Mist

From SciNet Users Documentation
Revision as of 14:55, 25 May 2022 by Rzon (talk | contribs)
Jump to navigation Jump to search
Mist
Mist.jpg
Installed Dec 2019
Operating System Red Hat Enterprise Linux 8.2
Number of Nodes 54 IBM AC922
Interconnect Mellanox EDR
Ram/Node 256 GB
GPUs/Node 4 V100-SMX2-32GB
Login/Devel Node mist.scinet.utoronto.ca
Vendor Compilers NVCC, IBM XL
Queue Submission Slurm

Specifications

Mist is a SciNet-SOSCIP joint GPU cluster consisting of 54 IBM AC922 servers. Each node of the cluster has 32 IBM Power9 cores, 256GB RAM and 4 NVIDIA V100-SMX2-32GB GPU with NVLINKs in between. The cluster has InfiniBand EDR interconnection providing GPU-Direct RMDA capability.

Important note: the majority of computer systems as of 2021 (laptops, desktops, and HPC) use the 64 bit x86 instruction set architecture (ISA) in their microprocessors produced by Intel and AMD. This ISA is incompatible with Mist, whose hardware uses the 64 bit PPC ISA (set to little endian mode). The practical meaning is that x86-compiled binaries (executables and libraries) cannot be installed on Mist. For this reason, the Niagara and Alliance (formerly Compute Canada) software stacks (modules) cannot be made available on Mist, and using closed-source software is only possible when the vendor provides a compatible version of their application. Python applications almost always rely on bindings to libraries originally written in C or C++, some of them are not available on PyPI or various Conda channels as precompiled binaries compatible with Mist. The recommended way to use Python on Mist is to create a Conda environment and install packages from the anaconda (default) channel, where most popular packages have a linux-ppc64le (Mist-compatible) version available. Some popular machine learning packages should be installed from the internal Open-CE channel. Where a compatible Conda package cannot be found, installing from PyPI (pip install) can be attempted. Pip will attempt to compile the package’s source code if no compatible precompiled wheel is available, therefore a compiler module (such as gcc/.core) should be loaded in advance. Some packages require tweaking of the source code or build procedure to successfully compile on Mist, please contact support if you need assistance.

Getting started on Mist

As of January 22 2022, authentication is only allowed via SSH keys. Please refer to this page to generate your SSH key pair and make sure you use them securely.

Mist can be accessed directly: 

ssh -i /path/to/ssh_private_key -Y MYCCUSERNAME@mist.scinet.utoronto.ca

Mist login node mist-login01 can also be accessed via Niagara cluster. 

ssh -i /path/to/ssh_private_key -Y MYCCUSERNAME@niagara.scinet.utoronto.ca
ssh -Y mist-login01

Storage

The filesystem for Mist is shared with Niagara cluster. See Niagara Storage for more details.

Loading software modules

You have two options for running code on Mist: use existing software, or compile your own. This section focuses on the former.

Other than essentials, all installed software is made available using module commands. These modules set environment variables (PATH, etc.), allowing multiple, conflicting versions of a given package to be available. A detailed explanation of the module system can be found on the modules page and a list of Modules for Mist is also available.

Common module subcommands are:

  • module load <module-name>: load the default version of a particular software.
  • module load <module-name>/<module-version>: load a specific version of a particular software.
  • module purge: unload all currently loaded modules.
  • module spider (or module spider <module-name>): list available software packages.
  • module avail: list loadable software packages.
  • module list: list loaded modules.

Along with modifying common environment variables, such as PATH, and LD_LIBRARY_PATH, these modules also create a SCINET_MODULENAME_ROOT environment variable, which can be used to access commonly needed software directories, such as /include and /lib.

There are handy abbreviations for the module commands. ml is the same as module list, and ml <module-name> is the same as module load <module-name>.

Tips for loading software

  • We advise against loading modules in your .bashrc. This can lead to very confusing behaviour under certain circumstances. Our guidelines for .bashrc files can be found here.
  • Instead, load modules by hand when needed, or by sourcing a separate script.
  • Load run-specific modules inside your job submission script.
  • Short names give default versions; e.g. cuda → cuda/11.0.3. It is usually better to be explicit about the versions, for future reproducibility.
  • Modules often require other modules to be loaded first. Solve these dependencies by using module spider.

