Input matrix contains Inf or NaN on Large Eigendecomposition

[rlangefe]

Description
I have been working with a slightly modified version of the eigenvalues demo script (included here). This code runs successfully (no errors, correct output as far as I can tell) when I run using up to a 10k x 10k matrix on a 8 node, 8 MPI process per node, 4 CPUs per MPI process grid. When I scale to a 100k x 100k matrix on a 10 node, 5 MPI process per node, 4 CPUs per MPI process grid, I get the error below (printed once for each of the 50 MPI processes):

terminate called after throwing an instance of 'std::domain_error'
  what():  Input matrix contains Inf or NaN
[r6401:3333756] *** Process received signal ***
[r6401:3333756] Signal: Aborted (6)
[r6401:3333756] Signal code:  (-6)
[r6401:3333758] [ 0] /lib/x86_64-linux-gnu/libc.so.6(+0x43090)[0x7f59276f9090]
[r6401:3333758] [ 1] /lib/x86_64-linux-gnu/libc.so.6(gsignal+0xcb)[0x7f59276f900b]
[r6401:3333758] [ 2] /lib/x86_64-linux-gnu/libc.so.6(abort+0x12b)[0x7f59276d8859]
[r6401:3333758] [ 3] /usr/lib/x86_64-linux-gnu/libstdc++.so.6(+0xa4ee6)[0x7f59279c2ee6]
[r6401:3333758] [ 4] /usr/lib/x86_64-linux-gnu/libstdc++.so.6(+0xb6f8c)[0x7f59279d4f8c]
[r6401:3333758] [ 5] /usr/lib/x86_64-linux-gnu/libstdc++.so.6(+0xb6ff7)[0x7f59279d4ff7]
[r6401:3333758] [ 6] /usr/lib/x86_64-linux-gnu/libstdc++.so.6(+0xb7258)[0x7f59279d5258]
[r6401:3333758] [ 7] /path/to/spack/install/spack/opt/spack/linux-ubuntu20.04-haswell/gcc-9.4.0/slate-2023.11.05-5ysppah/lib/libslate.so(_ZN5slate5stedcIfEEvRSt6vectorIT_SaIS2_EES5_RNS_6MatrixIS2_EERKSt3mapINS_6OptionENS_11OptionValueESt4lessISA_ESaISt4pairIKSA_SB_EEE+0x4e5)[0x7f5929c00935]
[r6403:3333758] [ 9] fast_eigenvals(+0xbecd)[0x563f5b231ecd]
[r6403:3333758] [10] fast_eigenvals(+0x5876)[0x563f5b22b876]
[r6403:3333758] [11] /lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0xf3)[0x7f36cef0e083]
[r6403:3333758] [12] fast_eigenvals(+0x59ee)[0x563f5b22b9ee]
[r6403:3333758] *** End of error message ***

At the end, I then see:

--------------------------------------------------------------------------
mpirun noticed that process rank 27 with PID 662523 on node r6406 exited on signal 6 (Aborted).
--------------------------------------------------------------------------
Command exited with non-zero status 134

I trimmed this because this message is just repeated over and over, but if helpful, I can give the whole output.

Steps To Reproduce

1. Compile the script (linking BLASpp, LAPACKpp, Intel MKL, and SLATE. I also use the unmodified utils.hh file from the examples SLATE tutorial (Tutorial Link).
2. Batch the script using OpenMPI.

C++ Code

// Solve Hermitian eigenvalues A = X Lambda X^H
#include <slate/slate.hh>

#include "util.hh"

#include <string>

// Include I/O
#include <iostream>
#include <fstream>
#include <unistd.h>

int mpi_size = 0;
int mpi_rank = 0;

//------------------------------------------------------------------------------
template <typename T>
void hermitian_eig(std::string& output_file)
{
    using real_t = blas::real_type<T>;

    print_func( mpi_rank );
    
    // Print hostname
    char hostname[256];
    gethostname( hostname, 256 );
    printf( "Rank %d of %d on %s\n", mpi_rank, mpi_size, hostname );

    // TODO: failing if n not divisible by nb?
    int64_t n=100000, nb=500, p=10, q=5;
    
    // Maybe can remove this
    assert(mpi_size == p*q);

    // Rank 0 prints out the parameters
    if (mpi_rank == 0) {
        printf( "Parameters: n %ld, nb %ld, p %ld, q %ld\n", n, nb, p, q );
    }

    //slate::HermitianMatrix<T> A( slate::Uplo::Lower, n, nb, p, q, MPI_COMM_WORLD );
    slate::SymmetricMatrix<T> A( slate::Uplo::Lower, n, nb, p, q, MPI_COMM_WORLD );
    A.insertLocalTiles();
    random_matrix( A );

    std::vector<real_t> Lambda( n );

    // A = X Lambda X^H, eigenvalues only
    // slate::eig_vals( A, Lambda );  // simplified API

