24 May 2013

432. NWChem 6.3 -- COSMO is now fast(er)!

I probably would've been more excited about this about a year ago when I 'believed' in implicit solvation models (nothing's perfect, and we'll use what is practical so they do fill a strong need. They just aren't very informative for a lot of systems) but it's still a Good Thing.

COSMO has been done using numerical gradients in nwchem 6.1.1 and earlier versions, which has meant that it's been horrendously slow in many cases, in particular if you need to optimise a structure using implicit solvation. COSMO has been -- and still is -- the only implicit solvation model implemented in NWChem, so slow COSMO puts a bit of a spanner in the solvation energy works. Sometimes the calculation even refuses to converge at all.

In contrast, Gaussian has had a number of implicit solvation models implemented, ranging from the quick and dirty PCM, to slower (and better?) C-PCM and I-PCM.

So this is great news.

A quick example:


The test:
Here's a test job (the default cosmo parameters aren't realistic, but this is for testing purposes):
scratch_dir /scratch start benzene geometry units angstroms C 0.100 1.396 0.000 C 1.209 0.698 0.000 C 1.209 -0.698 0.000 C 0.000 -1.396 0.000 C -1.209 -0.698 0.000 C -1.209 0.698 0.000 H 0.000 2.479 0.000 H 2.147 1.240 0.000 H 2.147 -1.240 0.000 H 0.000 -2.479 0.000 H -2.147 -1.240 0.000 H -2.147 1.240 0.000 end basis H library "6-31+g*" c library "6-31+g*" end dft direct end cosmo end scf maxiter 999 end task dft
Note that this is the same test job (plus cosmo, minus optimize) as shown here: http://verahill.blogspot.com.au/2013/05/430-briefly-crude-comparison-of.html

The results:
And here is what I see using nwchem 6.3. (w/ acml 5.3.1, AMD FX 8150/32 gb ram):
6.1.1 19.4 seconds
6.3   14.3 seconds

The difference isn't significant (in the sense that times are too variable so we can't really tell which is faster for such a short job).

But when we change task dft to task dft optimize we get
6.1.1 Fails after 2600 seconds
6.3   128.3 seconds

6.3 churns through the steps pretty efficiently:
@ Step Energy Delta E Gmax Grms Xrms Xmax Walltime @ ---- ---------------- -------- -------- -------- -------- -------- -------- @ 0 -230.09337488 0.0D+00 0.07376 0.01302 0.00000 0.00000 18.1 @ 1 -230.10523734 -1.2D-02 0.00903 0.00231 0.03627 0.10509 45.7 @ 2 -230.10619442 -9.6D-04 0.00491 0.00084 0.01898 0.06082 69.1 @ 3 -230.10628696 -9.3D-05 0.00176 0.00030 0.00737 0.02428 93.3 @ 4 -230.10629787 -1.1D-05 0.00023 0.00005 0.00219 0.00682 115.8 @ 5 -230.10629827 -4.0D-07 0.00004 0.00001 0.00047 0.00136 128.2 @ 5 -230.10629827 -4.0D-07 0.00004 0.00001 0.00047 0.00136 128.2
while 6.1.1 drags itself along for almost an hour:
@ Step Energy Delta E Gmax Grms Xrms Xmax Walltime @ ---- ---------------- -------- -------- -------- -------- -------- -------- @ 0 -230.09389924 0.0D+00 0.07389 0.01306 0.00000 0.00000 691.4 @ 1 -230.10680306 -1.3D-02 0.01081 0.00197 0.03065 0.10438 1378.3 @ 2 -230.10690186 -9.9D-05 0.01000 0.00167 0.00231 0.00803 2092.2
before failing with
6:6:driver: task_gradient failed:: 0 (rank:6 hostname:neon pid:4536):ARMCI DASSERT fail. ../../ga-5-1/armci/src/common/armci.c:ARMCI_Error():208 cond:0 ------------------------------------------------------------------------ There is an error related to the specified geometry ------------------------------------------------------------------------

Sure, the optimization takes 128 seconds instead of ca 44 seconds, but for anyone who's used NWCHEM with COSMO in the past, that's actually not too bad.

I ran another job to get a better feeling for how much longer COSMO vs no COSMO takes for optimization. Optimization of Arecoline (available in ECCE as a fragment) at rb3lyp/6-31+G* takes 2h 5 min with COSMO (33 optimization steps). Without COSMO it takes 37 minutes and uses 14 steps.

No comments:

Post a Comment