TURBOMOLE Users Forum

TURBOMOLE Modules => Ricc2 => Topic started by: evgeniy on June 07, 2016, 08:05:26 PM

Title: Good guess for the number of CPU-cores in parallel ricc2 calculations
Post by: evgeniy on June 07, 2016, 08:05:26 PM
Hello,

I wonder if there is a rule of thumb for the good number of CPU-cores in parallel ricc2 calculations.
Thanks!

Best regards,
Evgeniy
Title: Re: Good guess for the number of CPU-cores in parallel ricc2 calculations
Post by: christof.haettig on October 07, 2016, 02:05:05 PM
I'm not sure what you mean with 'good'...

The OpenMP parallelization of ricc2 is typically efficient as long as the number of occupied orbitals is large compared to the number of CPU cores assigned to the calculation.
Large means something like N_occ/N_cores > 10. But you need to decide what has the higher priority for you: high computational efficiency or a short wall time.

Christof
Title: Re: Good guess for the number of CPU-cores in parallel ricc2 calculations
Post by: evgeniy on October 10, 2016, 02:24:29 PM
Dear Christof,

Many thanks for your response. Yes, I meant the ratio "number of occupied orbitals" : "number of CPUs".
Does the ratio for the OpenMP parallelization also hold for the MPI one?

Best regards,
Evgeniy
Title: Re: Good guess for the number of CPU-cores in parallel ricc2 calculations
Post by: christof.haettig on November 17, 2016, 05:14:18 PM
yes. I does.

Also for MPI runs the number of CPU cores used should more than 10 times larger then the number of active occupied orbitals.  With fewer orbitals per CPU core the parallelization efficients becomes soon rather poor.

Christof