Available compilers and interpreters

  • cuda module has to be loaded first for GPU software.
  • For most compiled software, one should use the GNU compilers (gcc for C, g++ for C++, and gfortran for Fortran). Loading gcc module makes these available.
  • The IBM XL compiler suite (xlc_r, xlc++_r, xlf_r) is also available, if you load one of the xl modules.
  • To compile mpi code, you must additionally load an openmpi or spectrum-mpi module.

CUDA

The current installed CUDA Tookits are 11.0.3 and 10.2.2 (10.2.89) 

module load cuda/11.0.3
module load cuda/10.2.2
  • A compiler (GCC, XL or NVHPC/PGI) module must be loaded in order to use CUDA to build any code.

The current NVIDIA driver version is 450.119.04.

GNU Compilers

Available GCC modules are: 

gcc/9.3.0 (must load CUDA 11)
gcc/8.5.0 (must load CUDA 10 or 11)
gcc/10.3.0 (w/o CUDA)

IBM XL Compilers

To load the native IBM xlc/xlc++ and xlf (Fortran) compilers, run 

module load xl/16.1.1.10

IBM XL Compilers are enabled for use with NVIDIA GPUs, including support for OpenMP GPU offloading and integration with NVIDIA's nvcc command to compile host-side code for the POWER9 CPU. Information about the IBM XL Compilers can be found at the following links:IBM XL C/C++, IBM XL Fortran

OpenMPI

openmpi/<version> module is avaiable with different compilers including GCC and XL. spectrum-mpi/<version> module provides IBM Spectrum MPI.

NVHPC/PGI

PGI compiler is provided in NVHPC (NVIDIA HPC SDK). 

module load nvhpc/21.3

Software

Amber20

Users who hold Amber20 license can build Amber20 from its source code and run on Mist. SOSCIP/SciNet doesn't provide Amber license or source code.

Building Amber20

Modules that are needed for building Amber20: 

module load MistEnv/2021a cuda/10.2.2 gcc/8.5.0 anaconda3/2021.05 cmake/3.19.8

Cmake configuration: 

cmake .. -DCMAKE_INSTALL_PREFIX=$HOME/where-amber-install -DCOMPILER=GNU -DMPI=FALSE -DCUDA=TRUE -DINSTALL_TESTS=TRUE -DDOWNLOAD_MINICONDA=FALSE -DOPENMP=TRUE -DNCCL=FALSE -DAPPLY_UPDATES=TRUE

Running Amber20

NVIDIA Pascal P100 and later GPUs like V100 do not scale beyond a single GPU. It is highly suggested to run Amber20 as a single-gpu job. A job example: 

#!/bin/bash
#SBATCH --nodes=1
#SBATCH --gpus-per-node=1
#SBATCH --time=1:00:0
#SBATCH --account=soscip-<SOSCIP-project-ID>

module load MistEnv/2021a cuda/10.2.2 gcc/8.5.0 anaconda3/2021.05
export PATH=$HOME/where-amber-install/bin:$PATH
export LD_LIBRARY_PATH=$HOME/where-amber-install/lib:$LD_LIBRARY_PATH
pmemd.cuda .... <parameters> ...

Anaconda (Python)

Anaconda is a popular distribution of the Python programming language. It contains several common Python libraries such as SciPy and NumPy as pre-built packages, which eases installation. Anaconda is provided as modules: anaconda3

To install Anaconda locally, user need to load the module and create a conda environment: 

module load anaconda3
conda create -n myPythonEnv python=3.8
  • Note: By default, conda environments are located in $HOME/.conda/envs. Cache (downloaded tarballs and packages) is under $HOME/.conda/pkgs. User may run into problem with disk quota if there are too many environments created. To clean conda cache, please run: "conda clean -y --all" and "rm -rf $HOME/.conda/pkgs/*" after installation of packages.