    //slate::Matrix<T> Xempty;
    //slate::heev( slate::Job::NoVec, A, Lambda, Xempty );

    slate::Matrix<T> X( n, n, nb, p, q, MPI_COMM_WORLD );
    X.insertLocalTiles();
    slate::eig( A, Lambda, X );   // simplified API

    // // Write to file as csv
    // if (mpi_rank == 0) {
    //     std::ofstream file(output_file);
    //     for (int i = 0; i < n; i++) {
    //         file << Lambda[i] << "\n";
    //     }
    //     file << std::endl;
    //     file.close();
    // }
}

//------------------------------------------------------------------------------
int main( int argc, char** argv )
{   

    int provided = 0;
    int err = MPI_Init_thread( &argc, &argv, MPI_THREAD_MULTIPLE, &provided );
    assert( err == 0 );
    assert( provided == MPI_THREAD_MULTIPLE );

    err = MPI_Comm_size( MPI_COMM_WORLD, &mpi_size );
    assert( err == 0 );
    // if (mpi_size != 4) {
    //     printf( "Usage: mpirun -np 4 %s  # 4 ranks hard coded\n", argv[0] );
    //     return -1;
    // }

    err = MPI_Comm_rank( MPI_COMM_WORLD, &mpi_rank );
    assert( err == 0 );

    // so random_matrix is different on different ranks.
    srand( 100 * mpi_rank );

    std::string output_file = "eigenvals.csv";

    hermitian_eig< float >( output_file );

    err = MPI_Finalize();
    assert( err == 0 );

    // Print finished from rank 0
    if (mpi_rank == 0) {
        printf( "Finished\n" );
    }
}

Corresponding SLURM File

#!/bin/bash
#SBATCH --job-name="Eigendecomposition"
#SBATCH --partition=main
#SBATCH --time=72:00:00
#SBATCH --nodes=10
#SBATCH --tasks-per-node=5
#SBATCH --cpus-per-task=4
#SBATCH --mem-per-cpu=24GB
#SBATCH --output=logs/eigen-%A.o
#SBATCH --error=logs/eigen-%A.e

# Installation
# spack install -j 16 slate%gcc ~cuda ^intel-oneapi-mkl ^openmpi@4+legacylaunchers fabrics=auto schedulers=slurm

# Load software
spack load [email protected]
spack load blaspp
spack load lapackpp

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/path/to/spack/install/spack/opt/spack/linux-ubuntu20.04-haswell/gcc-9.4.0/blaspp-2023.11.05-jfb72an/lib:/path/to/spack/install/spack/opt/spack/linux-ubuntu20.04-haswell/gcc-9.4.0/lapackpp-2023.11.05-3ervrb4/lib

export MKL_NUM_THREADS=$SLURM_CPUS_PER_TASK

# Start time
STARTTIME=$(date)

# Run command
/usr/bin/time -p mpirun fast_eigenvals

# End time
ENDTIME=$(date)
echo "Job started at: $STARTTIME"
echo "Job ended at: $ENDTIME"

Makefile

# Makefile to compile fast_eigenvals.cpp with slate and lapack (oneapi-mkl)
# Assumes BLAS++ and LAPACK++ are installed in SLATE directory.
# Otherwise would need to add
#   -I${blaspp_dir}/include -I${lapackpp_dir}/include to CXXFLAGS,
#   -L${blaspp_dir}/lib -L${lapackpp_dir}/lib and corresponding rpaths to LDFLAGS.
#
# Assumes everything built as shared libraries, so transitive dependencies
# are handled implicitly. That is, linking with -lslate implicitly links with
# -lblaspp -llapackpp -lopenblas.
# Substitute your choice of BLAS/LAPACK library for -lopenblas.

-include make.inc

slate_dir ?= /path/to/spack/install/spack/opt/spack/linux-ubuntu20.04-haswell/gcc-9.4.0/slate-2023.11.05-5ysppah
scalapack_libs ?= -L${MKLROOT}/lib/intel64 -lmkl_scalapack_lp64 -lmkl_blacs_intelmpi_lp64

# CXXFLAGS add /opt/slate/include to include path.
# LDFLAGS  add /opt/slate/lib     to lib path.
# Using -rpath avoids need to add ${slate_dir}/lib to $LD_LIBRARY_PATH.
CXX      = mpicxx
CXXFLAGS = -fopenmp -Wall -O3 -std=c++17 -MMD -I${slate_dir}/include -I/path/to/spack/install/spack/opt/spack/linux-ubuntu20.04-haswell/gcc-9.4.0/blaspp-2023.11.05-jfb72an/include -I/path/to/spack/install/spack/opt/spack/linux-ubuntu20.04-haswell/gcc-9.4.0/lapackpp-2023.11.05-3ervrb4/include -I${MKLROOT}/include
LDFLAGS  = -fopenmp -L${slate_dir}/lib -Wl,-rpath,${slate_dir}/lib -L/path/to/spack/install/spack/opt/spack/linux-ubuntu20.04-haswell/gcc-9.4.0/blaspp-2023.11.05-jfb72an/lib -L/path/to/spack/install/spack/opt/spack/linux-ubuntu20.04-haswell/gcc-9.4.0/lapackpp-2023.11.05-3ervrb4/lib -L${MKLROOT}/lib/intel64 -Wl,-rpath,${MKLROOT}/lib/intel64
LIBS     = -lblaspp -llapackpp -lmkl_gnu_thread -lmkl_intel_lp64 -lmkl_sequential -lmkl_core -lpthread -lm -ldl

