VetNode 0.0.5

A node vetting cli for Distributed Workloads

Node Vetting for Distributed Workloads

Ensure allocated nodes are vetted before executing a distributed workload through a series of configurable sanity checks. These checks are designed to detect highly dynamic issues (e.g., GPU temperature) and should be performed immediately before executing the main distributed job.

Features

• ⚡ Fast and lightweight
• 🛠️ Modular and configurable
• 🚀 Easy to extend

Getting Started

# Install
pip install vetnode

# checks for dependencies and installs requirements
vetnode setup ./examples/local-test/config.yaml

# runs the vetting process
vetnode diagnose ./examples/local-test/config.yaml

Workflow Usage Example

The vetnode cli is intended to be embedded into your HPC workflow. The following is a node vetting example for a ML (machine learning) workflow on a Slurm HPC cluster.

#!/bin/bash

#SBATCH --nodes=6
#SBATCH --time=0-00:15:00
#SBATCH --account=a-csstaff

REQUIRED_NODES=4
MAIN_JOB_COMMAND="python -m torch.distributed.torchrun --nproc_per_node=$(wc -l < vetted-nodes.txt) main.py"

vetnode setup ../examples/slurm-ml-vetting/config.yaml
srun vetnode diagnose ../examples/slurm-ml-vetting/config.yaml >> results.txt

# Extract node lists
grep '^Cordon:' results.txt | awk '{print $2}' > cordoned-nodes.txt
grep '^Vetted:' results.txt | awk '{print $2}' > vetted-nodes.txt

#Run on healthy nodes only
if [ $(wc -l < vetted-nodes.txt) -ge $REQUIRED_NODES ]; then
    srun -N $REQUIRED_NODES --exclude=./cordoned-nodes.txt $MAIN_JOB_COMMAND
else
    echo "Job canceled!"
    echo "Reason: too few vetted nodes."
fi

Quick Run

The following is a Slurm job example you can download and run as a test.

curl -o job.sh  https://raw.githubusercontent.com/theely/vetnode/refs/heads/main/examples/slurm-ml-vetting/job.sh
sbatch --account=a-csstaff job.sh

#check job status
squeue -j {jobid} --long

#check vetting results
cat vetnode-{jobid}/results.txt

Development

Set-up Python Virtual environement

Create a virtual environment:

python3.11 -m venv .venv
source .venv/bin/activate
python3 -m pip install --upgrade pip
pip install -r requirements.txt

Run the CLI

cd src
python -m vetnode setup ../examples/local-test/config.yaml
python -m vetnode diagnose ../examples/local-test/config.yaml

Running Tests

From the FirecREST root folder run pytest to execute all unit tests.

source .venv/bin/activate
pip install -r ./requirements.txt -r ./requirements-testing.txt
pytest

Distribute

pip install -r ./requirements-testing.txt
python3 -m build --wheel
twine upload dist/*         

Info dump

Clariden Distro:

NAME="SLES" VERSION="15-SP5" VERSION_ID="15.5" PRETTY_NAME="SUSE Linux Enterprise Server 15 SP5" ID="sles" ID_LIKE="suse" ANSI_COLOR="0;32" CPE_NAME="cpe:/o:suse:sles:15:sp5" DOCUMENTATION_URL="https://documentation.suse.com/"

./configure --prefix=/users/palmee/aws-ofi-nccl/install --disable-tests --without-mpi --enable-cudart-dynamic --with-libfabric=/opt/cray/libfabric/1.15.2.0 --with-cuda=/opt/nvidia/hpc_sdk/Linux_aarch64/24.3/cuda/12.3/

https://download.opensuse.org/repositories/home:/aeszter/openSUSE_Leap_15.3/x86_64/libhwloc5-1.11.8-lp153.1.1.x86_64.rpm https://download.opensuse.org/repositories/home:/aeszter/15.5/x86_64/libhwloc5-1.11.8-lp155.1.1.x86_64.rpm

Build plugin in image

export DOCKER_DEFAULT_PLATFORM=linux/amd64 export #export DOCKER_DEFAULT_PLATFORM=linux/arm64 docker run -i -t registry.suse.com/suse/sle15:15.5
zypper install -y libtool git gcc awk make wget

zypper addrepo https://developer.download.nvidia.com/compute/cuda/repos/opensuse15/x86_64/cuda-opensuse15.repo zypper addrepo https://developer.download.nvidia.com/compute/cuda/repos/sles15/sbsa/cuda-sles15.repo zypper --non-interactive --gpg-auto-import-keys refresh zypper install -y cuda-toolkit-12-3

