[ERROR] LINEAR ALGEBRA driver

[Biglinn]

Dear all,

When performing BSE calculations, I get an erro like:
...
[06.02] BSE solver(s) @q1
=========================
[06.03] Diago Solver @q1
========================
[ERROR] STOP signal received while in[06.03] Diago Solver @q1
[ERROR] LINEAR ALGEBRA driver [SERIAL_HERMITIAN_diagonalization]performing CHEEV
...

I have read several previous posts with similar problems. It seems that this erro is related to compilation of Yambo.

I think it makes sence. Here's a case:
I keep doing calculations serially with just ONE processor. As long as I use several processors for some calculations in parallel, it gives an erro "[ERROR] Writing File ./run_MPI4_OMP1//ndb.dipoles; Variable NOT DEFINED; NetCDF: Parallel operation on file opened for non-parallel access"

This shows that I use some wrong dependencies when compiling. So, what dependencies do I need? like parallel version of NetCDF, or what else?

Thanks!
Zhicheng

[Daniele Varsano]

Dear Zhicheng,

the Diago solver issue and dipoles seems to me different problem.
What version of Yambo are you using?

Best,

Daniele

[Biglinn]

Dear Daniele,

Thanks to your reply!

I use the Yambo 5.2.4. I also think they are two problems. Its just whether Diago solver have dependencies problems like dipoles.

Best,
Zhicheng

[Daniele Varsano]

Dear Zhicheng,

my advice is to update to the yambo release 5.3. Please report if the problem persists.

Best,

Daniele

[Biglinn]

Dear Daniele,

I follow your suggestion to install Yambo-5.3.0 with the same configure options for 5.2.4. But I get the same erro when performing calculations.

The configure options are the following:
./configure --enable-mpi --enable-open-mp --with-fft-libs="-mkl" --with-blas-libs="-mkl" --with-lapack-libs="-mkl" --with-scalapack-libs=mkl --with-hdf5-path=/public/software/mathlib/hdf5/intel/1.12.0 --with-netcdf-path=/public/software/mathlib/netcdf/intel/4.7.4

I find there is one option "--enable-par-linalg". I dont know whether the problem is that I ignore this option.

Best,
Zhicheng

[Biglinn]

Dear Daniele,

I follow your suggestion to install Yambo-5.3.0 with the same configure options for 5.2.4. But I get the same erro when performing calculations.

The configure options are the following:
./configure --enable-mpi --enable-open-mp --with-fft-libs="-mkl" --with-blas-libs="-mkl" --with-lapack-libs="-mkl" --with-scalapack-libs=mkl --with-hdf5-path=/public/software/mathlib/hdf5/intel/1.12.0 --with-netcdf-path=/public/software/mathlib/netcdf/intel/4.7.4

I find there is one option "--enable-par-linalg". I dont know whether the problem is that I ignore this option.

Best,
Zhicheng

Or I I shouldn't use the lapack, NetCDF, HDF5, ect. on the cluster, but install the external libararies?

[Daniele Varsano]

Dear Zicheng,

the libraries on the cluster are fine.
"-enable-par-linalg"" allows using scalapack for parallel linear algebra.
Can you post your input/report files one of the log files and the config.log?

Moreover can you check using "ncdump" if some NAN is present in the BS databse?

Best,

Daniele

[Biglinn]

Dear Daniele,

The compile file is the following, and the input/report file.

Best,
Zhicheng

[Biglinn]

And I use ncdump to check. These are NaN in the file "ndb.BS_PAR_Q1". I attach it as "bs_par.txt".(The file is two large, I only keep the first few lines.)

And I find that the "ndb.em1s_fragment_1" also has many NaN. I attach is as "emls_fragment_1.txt"

[Daniele Varsano]

Dear Biglinn,

ok the error is due to the fact that BSE was not properly calculated as evident by the NaN present already in the screening.

I noticed from your file that you generated the input file with the Lumen fork and used yambo for the calculation, I definitely would try to be consistent.
From the report I can see you are not calculating the BSE matrix or the screening but reading pre-existing databases (which contain NaN), if they were generated with a different version the code, again I would try to be consistent, i.e. calculate them with the same version.

My suggestion is to start the calculation from scratch with the version you compiled, generating also the input file with the same version. If it fails and NaN are still present you can try to recompile the code using the option "--enable-int-linalg" , this will force to compile internal linear algebra libraries, in this way you can check if there is some incompatibility with the libraries compiled in your cluster.

Best,
Daniele