Having fun causing my laptop to shutdown trying to run some density functional theory code (CASTEP) on my laptop. Usually firefox is the culprit.

I installed cpufreq tools which seem to make no difference.

Installing thermald seems to help.

The fist time I ran simulations with castep.serial, i.e. compiled to run on one processor core, and it ran fine. To try and speed up calculations I managed to compile using intel ifort and supporting mpi. I actually managed to slow down the calculation the first time trying to run on 4 cores… the i7-620M only has 2 cores, but supports hyper-threading to add an additional 2 virtual cores.

After that I also need to make sure I don’t do a bandstructure calculation using multi-cores again, CASTEP will happily perform the calculation on 4 cores, but the output file informs there is no parallelisation implemented yet for this task.

This command lets me know temperatures, fanspeed, cpu frequency etc.

watch -n 2 -d cpufreq-info -c 0 -fm; cpufreq-info -c 1 -fm; cpufreq-info -c 2 -fm; cpufreq-info -c 3 -fm; sensors


Running castep.serial compiled with gfortran / fftw3 ?

Pseudo atomic calculation performed for Ni 3d8 4s2

Converged in 38 iterations to a total energy of -1355.3473 eV

Charge spilling parameter for spin component 1 = 0.20%
Charge spilling parameter for spin component 2 = 0.27%

Initialisation time = 17.13 s
Calculation time = 2663.64 s
Finalisation time = 0.42 s
Total time = 2681.19 s
Peak Memory Use = 397444 kB


Running castep.mpi compiled with ifort / mkl

Pseudo atomic calculation performed for Ni 3d8 4s2

Converged in 38 iterations to a total energy of -1355.3473 eV

Charge spilling parameter for spin component 1 = 0.20%
Charge spilling parameter for spin component 2 = 0.27%

Initialisation time = 10.33 s
Calculation time = 996.47 s
Finalisation time = 1.26 s
Total time = 1008.06 s
Peak Memory Use = 430140 kB


You will notice I haven’t yet tried mpi with gfortran or serial with ifort. These are just the options I tried so far. Also thermal management was different between these two calculations. But it seems playing around with compilation, thermal management, and using parallelisation can more than double the speed of the calculations.

System details:
3.16.0-4-amd64 #1 SMP Debian 3.16.7-ckt25-2+deb8u3 (2016-07-02) x86_64 GNU/Linux