TURBOMOLE Users Forum
TURBOMOLE Modules => Aoforce and Numforce => Topic started by: martijn on October 16, 2019, 01:06:42 PM
-
Hi,
I am trying to use the raman script in Turbomole 7.3 to calculate the Raman intensities for an open-shell molecule. The vibrational part of the calculation works fine but during the escf part of the calculation to calculate the dynamic polarisability the calculations hangs at:
Nonorthonormal Krylov Space Iteration
total number of roots to be determined: 3
maximum core memory set to 3000 MB,
corresponding to 84 vectors in CAO basis
maximum number of simultaneously treated vectors (including degeneracy): 3
Iteration IRREP Converged Max. Euclidean
roots residual norm
for days. The top command suggests that escf_smp is running full steam but no output beyond this point is produced.
In contrast a stand alone dynamic polarisability escf calculation runs just fine and converges in less than half an hour. What could be the problem?
-
Spoke to soon. While a separate escf calculation runs fine an egrad calculation, that which the raman script calls, seem to suffer from the same problem. So either calculating the derivative of the dynamic polarisability is a couple of orders of magnitude more expensive than calculating the straight dynamic polarisability or there is some problem with this calculation.
-
Hi,
sounds like a technical problem. I have seen similar issues with other tools on newer Intel CPUs - the math library (Intel's MKL) runs into an infinitive loop when trying to start threads. But interestingly only on CPUs with AVX2 or AVX512 support.
You could try to set:
export MKL_ENABLE_INSTRUCTIONS=SSE4.2
and rerun the same job. In case it runs without problems, please contact the Turbomole Support team - the latest Turbomole version should not have this problem any more.
Regards,
Uwe
-
Thanks Uwe! Adding "export MKL_ENABLE_INSTRUCTIONS=SSE4.2", at least for version 7.3 on my computer, didn't change anything the egrad calculation ended up hanging inexactly the same place in the calculation. Going back to 7.01, however, allowed me to successfully complete the calculation. Next I'll try the latest 7.4 version.