4:06:03 PM PDT - Sat, May 17th 2014 |
|
Hello,
Is there any alternative to storing 2 electron integrals on disk when calculating MP2 gradients? Ideally, I would like to cache them in physical memory and avoid disk IO altogether (as is possible in the SCF module with directive "semidirect filesize 0 memsize xxxxxx"). If I try to specify "scratchdisk 0" in the mp2 input deck, unsurprisingly, I receive an "insufficient resources" error.
The reason I wish to avoid IO is that I'm running an application on an HPC system that calls several instances of NWChem in parallel. On this particular system, there is abundant physical memory (64 GB/node) but scratch disk IO is quite slow due to the virtual filesystem. I am seeing Walltimes almost double the CPU time in these fragment MP2 gradient calculations. Any suggestions on getting around this bottleneck, or is coding a numerical gradient my best bet here?
Thanks for your advice.
|