Add missing lib path required by hwloc

echo "/usr/local/cuda/targets/x86_64-linux/lib/stubs/" | tee /etc/ld.so.conf.d/nvidiaml-x86_64.conf echo "/usr/local/cuda/targets/sbsa-linux/lib/stubs/" | tee /etc/ld.so.conf.d/nvidiaml-sbsa.conf

ldconfig ldconfig -p | grep libnvidia

git clone -b v1.19.0 https://github.com/ofiwg/libfabric.git cd libfabric autoupdate ./autogen.sh #CC=gcc ./configure --prefix=/users/palmee/libfabric/install CC=gcc ./configure make make install

wget https://download.open-mpi.org/release/hwloc/v2.12/hwloc-2.12.0.tar.gz tar -xvzf hwloc-2.12.0.tar.gz cd hwloc-2.12.0 #CC=gcc ./configure --prefix=/users/palmee/hwloc-2.12.0/install CC=gcc ./configure make make install

git clone -b v1.14.0 https://github.com/aws/aws-ofi-nccl.git cd aws-ofi-nccl mkdir install GIT_COMMIT=$(git rev-parse --short HEAD)

./autogen.sh CC=gcc ./configure --disable-tests --without-mpi
--enable-cudart-dynamic
--prefix=./install/v1.14.0-${GIT_COMMIT}/x86_64/12.3/
--with-cuda=/usr/local/cuda

TODO: consider building an rpm: https://www.redhat.com/en/blog/create-rpm-package

CC=gcc ./configure --disable-tests --without-mpi
--enable-cudart-dynamic
--prefix=/users/palmee/aws-ofi-nccl/install_2/
--with-libfabric=/opt/cray/libfabric/1.15.2.0 --with-cuda=/opt/nvidia/hpc_sdk/Linux_aarch64/24.3/cuda/12.3/ --with-hwloc=/users/palmee/hwloc-2.12.0/install

export LD_LIBRARY_PATH=/opt/cray/libfabric/1.15.2.0/lib64/:$LD_LIBRARY_PATH export LD_LIBRARY_PATH=/opt/nvidia/hpc_sdk/Linux_aarch64/24.3/cuda/12.3/lib64/:$LD_LIBRARY_PATH ld /users/palmee/aws-ofi-nccl/install_2/lib/libnccl-net.so

Install NCCL

git clone https://github.com/NVIDIA/nccl.git git checkout v2.20.3-1 #looks like this is the version compatible with cuda/12.3/ cd nccl make src.build CUDA_HOME=/opt/nvidia/hpc_sdk/Linux_aarch64/24.3/cuda/12.3/

WORKING LIB (job 327119)

payload	busbw	algbw
1GiB	90.94GBps	46.94GBps
2GiB	91.24GBps	47.09GBps
4GiB	91.35GBps	47.15GBps

wget https://download.open-mpi.org/release/hwloc/v2.12/hwloc-2.12.0.tar.gz tar -xvzf hwloc-2.12.0.tar.gz cd hwloc-2.12.0 ./configure --prefix=/users/palmee/hwloc-2.12.0/install make make install

git clone -b v1.14.0 https://github.com/aws/aws-ofi-nccl.git cd aws-ofi-nccl mkdir install

./autogen.sh ./configure --disable-tests --without-mpi
--enable-cudart-dynamic
--prefix=/users/palmee/aws-ofi-nccl/install/
--with-libfabric=/opt/cray/libfabric/1.15.2.0 --with-cuda=/opt/nvidia/hpc_sdk/Linux_aarch64/24.3/cuda/12.3/ --with-hwloc=/users/palmee/hwloc-2.12.0/install

TEST with gcc (job 327124) - working

payload	busbw	algbw
1GiB	91.06GBps	47.00GBps
2GiB	91.24GBps	47.09GBps
4GiB	91.34GBps	47.15GBps

wget https://download.open-mpi.org/release/hwloc/v2.12/hwloc-2.12.0.tar.gz tar -xvzf hwloc-2.12.0.tar.gz cd hwloc-2.12.0 CC=gcc ./configure --prefix=/users/palmee/hwloc-2.12.0/install make make install

git clone -b v1.14.0 https://github.com/aws/aws-ofi-nccl.git cd aws-ofi-nccl mkdir install