# auto-detect OS
# $OSTYPE may not be exported from the shell, so echo it
ostype := $(shell echo $${OSTYPE})
ifneq ($(findstring darwin, $(ostype)),)
    # MacOS is darwin
    macos = 1
    LDD = otool -L
else
    LDD = ldd
endif


# ------------------------------------------------------------------------------
# Check if SLATE was compiled with cublas. If so, explicitly link cublas
# since it is called in the SLATE headers.
grep_cuda := ${shell ${LDD} ${slate_dir}/lib/libslate.so | grep cublas}
ifeq (${grep_cuda},)
    ${info SLATE without CUDA}
    CXXFLAGS += -DSLATE_NO_CUDA
else
    cuda_nvcc := ${shell which nvcc}
    CUDA_ROOT ?= $(cuda_nvcc:/bin/nvcc=)
    ${info SLATE with CUDA ${CUDA_ROOT}}
    ifeq (${cuda_nvcc},)
        ${error SLATE compiled with CUDA, but cannot find nvcc.}
    endif

    CXXFLAGS += -I${CUDA_ROOT}/include
    LDFLAGS  += -L${CUDA_ROOT}/lib -Wl,-rpath,${CUDA_ROOT}/lib
    LIBS     += -lcublas -lcudart
endif


# ------------------------------------------------------------------------------
# SLATE examples
# Link with -lslate.
# Implicitly links with -lblaspp -llapackpp -lopenblas.

slate_src = fast_eigenvals.cpp
slate_exe = ${basename ${slate_src}}

# ${slate_exe}: %: %.o
# 	${CXX} -o $@ $^ \
# 		${LDFLAGS} \
# 		-lslate ${LIBS}

${slate_exe}:
	${CXX} ${CXXFLAGS} -c -o ${slate_exe}.o ${slate_src} ${LDFLAGS} -lslate ${LIBS}
	${CXX} -o ${slate_exe} ${slate_exe}.o ${LDFLAGS} -lslate ${LIBS}


# ------------------------------------------------------------------------------
# Generic rules.

exe = ${slate_exe}

all: ${exe}

.DEFAULT_GOAL := all
.SUFFIXES:

%.o: %.cc
	${CXX} ${CXXFLAGS} -c -o $@ $<

clean:
	-rm -f ${exe} *.o

clean_exe:
	-rm -f ${exe}

distclean: clean clean_run
	-rm -f *.d

-include *.d

# ------------------------------------------------------------------------------
# Debugging

echo:
	@echo "CXX       = ${CXX}"
	@echo "CXXFLAGS  = ${CXXFLAGS}"
	@echo "LDFLAGS   = ${LDFLAGS}"
	@echo "LIBS      = ${LIBS}"
	@echo "LDD       = ${LDD}"
	@echo
	@echo "macos     = ${macos}"
	@echo "grep_cuda = ${grep_cuda}"
	@echo "cuda_nvcc = ${cuda_nvcc}"
	@echo "CUDA_ROOT = ${CUDA_ROOT}"
	@echo
	@echo "slate_dir = ${slate_dir}"
	@echo "scalapack_libs = ${scalapack_libs}"
	@echo
	@echo "slate_src = ${slate_src}"
	@echo "slate_exe = ${slate_exe}"
	@echo
	@echo "scalapack_src = ${scalapack_src}"
	@echo "scalapack_exe = ${scalapack_exe}"
	@echo
	@echo "blaspp_src = ${blaspp_src}"
	@echo "blaspp_exe = ${blaspp_exe}"
	@echo
	@echo "lapackpp_src = ${lapackpp_src}"
	@echo "lapackpp_exe = ${lapackpp_exe}"
	@echo
	@echo "exe = ${exe}"
	@echo
	@echo "txt = ${txt}"

Environment
All software other than SLURM and gcc was installed through spack:

• SLATE version / commit ID (e.g., git log --oneline -n 1): [email protected]
• How installed:
  • git clone
  • release tar file
  • Spack
  • module
• How compiled:
  • makefile (include your make.inc)
  • CMake (include your command line options)
• Compiler & version (e.g., mpicxx --version): [email protected]
• BLAS library (e.g., MKL, ESSL, OpenBLAS) & version: [email protected]
• MPI library & version (MPICH, Open MPI, Intel MPI, IBM Spectrum, Cray MPI, etc. Sometimes mpicxx -v gives info.): [email protected]
• OS: Ubuntu 20.04.6 LTS
• Hardware (CPUs, GPUs, nodes): CPU - Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz