Yambo Community Forum

This issue has been resolved, but I do not know how to delete the post.

Dear Xiangjun,

Solved issue should not be deleted. It is useful to all the community to read threads where issues are solved, so better to post the solution more than delete the post.

Best,
Daniele

Daniele Varsano wrote: ↑Fri Feb 28, 2025 10:16 am Dear Xiangjun,

Solved issue should not be deleted. It is useful to all the community to read threads where issues are solved, so better to post the solution more than delete the post.

Best,
Daniele

Dear Daniele,
Thank you for your reminder! In fact, I am not entirely sure where the issue with my code lies. I only managed to get it running successfully by repeatedly adjusting the number of nodes, cores, and parallel parameters. When calculating GW, I sequentially set BndsRnXp to 120, 240, and 360, using one node with 62 cores. The first two attempts ran smoothly, but the calculation unexpectedly terminated when BndsRnXp was set to 360. To avoid memory insufficiency issues, I tried to expand the memory per core by increasing the number of nodes without increasing the number of cores. I made the following multiple attempts:

**Attempt 1:** <---> P10: MPI Cores-Threads : 62(CPU)-2(threads)
<---> P10: MPI Cores-Threads : X_and_IO(environment)-1 1 62 1(CPUs)-q k c v(ROLEs)
<---> P10: MPI Cores-Threads : SE(environment)-1 1 62(CPUs)-q qp b(ROLEs)

**Attempt 2:** <02s> P10: MPI Cores-Threads : 64(CPU)-2(threads)
<02s> P10: MPI Cores-Threads : X_and_IO(environment)-1 1 64 1(CPUs)-q k c v(ROLEs)

**Attempt 3:** <02s> P10: MPI Cores-Threads : 64(CPU)-2(threads)
<02s> P10: MPI Cores-Threads : X_and_IO(environment)-1 1 32 2(CPUs)-q k c v(ROLEs)


The third attempt succeeded. Could you please help me understand what might have gone wrong?

Additionally, is it possible to use the `-J` flag to point to the folder where partial results have already been calculated, in order to avoid redundant calculations after an unexpected interruption?

Thanks!

Yours,
Xiao

Dear Xiao,

in the third attempts you are distributing in more efficient way the memory of the wave functions, this depends on the number of occupied and unoccupied states in your system. In general a good strategy is to have Nv/(cpu-v) ~ Nc/(cpu-c)

where cpu-v and cpu-c are the tasks assigned to "v" and "c".

Best,

Daniele