Installation and usage of TURBOMOLE => Parallel Runs => Topic started by: sampsa on February 28, 2012, 12:44:45 PM

Title: ridft and hybrid functionals .. parallel?
Post by: sampsa on February 28, 2012, 12:44:45 PM
Dear Forum,

While trying to run a b3-lyp calculation using ridft, I ran into the following problem:
ERROR: HF-exchange within ridft is not yet implemented in parallel mode.
ridft ended abnormally
I tried mp parallelization, i.e. added to my batch job script the line:


but the scf iteration is, even with 12 threads, painstakingly slow.

Are there any plans to parallelize ridft for hybrid functionals someday?


Title: Re: ridft and hybrid functionals .. parallel?
Post by: antti_karttunen on February 29, 2012, 07:45:57 AM

PARA_ARCH=SMP invokes the "Global array" version of ridft, which is indeed parallelized also for HF exchange. However, as you have noticed, it's significantly slower than pure DFT without HF exchange. You could check the performance with respect to running non-RI hybrid-DFT with dscf. The performance should still be better than here (if I remember correctly, I have used the Global Arrays version only very little). If ridft is slower for your systems than dscf, then it's better to use dscf for hybrid calculations. Maybe additional memory could also speed-up the Global arrays version, so increase $ricore and check that other memory requirements are OK: http://www.cosmologic.de/data/DOK_HTML/node29.html