TwoStageTruncation = true" in the configuration file, Actually, don't do this, it seems to not work so well.

3. Set "ExtraFrac = 0.2" (this needs more experimentation)

The idea behind the two-stage truncation is to get at least m well-converged density matrix eigenstates, and then keep some extra states that we guess will accelerate the convergence. The actual algorithm is that it keeps at least one state in each quantum number subspace, and adds Steve's single-site density matrix mixing term to get the other states. (This is done in the Hilbert space left after removing the states kept in stage 1). the ExtraFrac is the proportion of extra states to keep in stage 2. That is, if you keep m=1000 states with ExtraFrac = 0.2, then the actual number of states kept will be around 1200. I'm not sure what the optimal number is here. 0.2 seems to work OK, larger numbers might work even better, but will be slower so this is a tradeoff.

My other suggestion is to do a fairly large number of sweeps with a small number of states kept (perhaps 20 - 50 states) before increasing the number of states to the maximum.

The trunk/ version should have a dramatic effect on the file size of the lattice file. The wavefunction files and checkpoint files are already pretty close to optimal.

You should mkae sure to use the LOCAL disk space (ie. /tmp) to store the checkpoint files. The job script can handle copying the checkpoint file to/from the data directory.

My experience has been that increasing the number of Lanczos iterations per step gives an overall speed increase. Perhaps try NumIterations = 8. There is a Davidson solver that should be faster than Lanczos. It does work, but so far I have not been seeing the expected improvement. I suspect there is some problem with it. But if you want to experiment, set "Solver = Davidson" in the configuration file.