Is it reasonable that time cost of diagonalization by scalapack_gvx is 7 times longer than that by genelpa

[pxlxingliang]

Describe the bug

For the below example, the total time of the LCAO SCF calculation by scalapack_gvx is about 3 times longer than that by genelpa, and the longer of diagonalization (hamiltSolvePsiK) is 7 times.

The detail of the time cost is:

The left one is for scalapack_gvx and the right one is for genelpa.

006_Ti15.zip

Expected behavior

No response

To Reproduce

set ks_solver to be scalapack_gvx and genelpa respectively, and run the job with command OMP_NUM_THREADS=1 mpirun -np 16 abacus

Environment

Bohrium image: registry.dp.tech/dptech/abacus:3.1.0

Additional Context

No response

[Alcanderian]

Comparing with this commit
, I found test/performance/P105_si512_lcao takes 20 time slower in the lastest develop branch.

Perhaps the same reason for this issue?

[pxlxingliang]

It seems not a same issue. In this issue the slower is compared to genelpa, and is found in version V3.1.0 (may also exist in earlier versions).