topic Re: Re:Illegal instruction (core dumped) when execute HPL in Intel® oneAPI Math Kernel Library

Illegal instruction (core dumped) when execute HPL

jliu41 — Thu, 09 May 2024 04:43:29 GMT

OS: RHEL 9.3 5.14.0-362.8.1.el9_3.x86_64

gcc: 11.4.1 20230605 (Red Hat 11.4.1-2)

oneAPI Base Toolkit: BaseKit_p_2024.1.0.596

HPC Toolkit: HPCKit_p_2024.1.0.560

CPU: Sierra Forrest-SP

HPL version: l_onemklbench_p_2024.1.0_517

Always get Illegal instruction (core dumped) when execute HPL.

Tried to add export MKL_ENABLE_INSTRUCTIONS=AVX2 and -march=native into build.sh script

But got Illegal instruction with compiled binary

core dump msg: kernel: traps: xhpl_intel64_dy[680773] trap invalid opcode ip:7f1eeccc5a77 sp:7ffcb9522ca8 error:0 in libmpi.so.12.0.0[7f1eec800000+b9c000]

Re:Illegal instruction (core dumped) when execute HPL

Gennady_F_Intel — Fri, 10 May 2024 05:47:04 GMT

This is MPI issue. Could you please try to

export I_MPI_COLL_INTRANODE=pt2pt

export I_MPI_SHM = clx_avx2

Re: Re:Illegal instruction (core dumped) when execute HPL

jliu41 — Fri, 10 May 2024 06:11:58 GMT

Yes, the error disappear with these two MPI environment variable.

Will MPI fix this mismatch with CPU flags?

Re: Illegal instruction (core dumped) when execute HPL

Gennady_F_Intel — Fri, 10 May 2024 06:18:45 GMT

The next version of MPI ( 2021.13) will fix all issues.

Re: Illegal instruction (core dumped) when execute HPL

Gennady_F_Intel — Fri, 28 Jun 2024 16:24:07 GMT

Intel MPI ( 2021.13) and oneMKL ( 2024.2) has been released this week an available for download. You could take it and check the problem with these versions.