./autogen.sh CC=gcc ./configure --disable-tests --without-mpi
--enable-cudart-dynamic
--prefix=/users/palmee/aws-ofi-nccl/install_3/
--with-libfabric=/opt/cray/libfabric/1.15.2.0 --with-cuda=/opt/nvidia/hpc_sdk/Linux_aarch64/24.3/cuda/12.3/ --with-hwloc=/users/palmee/hwloc-2.12.0/install make make install

TEST with gcc all (job 327130) - running

payload	busbw	algbw
1GiB	91.05GBps	46.99GBps
2GiB	91.24GBps	47.09GBps
4GiB	91.34GBps	47.14GBps

wget https://download.open-mpi.org/release/hwloc/v2.12/hwloc-2.12.0.tar.gz tar -xvzf hwloc-2.12.0.tar.gz cd hwloc-2.12.0 CC=gcc ./configure --prefix=/users/palmee/hwloc-2.12.0/install make make install

git clone -b v1.14.0 https://github.com/aws/aws-ofi-nccl.git cd aws-ofi-nccl mkdir install

./autogen.sh CC=gcc ./configure --disable-tests --without-mpi
--enable-cudart-dynamic
--prefix=/users/palmee/aws-ofi-nccl/install_3/
--with-libfabric=/opt/cray/libfabric/1.15.2.0 --with-cuda=/opt/nvidia/hpc_sdk/Linux_aarch64/24.3/cuda/12.3/ --with-hwloc=/users/palmee/hwloc-2.12.0/install make make install

TEST with local libfabric only for compile (job 327145) -

payload	busbw	algbw
1GiB	91.08GBps	47.01GBps
2GiB	91.21GBps	47.08GBps
4GiB	91.34GBps	47.14GBps

git clone -b v1.19.0 https://github.com/ofiwg/libfabric.git cd libfabric autoupdate ./autogen.sh CC=gcc ./configure --prefix=/users/palmee/libfabric/install make make install

wget https://download.open-mpi.org/release/hwloc/v2.12/hwloc-2.12.0.tar.gz tar -xvzf hwloc-2.12.0.tar.gz cd hwloc-2.12.0 CC=gcc ./configure --prefix=/users/palmee/hwloc-2.12.0/install make make install

git clone -b v1.14.0 https://github.com/aws/aws-ofi-nccl.git cd aws-ofi-nccl mkdir install

./autogen.sh CC=gcc ./configure --disable-tests --without-mpi
--enable-cudart-dynamic
--prefix=/users/palmee/aws-ofi-nccl/install_4/
--with-libfabric=/users/palmee/libfabric/install --with-cuda=/opt/nvidia/hpc_sdk/Linux_aarch64/24.3/cuda/12.3/ --with-hwloc=/users/palmee/hwloc-2.12.0/install make make install

TEST with local libfabric compile and job run (job 327161) -

NOT WORKING!!! We need to use the crey lib fabric

Build plugin in image

export DOCKER_DEFAULT_PLATFORM=linux/amd64 docker run -i -t registry.suse.com/suse/sle15:15.5
zypper install -y libtool git gcc awk make wget

zypper addrepo https://developer.download.nvidia.com/compute/cuda/repos/opensuse15/x86_64/cuda-opensuse15.repo zypper refresh zypper install -y cuda-toolkit-12-3

Build in Clariden

wget https://download.open-mpi.org/release/hwloc/v2.12/hwloc-2.12.0.tar.gz tar -xvzf hwloc-2.12.0.tar.gz cd hwloc-2.12.0 CC=gcc ./configure --prefix=/users/palmee/hwloc-2.12.0/install #CC=gcc ./configure make make install

git clone -b v1.14.1 https://github.com/aws/aws-ofi-nccl.git cd aws-ofi-nccl mkdir install

./autogen.sh CC=gcc ./configure --disable-tests --without-mpi
--enable-cudart-dynamic
--prefix=/users/palmee/aws-ofi-nccl/install/
--with-libfabric=/opt/cray/libfabric/1.22.0/
--with-cuda=/opt/nvidia/hpc_sdk/Linux_aarch64/24.3/cuda/12.3/
--with-hwloc=/users/palmee/hwloc-2.12.0/install

make make install

export LD_LIBRARY_PATH=/opt/cray/libfabric/1.15.2.0/lib64/:$LD_LIBRARY_PATH export LD_LIBRARY_PATH=/opt/nvidia/hpc_sdk/Linux_aarch64/24.3/cuda/12.3/lib64/:$LD_LIBRARY_PATH ld /users/palmee/aws-ofi-nccl/install/lib/libnccl-net.so

TODO: consider building an rpm: https://www.redhat.com/en/blog/create-rpm-package