To activate the conda environment: (should be activated before running python) 

source activate myPythonEnv

Note that you SHOULD NOT use conda activate myPythonEnv to activate the environment. This leads to all sorts of problems. Once the environment is activated, user can update or install packages via conda or pip 

conda install  <package_name> (preferred way to install packages)
pip install <package_name>
  • Once the installation finishes, please clean the cache:
conda clean -y --all
rm -rf $HOME/.conda/pkgs/*

To deactivate: 

source deactivate

To remove a conda environment: 

conda remove --name myPythonEnv --all

To verify that the environment was removed, run: 

conda info --envs

Submitting Python Job

A single-gpu job example: 

#!/bin/bash
#SBATCH --nodes=1
#SBATCH --gpus-per-node=1
#SBATCH --time=1:00:0
#SBATCH --account=soscip-<SOSCIP_PROJECT_ID> #For SOSCIP projects only

module load anaconda3
source activate myPythonEnv
python code.py ...

CuPy

CuPy is an open-source matrix library accelerated with NVIDIA CUDA. It also uses CUDA-related libraries including cuBLAS, cuDNN, cuRand, cuSolver, cuSPARSE, cuFFT and NCCL to make full use of the GPU architecture. CuPy is an implementation of NumPy-compatible multi-dimensional array on CUDA. CuPy consists of the core multi-dimensional array class, cupy.ndarray, and many functions on it. It supports a subset of numpy.ndarray interface.

CuPy can be install into any conda environment. Python packages: numpy, six and fastrlock are required. cuDNN and NCCL are optional. 

module load MistEnv/2021a cuda/11.2.2 gcc/8.5.0 cudnn nccl anaconda3/2021.05
conda create -n cupy-env python=3.8 numpy six fastrlock
source activate cupy-env
CFLAGS="-I$MODULE_CUDNN_PREFIX/include -I$MODULE_NCCL_PREFIX/include -I$MODULE_CUDA_PREFIX/include" LDFLAGS="-L$MODULE_CUDNN_PREFIX/lib64 -L$MODULE_NCCL_PREFIX/lib" CUDA_PATH=$MODULE_CUDA_PREFIX pip install cupy
#building/installing CuPy will take a few minutes

Gromacs

GROMACS is a versatile package to perform molecular dynamics, i.e. simulate the Newtonian equations of motion for systems with hundreds to millions of particles. It is primarily designed for biochemical molecules like proteins, lipids and nucleic acids that have a lot of complicated bonded interactions, but since GROMACS is extremely fast at calculating the nonbonded interactions (that usually dominate simulations) many groups are also using it for research on non-biological systems, e.g. polymers.

  • GROMACS 2019
module load MistEnv/2021a cuda/10.2.2 gcc/8.5.0 gromacs/2019.6
  • GROMACS 2020 and later Thread-MPI version supports full GPU enablement of all key computational sections. The GPU is used throughout the timestep and repeated CPU-GPU transfers are eliminated. Users are suggested to carefully verify the results.
module load MistEnv/2021a cuda/10.2.2 gcc/8.5.0 gromacs/2020.4
module load MistEnv/2021a cuda/10.2.2 gcc/8.5.0 gromacs/2020.6
module load MistEnv/2021a cuda/11.0.3 gcc/9.4.0 gromacs/2021.2
module load MistEnv/2021a cuda/11.0.3 gcc/9.4.0 openmpi/4.1.1+ucx-1.10.0 gromacs/2021.2 (testing purpose only)
module load MistEnv/2021a cuda/11.0.3 gcc/9.4.0 gromacs/2021.4
module load MistEnv/2021a cuda/11.0.3 gcc/9.4.0 gromacs/2021.5
module load MistEnv/2021a cuda/11.0.3 gcc/9.4.0 gromacs/2022

Small/Medium Simulation

Due to the lack of PME domain decomposition support on GPU, Gromacs uses CPU to calculate PME when using multiple GPUs. It is always recommended to use a single GPU to do small and medium sized simulations with Gromacs. By using only 1 tMPI thread (w/ multiple OpenMP threads) on a single GPU, both non-bonded PP and PME are atomically offloaded to GPU when possible.

  • Gromacs 2019 example:
#!/bin/bash
#SBATCH --time=20:00
#SBATCH --nodes=1
#SBATCH --gpus-per-node=1

module load MistEnv/2021a cuda/10.2.2 gcc/8.5.0 gromacs/2019.6
export OMP_NUM_THREADS=8
export OMP_PLACES=cores
gmx mdrun -pin off -ntmpi 1 -ntomp 8  ... <other parameters>
  • Gromacs 2020 or later example:
#!/bin/bash
#SBATCH --time=20:00
#SBATCH --nodes=1
#SBATCH --gpus-per-node=1

module load MistEnv/2021a cuda/11.0.3 gcc/9.4.0 gromacs/2021.5
export OMP_NUM_THREADS=8
export OMP_PLACES=cores
export GMX_FORCE_UPDATE_DEFAULT_GPU=true
gmx mdrun -pin off -ntmpi 1 -ntomp 8  ... <other parameters>

Large Simulation

If memory size (~58GB) for single-gpu job is not sufficient for the simulation, multiple GPUs can be used. It is suggested to test starting with one full node with 4GPUs and force PME on GPU. Multiple PME ranks are not supported with PME on GPU, so if GPU is used for the PME calculation -npme (number of PME ranks) must be set to 1. If PME has less work than PP, it is suggested to run multiple ranks per GPU, so the GPU for PME rank can also do some work on PP rank(s). If your simulation can fit in a single GPU job, please use single GPU to get much higher efficiency. Do not waste 3 additional GPU resource for getting only a small performance improvement.

  • An example using 4 GPUs, 7 PP ranks/tmpi threads + 1 PME rank/tmpi thread: (-pin on -pme gpu -npme 1 must be added to mdrun command in order to force GPU to do PME)
#!/bin/bash
#SBATCH --time=20:00
#SBATCH --gpus-per-node=4
#SBATCH --ntasks=1
#SBATCH --nodes=1
#SBATCH -p compute_full_node

module load MistEnv/2021a cuda/11.0.3  gcc/9.4.0 gromacs/2021.5

export OMP_NUM_THREADS=4
gmx mdrun -ntmpi 8 -pin on -pme gpu -npme 1 ... <add your parameters>
  • It is suggested to also test using -ntmpi 4 and export OMP_NUM_THREADS=8 if you receive a NOTE in Gromacs output saying "% performance was lost because the PME ranks had more work to do than the PP ranks". In this case, NVIDIA MPS is not needed since there is only one MPI rank per GPU.
  • Please note that the solving of PME on GPU is still only the initial version supporting this behaviour, and comes with a set of limitations outlined further below.
* Only a PME order of 4 is supported on GPUs.
* PME will run on a GPU only when exactly one rank has a PME task, ie. decompositions with multiple ranks doing PME are not supported.
* Only single precision is supported.
* Free energy calculations where charges are perturbed are not supported, because only single PME grids can be calculated.
* Only dynamical integrators are supported (ie. leap-frog, Velocity Verlet, stochastic dynamics)
* LJ PME is not supported on GPUs.
  • An example using 4 GPUs, PME on CPU: (-pin on must be added to mdrun command for proper CPU thread bindings)
#!/bin/bash
#SBATCH --time=20:00
#SBATCH --gpus-per-node=4
#SBATCH --ntasks=1
#SBATCH --nodes=1
#SBATCH -p compute_full_node

module load MistEnv/2021a cuda/11.0.3  gcc/9.4.0 gromacs/2021.5

export OMP_NUM_THREADS=4
gmx mdrun -ntmpi 8 -pin on  ... <add your parameters>

# "-ntmpi 16, OMP_NUM_THREADS=2" and "-ntmpi 4, OMP_NUM_THREADS=8" should also be tested.  
# num_thread_MPI_ranks(-ntmpi) * num_OpenMP_threads = 32
  • If your simulation can fit in a single GPU job, please use single GPU to get much higher efficiency. Do not waste 3 additional GPU resource for getting only a small performance improvement.
  • NOTE: The above examples will NOT work with multiple nodes. If simulation is too large for a single GPU node, please contact SciNet/SOSCIP support.

NAMD

NAMD is a parallel, object-oriented molecular dynamics code designed for high-performance simulation of large biomolecular systems.

2.14

module load MistEnv/2021a cuda/11.0.3 gcc/9.4.0 spectrum-mpi/10.4.0 namd/2.14

Running with single GPU

If you have many jobs to run, it is always suggested to run with a single gpu per job. This makes jobs easier to be scheduled and gives better overall performance. 

