Mist

From SciNet Users Documentation
Jump to: navigation, search
Mist
Mist.jpg
Installed Dec 2019
Operating System Red Hat Enterprise Linux 7.6
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.

Getting started on Mist

Mist can be accessed directly.

ssh -Y MYCCUSERNAME@mist.scinet.utoronto.ca

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

ssh -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.

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. cudacuda/10.1.243. 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 softwares.
  • 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 10.1.243 and 10.2.89 (default)

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

The current NVIDIA driver version is 440.33.01.

GNU Compilers

Available GCC modules are:

gcc/7.5.0
gcc/8.3.0

IBM XL Compilers

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

module load xl/16.1.1.3

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.

PGI

To load PGI compiler and its own OpenMPI environment, run:

module load pgi/19.10
module load pgi-openmpi/3.1.3

Softwares

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=3.7
  • 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>

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 -A 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 anaconda3/2019.10 cuda/10.2.89 gcc/7.5.0 cudnn/7.6.5.32  nccl/2.5.6 
conda create -n cupy-env python=3.7 numpy six fastrlock
source activate cupy-env
CFLAGS="-I$SCINET_CUDNN_ROOT/include -I$SCINET_NCCL_ROOT/include -I$SCINET_CUDA_ROOT/include" LDFLAGS="-L$SCINET_CUDNN_ROOT/lib64 -L$SCINET_NCCL_ROOT/lib" CUDA_PATH=$SCINET_CUDA_ROOT 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.

module load cuda/10.2.89  gcc/8.3.0  openmpi/3.1.5 gromacs/2019.5
module load cuda/10.2.89  gcc/8.3.0  openmpi/3.1.5 gromacs/2019.6
  • GROMACS 2020 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. Currently only single-GPU is supported on Mist. Users are suggested to carefully verify the results.
module load cuda/10.2.89 gcc/8.4.0 openmpi/4.0.3 gromacs/.experimental-2020.2

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 MPI rank (w/ OpenMP threads) on a single GPU, both non-bonded PP and PME are atomically offloaded to GPU when possible.

  • A Single-GPU Gromacs job must ask --ntasks=32 even only 1 MPI rank will be launched by mpirun command. OMP_PLACES must be set to core to force OpenMP threads on physical CPU cores. -bind-to none and -pin off must be set to avoid CPU affiliate conflicts among OpenMP, MPI and Gromacs. OMP_NUM_THREADS must be set to 8 to get optimal performance.
#!/bin/bash
#SBATCH --time=20:00
#SBATCH --gpus-per-node=1
#SBATCH --ntasks=32

module load cuda/10.2.89  gcc/8.3.0  openmpi/3.1.5 gromacs/2019.6
export OMP_NUM_THREADS=8
export OMP_PLACES=cores
mpirun -np 1 -bind-to none gmx_mpi mdrun -pin off ... <other parameters>
  • Groamcs 2020 example: (OpenMPI module should to be loaded, but mpirun should NOT be used)
#!/bin/bash
#SBATCH --time=20:00
#SBATCH --nodes=1
#SBATCH --gpus-per-node=1

module load cuda/10.2.89 gcc/8.4.0 openmpi/4.0.3 gromacs/.experimental-2020.2
export OMP_NUM_THREADS=8
export OMP_PLACES=cores
gmx mdrun -pin off -ntmpi 1 -ntomp 8 -update gpu ... <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). When running multiple MPI ranks on the same GPU, NVIDIA Multi-Process Service (MPS) must be enabled.

