bug in BS solve or kernel

[gkedziora]

Dear Yambo Community,

I think I identified a bug, or perhaps it has to do with incorrect usage of the parallel parameters in the BS part without helpful error traps. I'm using the latest svn code in the development branch (revision 94).

I start with solid Si using an 8x8x8 k-point grid from Abinit. I did a previous G0W0 calculation and use it for QP correction. The inputs are attached. The BSE calculation that fails has

BS_ROLEs= "k eh t"
BS_CPU= "32 1 8"

It hangs when calling section() from yambo_driver() with the "Game Over" message. It actually is waiting for half the processors (128 of 256) in at mpi_barrier in PARALLEL_wait. If instead I use

BS_CPU= "32 2 4"

it works ok, I suppose, even though a warning message is produced:

<02s> P0001: [WARNING] n_eh_CPU > 1 in a system with symmetries. This is not efficient, in particular if you have many
kpts.
<02s> P0001: [WARNING] It is suggested to first distribute on "k" and "t". Maximum value for both is: 65

The dielectric function looks reasonable, even though it is not converged, in this case. In the previous case, the dielectric function does not match.

Let me know if I can do anything to help with solving this.

Thanks,
Gary

[Daniele Varsano]

Dear Gary,
thanks for reporting it. We will try to reproduce the problem and let you know.
Best,
Daniele

[Daniele Varsano]

Dear Gary,
at the moment I do not have that number of processors available so I cannot reproduce your error.
Can you tell us if you encounter the same problem when using a total of 64 cpus. If this is the case please report it. In that case I can reproduce the calculation and spot the problem.
Thanks a lot,
Daniele

[gkedziora]

Hi Daniele,

With 64 process, when I used BS_CPU= "16 1 4", the program terminates normally. When I use BS_CPU= "8 1 8", the program hangs after the Haydock solver. Again the results are different. In the attached plot BSE_1f is the case with BS_CPU= "16 1 4".

Best,
Gary

[amolina]

Dear Gary,

can you upload the pseudo file?

Thanks,

Alejandro.