#!/bin/bash
#SBATCH --time=20:00
#SBATCH --gpus-per-node=1
#SBATCH --nodes=1

module load MistEnv/2021a cuda/11.0.3 gcc/9.4.0 spectrum-mpi/10.4.0 namd/2.14
scontrol show hostnames > nodelist-$SLURM_JOB_ID

`which charmrun` -npernode 1 -bind-to none -hostfile nodelist-$SLURM_JOB_ID `which namd2` +idlepoll +ppn 8 +p 8 stmv.namd

Running with one process per node (4 GPUs)

An example of the job script (using 1 node, one process per node, 32 CPU threads per process + 4 GPUs per process): 

#!/bin/bash
#SBATCH --time=20:00
#SBATCH --gpus-per-node=4
#SBATCH --ntasks=1
#SBATCH --nodes=1
#SBATCH -p compute_full_node

module load MistEnv/2021a cuda/11.0.3 gcc/9.4.0 spectrum-mpi/10.4.0 namd/2.14
scontrol show hostnames > nodelist-$SLURM_JOB_ID

`which charmrun` -npernode 1 -hostfile nodelist-$SLURM_JOB_ID `which namd2` +setcpuaffinity +pemap 0-127:4 +idlepoll +ppn 32 +p $((32*SLURM_NTASKS)) stmv.namd

Running with one process per GPU (4 GPUs)

NAMD may scale better if using one process per GPU. Please do your own benchmark. An example of the job script (using 1 node, one process per GPU, 8 CPU threads per process): 

#!/bin/bash
#SBATCH --time=20:00
#SBATCH --gpus-per-node=4
#SBATCH --ntasks=4
#SBATCH --nodes=1
#SBATCH -p compute_full_node

module load MistEnv/2021a cuda/11.0.3 gcc/9.4.0 spectrum-mpi/10.4.0 namd/2.14
scontrol show hostnames > nodelist-$SLURM_JOB_ID

`which charmrun` -npernode 4 -hostfile nodelist-$SLURM_JOB_ID `which namd2` +setcpuaffinity +pemap 0-127:4 +idlepoll +ppn 8 +p $((8*SLURM_NTASKS)) stmv.namd

Open-CE

Open-CE is an IBM repo for feedstock collection, environment data, and scripts for building Tensorflow, Pytorch, and other machine learning packages and dependencies. Open-CE is distributed as a conda channel on Mist cluster. Available packages and versions are listed here Open-CE Releases. Currently only python 3.8 and CUDA 11.2 are supported. If you need a different python or cuda version, please contact SOSCIP or SciNet support.

  • Packages can be installed by setting Open-CE conda channel:
conda install -c /scinet/mist/ibm/open-ce python=3.8 cudatoolkit=11.2 PACKAGE
  • Once the installation finishes, please clean the cache:
conda clean -y --all
rm -rf $HOME/.conda/pkgs/*
Available Packages:
Tensorflow TensorFlow Estimators TensorFlow Probability TensorBoard TensorBoard Data Server TensorFlow Text TensorFlow Model Optimizations TensorFlow Addons (tensorflow-addons) TensorFlow Datasets TensorFlow Hub
TensorFlow MetaData PyTorch TorchText TorchVision PyTorch Lightning PyTorch Lightning Bolts ONNX Onnx-runtime skl2onnx tf2onnx
onnxmltools onnxconverter-common XGBoost LightGBM Transformers Tokenizers SentencePiece Spacy DALI OpenCV
Horovod PyArrow grpc uwsgi ORC Mamba Ray (ray-tune) pytorch_geometric

PyTorch

Installing from IBM Open-CE Conda Channel

The easiest way to install PyTorch on Mist is using IBM's Conda channel. User needs to prepare a conda environment and install PyTorch using IBM's Open-CE Conda channel. 

module load anaconda3
conda create -n pytorch_env python=3.8
source activate pytorch_env

#must force to use Open-CE channel to avoid the cpu-only version of PyTorch from default Anaconda channel
conda config --prepend channels /scinet/mist/ibm/open-ce
conda config --set channel_priority strict

conda install -c /scinet/mist/ibm/open-ce pytorch=1.10.2 cudatoolkit=11.2
or
conda install -c /scinet/mist/ibm/open-ce-1.2 pytorch=1.7.1 cudatoolkit=11.0 (or 10.2)