  • An example using 4 GPUs, 7 PP ranks + 1 PME rank: (-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=8
#SBATCH --nodes=1
#SBATCH -p compute_full_node

module load cuda/10.2.89  gcc/8.3.0  openmpi/3.1.5 gromacs/2019.6

mkdir -p /dev/shm/nvidia-mps
export CUDA_MPS_PIPE_DIRECTORY=/dev/shm/nvidia-mps
mkdir -p /dev/shm/nvidia-log
export CUDA_MPS_LOG_DIRECTORY=/dev/shm/nvidia-log
nvidia-cuda-mps-control -d

export OMP_NUM_THREADS=4
mpirun  -bind-to none gmx_mpi mdrun -pin on -pme gpu -npme 1 ... <add your parameters>
  • It is suggested to also test using --ntasks=4 and 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=8
#SBATCH --nodes=1
#SBATCH -p compute_full_node

module load cuda/10.2.89  gcc/8.3.0  openmpi/3.1.5 gromacs/2019.6

mkdir -p /dev/shm/nvidia-mps
export CUDA_MPS_PIPE_DIRECTORY=/dev/shm/nvidia-mps
mkdir -p /dev/shm/nvidia-log
export CUDA_MPS_LOG_DIRECTORY=/dev/shm/nvidia-log
nvidia-cuda-mps-control -d

export OMP_NUM_THREADS=4
mpirun -bind-to none gmx_mpi mdrun -pin on  ... <add your parameters>

# "--ntasks=16, OMP_NUM_THREADS=2" and "--ntasks=4, OMP_NUM_THREADS=8" should also be tested.  
# num_Tasks(MPI_ranks) * num_OpenMP_threads = 32
  • 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.

IBM Watson Machine Learning Community Edition (PowerAI)

IBM Watson Machine Learning Community Edition (PowerAI) contains many popular ML packages including TensorFlow, PyTorch, XGBoost and RAPIDS. It is distributed through IBM Conda channel. To install packages from PowerAI, user needs to specify IBM Conda channel when using Anaconda.

module load anaconda3

conda create --name wmlce_env -c https://public.dhe.ibm.com/ibmdl/export/pub/software/server/ibm-ai/conda <package_name> (e.g. powerai, tensorflow-gpu, keras, pytorch, powerai-rapids, py-xgboost-gpu,  etc)

source activate wmlce_env 

NAMD

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

v2.13

module load cuda/10.2.89 gcc/7.5.0 fftw/3.3.8 spectrum-mpi/10.3.1  namd/2.13

Running with one process per node

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 cuda/10.2.89 gcc/7.5.0 fftw/3.3.8 spectrum-mpi/10.3.1  namd/2.13
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

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 cuda/10.2.89 gcc/7.5.0 fftw/3.3.8 spectrum-mpi/10.3.1  namd/2.13
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

PyTorch

Installing from IBM Conda Channel

The easiest way to install PyTorch on Mist is using IBM's Conda channel. User needs to prepare a conda environment with Python 3.6 or 3.7 and install PyTorch using IBM's Conda channel.

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

Once the installation finishes, please clean the cache:

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

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.

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.

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/ 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 Conda channel. User needs to prepare a conda environment with Python 3.6 or 3.7 and install TensorFlow-gpu using IBM's Conda channel.

module load anaconda3
conda create -n tf_env python=3.7
source activate tf_env
conda install -c https://public.dhe.ibm.com/ibmdl/export/pub/software/server/ibm-ai/conda/ tensorflow-gpu==2.1.0 
If you need TF 1.x version:
conda install -c https://public.dhe.ibm.com/ibmdl/export/pub/software/server/ibm-ai/conda/ tensorflow-gpu==1.15.0

Once the installation finishes, please clean the cache:

conda clean -y --all
rm -rf $HOME/.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 30 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.

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 -A 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 2-node, 8-rank OpenMPI job: (Each rank binds to 1 GPU and 8 physical CPU cores in this case)
#!/bin/bash
#SBATCH --nodes=2
#SBATCH --gpus-per-node=4
#SBATCH --ntasks=8
#SBATCH --time=1:00:00
#SBATCH -p compute_full_node
#SBATCH -A soscip-<SOSCIP_PROJECT_ID> #For SOSCIP projects only

module load cuda/10.2.89 gcc/8.3.0 openmpi/3.1.5

mpirun -bind-to core -map-by slot:PE=8 -report-bindings ./program

Support

SciNet inquiries:

SOSCIP inquiries: