QuantumATK Forum

QuantumATK => General Questions and Answers => Topic started by: yasheng on August 25, 2016, 11:59

Title: processes_per_kpoint in 2016 version
Post by: yasheng on August 25, 2016, 11:59

Hi,

In 2016 version, the default value of processes_per_kpoint=None. what does this setting mean? If we want to run parallel calculations, let say using 64 cpu, what should be the value of processes_per_kpoint ? 

Thank you,
Yasheng

Title: Re: processes_per_kpoint in 2016 version
Post by: Ulrik G. Vej-Hansen on August 25, 2016, 13:18

The value of processes_per_kpoint depends on the number of k-points you use, but in 2016 it is now determined automatically. So unless you find that the default parallelization scheme is sub-optimal for your system and cluster setup, it should not be necessary to manually set processes_per_kpoint.

Title: Re: processes_per_kpoint in 2016 version
Post by: Jess Wellendorff on August 26, 2016, 09:36

Just to elaborate: The processes_per_kpoint parameter is used to define groups of MPI processes, each working on a single k-point. This is useful if you have more processes than k-points in the calculation (16 MPI processes can work together on computing 4 k-points, processes_per_kpoint=4). In ATK 2016, the parameter is automatically set, and it should rarely be changed manually, unless you really know what you are doing.