Once the installation finishes, please clean the cache: 

conda clean -y --all
rm -rf $HOME/.conda/pkgs/*
#remove .condarc to reset conda channel priority
rm -f $HOME/.condarc

Add below command into your job script before python command to get deterministic results, see details here: [1] 

export CUBLAS_WORKSPACE_CONFIG=:4096:2

RAPIDS

The RAPIDS is a suite of open source software libraries that gives you the freedom to execute end-to-end data science and analytics pipelines entirely on GPUs. The RAPIDS data science framework includes a collection of libraries: cuDF(GPU DataFrames), cuML(GPU Machine Learning Algorithms), cuStrings(GPU String Manipulation), etc.

  • The most recent version supported is 0.13.0. Newer version is no longer provided by IBM's powerai channel.

Installing from IBM Conda Channel

The easiest way to install RAPIDS on Mist is using IBM's Conda channel. User needs to prepare a conda environment with Python 3.6 or 3.7 and install powerai-rapids using IBM's Conda channel. Python 3.8+ is not supported. 

module load anaconda3
conda create -n rapids_env python=3.7
source activate rapids_env
conda install -c https://public.dhe.ibm.com/ibmdl/export/pub/software/server/ibm-ai/conda-early-access/ powerai-rapids

Once the installation finishes, please clean the cache: 

conda clean -y --all
rm -rf $HOME/.conda/pkgs/*

TensorFlow and Keras

Installing from IBM Conda Channel

The easiest way to install TensorFlow and Keras on Mist is using IBM's Open-CE Conda channel. User needs to prepare a conda environment and install TensorFlow using IBM's Open-CE Conda channel. 

module load anaconda3
conda create -n tf_env python=3.8
source activate tf_env
conda install -c /scinet/mist/ibm/open-ce tensorflow==2.7.1 cudatoolkit=11.2
or
conda install -c /scinet/mist/ibm/open-ce-1.2 tensorflow==2.4.3 cudatoolkit=11.0 (or 10.2)

Once the installation finishes, please clean the cache: 

conda clean -y --all
rm -rf $HOME/.conda/pkgs/*

Ray

Ray is an API for building distributed applications. A local wheel is available for Python 3.8, but some tinckering is required to succesfully install it. Please activate your Conda environment and follow this installation recipe: 

conda install tabulate tensorboardX pandas dataclasses aiohttp aioredis click colorama colorful filelock gpustat grpcio jsonschema numpy protobuf py-spy pyyaml requests redis opencensus prometheus_client beautifulsoup4 soupsieve cython wheel
pip install msgpack google aiohttp_cors
# Manually add py-spy version info, since Conda forgot to do that
PYSPYVERSION=$(py-spy --version | cut -d ' ' -f 2)
mkdir $CONDA_PREFIX/lib/python3.8/site-packages/py_spy-$PYSPYVERSION.dist-info
echo -e "Metadata-Version: 2.1\nName: py-spy\nVersion: $PYSPYVERSION" > $CONDA_PREFIX/lib/python3.8/site-packages/py_spy-$PYSPYVERSION.dist-info/METADATA
pip install /scinet/mist/wheelhouse/experimental/2021a/ray-1.0.1.post1-cp38-cp38-linux_ppc64le.whl
conda clean -y --all
rm -rf ~/.conda/pkgs/*

Testing and debugging

You really should test your code before you submit it to the cluster to know if your code is correct and what kind of resources you need.

  • Small test jobs can be run on the login node. Rule of thumb: tests should run no more than a couple of minutes, taking at most about 1-2GB of memory, and use no more than one gpu and a few cores.
  • Short tests that do not fit on a login node, or for which you need a dedicated node, request an interactive debug job with the debug command:
mist-login01:~$ debugjob --clean -g G

where G is the number of gpus, If G=1, this gives an interactive session for 2 hours, whereas G=4 gets you a single node with 4 gpus for 30 minutes, and with G=8 (the maximum) gets you 2 nodes each with 4 gpus for 15 minutes. The --clean argument is optional but recommended as it will start the session without any modules loaded, thus mimicking more closely what happens when you submit a job script. Users needs to load module and activate the conda environment after a debug job starts. It is recommended to do a 'conda clean' before 'source activate ENV' in a debug job if --clean flag is missed.

Submitting jobs

Once you have compiled and tested your code or workflow on the Mist login nodes, and confirmed that it behaves correctly, you are ready to submit jobs to the cluster. Your jobs will run on some of Mist's 53 compute nodes. When and where your job runs is determined by the scheduler.

Mist uses SLURM as its job scheduler. It is configured to allow only Single-GPU jobs and Full-node jobs (4 GPUs per node).

You submit jobs from a login node by passing a script to the sbatch command:

mist-login01:scratch$ sbatch jobscript.sh

This puts the job in the queue. It will run on the compute nodes in due course. In most cases, you should not submit from your $HOME directory, but rather, from your $SCRATCH directory, so that the output of your compute job can be written out (as mentioned above, $HOME is read-only on the compute nodes).

Example job scripts can be found below. Keep in mind:

  • Scheduling is by single gpu or by full node, so you ask only 1 gpu or 4 gpus per node.
  • Your job's maximum walltime is 24 hours.
  • Jobs must write their output to your scratch or project directory (home is read-only on compute nodes).
  • Compute nodes have no internet access.
  • Your job script will not remember the modules you have loaded, so it needs to contain "module load" commands of all the required modules (see examples below).

SOSCIP Users

  • SOSCIP is a consortium to bring together industrial partners and academic researchers and provide them with sophisticated advanced computing technologies and expertise to solve social, technical and business challenges across sectors and drive economic growth.

If you are working on a SOSCIP project, please contact soscip-support@scinet.utoronto.ca to have your user account added to SOSCIP project accounts. SOSCIP users need to submit jobs with additional SLURM flag to get higher priority: 

#SBATCH -A soscip-<SOSCIP_PROJECT_ID>    #e.g. soscip-3-001
OR
#SBATCH --account=soscip-<SOSCIP_PROJECT_ID>

Single-GPU job script

For a single GPU job, each will have a quarter of the node which is 1 GPU + 8/32 CPU Cores/Threads + ~58GB CPU memory. Users should never ask CPU or Memory explicitly. If running MPI program, user can set --ntasks to be the number of MPI ranks. Do NOT set --ntasks for non-MPI programs.

  • It is suggested to use NVIDIA Multi-Process Service (MPS) if running multiple MPI ranks on one GPU.
#!/bin/bash
#SBATCH --nodes=1
#SBATCH --gpus-per-node=1
#SBATCH --time=1:00:0
#SBATCH --account=soscip-<SOSCIP_PROJECT_ID> #For SOSCIP projects only

module load anaconda3
source activate conda_env
python code.py ...

Full-node job script

If you are not sure the program can be executed on multiple GPUs, please follow the single-gpu job instruction above or contact SciNet/SOSCIP support.

Multi-GPU job should ask for a minimum of one full node (4 GPUs). User need to specify "compute_full_node" partition in order to get all resource on a node.

  • An example for a 1-node job:
#!/bin/bash
#SBATCH --nodes=1
#SBATCH --gpus-per-node=4
#SBATCH --ntasks=4 #this only affects MPI job
#SBATCH --time=1:00:00
#SBATCH -p compute_full_node
#SBATCH --account=soscip-<SOSCIP_PROJECT_ID> #For SOSCIP projects only

module load <modules you need>
Run your program

Limits

There are limits to the size and duration of your jobs, the number of jobs you can run and the number of jobs you can have queued.

Usage Partition Running jobs Jobs in queue Min. size of jobs Max. size of jobs Min. walltime Max. walltime
Compute jobs compute 100 GPUs 1000 1 GPU (8 cores) default: 4 nodes (16 GPUs)
with allocation: 4 nodes (16 GPUs)		 15 minutes 24 hours
Testing or troubleshooting debug 1 1 1 GPU (8 cores) 2 nodes (8 GPUs) N/A 2/ngpu hours

Even if you respect these limits, your jobs will still have to wait in the queue. The waiting time depends on many factors such as your group's allocation amount, how much allocation has been used in the recent past, the number of requested nodes and walltime, and how many other jobs are waiting in the queue.

Jupyter Notebooks

SciNet’s Jupyter Hub is a Niagara-type node; it has a different CPU architecture and no GPUs. Conda environments prepared on Mist will not work there properly. Users who need to use Jupyter Notebook to develop and test some aspects of their workflow can create their own server on the Mist login node and use an SSH tunnel to connect to it from outside. Users who choose to do so have to keep in mind that the login node is a shared resource, and heavy calculations should be done only on compute nodes. Processes (including iPython kernels used by the notebooks) are limited to one hour of total CPU time: idle time will not be counted toward this one hour, and use of multiple cores will count proportionally to the number of cores (i.e. a kernel using all 128 virtual cores on the node will be killed after 28 seconds). Idle notebooks can still burden the node by hogging system and GPU memory, please be mindful of other users and terminate notebooks when work is done.

As an example, let us create a new Conda environment and activate it: 

module load anaconda3
conda create -n jupyter_env python=3.7
source activate jupyter_env

Install the Jupyter Notebook server: 

conda install notebook

Running the notebook server

When the Conda environment is active, enter: 

jupyter-notebook

By default, the Jupyter Notebook server uses port 8888 (can be overridden with the --port option). If another user has already started their own server, the default port may be busy, in which case the server will be listening on a different port. Once launched, the server will output some information to the terminal that will include the actual port number used and a 48-character token. For example: 

http://localhost:8890/?token=54c4090d……

In this example, the server is listening on port 8890.

Creating a tunnel

In order to access this port remotely (i.e. from your office or home), an SSH tunnel has to be established. Please refer to your SSH client’s documentation for instructions on how to do that. For the OpenSSH client (standard in most Linux distributions and macOS), a tunnel can be opened in a separate terminal session to the one where the Jupyter Notebook server is running. In the new terminal, issue this command: 

ssh -L8888:localhost:8890 <username>@mist.scinet.utoronto.ca

(replace <username> with your actual username) The tunnel is open as long as this SSH connection is alive. In this example, we tunnel Mist login node’s port 8890 (where our server is assumed to be running) to our home computer’s port 8888 (any other free port is fine). The notebook can be accessed in the browser at the http://localhost:8888 address (followed by /?token=54c4090d……, or the token can be input on the webpage).

Using Jupyter on compute nodes

You can use the instructions here to set up a Jupyter Notebook server on a compute node (including a debugjob). We strongly discourage you from running an interactive notebook on a compute node (other than for a debugjob), scheduled jobs run in arbitrary times and are not meant to be interactive. Jupyter notebooks can be run non-interactively or converted to Python scripts.

To launch the Jupyter Notebook server, load the anaconda3 module and activate your environment as before (by adding the appropriate lines to the submission script, if you are not using the compute node with an interactive shell). Launching the server has to be done like so: 

HOME=/dev/shm/$USER jupyter-notebook

That is because Jupyter will fail unless it can write to the home folder, which is read-only from compute nodes. This modification of the $HOME environment variable will carry over into the notebooks, which is usually not a problem, but in case the notebook relies on this environment variable (e.g. to read certain files), it can be reset manually in the notebook (import os; os.environ['HOME']=……).

Because compute nodes are not accessible from the Internet, tunneling has to be done twice, once from the remote location (office or home) to the Mist login node, and then from the login node to the compute node. Assuming the server is running on port 8890 of the mist006 node, open the first tunnel in a new terminal session in the remote computer: 

ssh -L8888:localhost:9999 <username>@mist.scinet.utoronto.ca

where 9999 is any available port on the Mist login node (to test port availability enter ss -Hln src :9999 in the terminal when connected to the Mist login node; an empty output indicates that the port is free). In the same session in the login node that was created with the above command, open the second tunnel to the compute node: 

ssh -L9999:localhost:8890 mist006

Be aware that the second tunnel will automatically disconnect once the job on the compute node times out or is relinquished. The Jupyter Notebook server running on the compute node can now be accessed from the browser as in the previous subsection.


Support

SciNet inquiries:

SOSCIP inquiries:

Retrieved from "https://docs.scinet.utoronto.ca/index.php?title=Mist&oldid=3953"