Installing PETSc and Torchdiffeq on Theta

Detailed step-by-step instructions on how to install PETSc and torchdiffeq on Argonne’s supercomputer Theta.

Image credit: ANL

📚 Table of Contents

  1. 👉Prerequistie
  2. 🦄Install PETSc
  3. ✨Install torchdiffeq

👉 Prerequisite on ThetaGPU

  1. Request an interactive node

Once you login to Theta, do the following

ssh thetagpusn1
qsub -n 1 -t <NODE_TIME_IN_MINUTES> -A <ALLOCATION_NAME> -I
  1. Enable python environment:
source /lus/theta-fs0/software/thetagpu/conda/2021-06-28/mconda3/setup.sh

Python is not avaible on thetagpusn nodes and GPU nodes. This will setup a conda environment with a recent “from scratch” build of the PyTorch repository on the master branch. The location is subject to change.

  1. Set up internet connection:
export http_proxy=http://proxy.tmi.alcf.anl.gov:3128
export https_proxy=http://proxy.tmi.alcf.anl.gov:3128

Alternatively, you can add --attrs=pubnet when requesting nodes with qsub .

Now you can install python packages with pip install, e.g.

pip install Cython matplotlib scipy

💬More info can be found here

  1. Lunch jobs on Theta:
aprun -n 64 -N 64 ./<your_program> <program_arguments>

You need to use aprun to execute your program on Theta.

  • -n specifies the total number of MPI ranks
  • -N specifies the ranks per node
  • Add -q if you want to silence the output such as Application 22305720 resources: utime ~10s, stime ~3s, Rss ~18776, inblocks ~0, outblocks ~8

🦄 Installation of PETSc

  • GPU build
git clone https://gitlab.com/petsc/petsc.git
cd petsc
./configure --download-revolve --with-shared-libraries COPTFLAGS=-O3 CXXOPTFLAGS=-O3 FOPTFLAGS=-O3 PETSC_ARCH=arch-theta-gpu-opt --with-cuda-gencodearch=80 --with-cuda=1 --with-cudac=nvcc --with-petsc4py --download-fblaslapack

Follow the printed instructions at the end of configure to do a make. If you enabled petsc4py (--with-petsc4py), add <path_to_petsc>/petsc/arch-theta-cuda-opt/lib to PYTHONPATH. For example, the following can be run directly or added into your .bash_profile script.

export PYTHONPATH=<path_to_petsc>/petsc/arch-theta-cuda-opt/lib:PYTHONPATH
  • KNL build
git clone https://gitlab.com/petsc/petsc.git
cd petsc
./configure --LIBS=-lstdc++ --known-64-bit-blas-indices=0 --known-bits-per-byte=8 --known-has-attribute-aligned=1 --known-level1-dcache-assoc=8 --known-level1-dcache-linesize=64 --known-level1-dcache-size=32768 --known-memcmp-ok=1 --known-mklspblas-supports-zero-based=0 --known-mpi-c-double-complex=1 --known-mpi-int64_t=1 --known-mpi-long-double=1 --known-mpi-shared-libraries=1 --known-sdot-returns-double=0 --known-sizeof-MPI_Comm=4 --known-sizeof-MPI_Fint=4 --known-sizeof-char=1 --known-sizeof-double=8 --known-sizeof-float=4 --known-sizeof-int=4 --known-sizeof-long-long=8 --known-sizeof-long=8 --known-sizeof-short=2 --known-sizeof-size_t=8 --known-sizeof-void-p=8 --known-snrm2-returns-double=0 --with-cc=cc --with-clib-autodetect=0 --with-cxx=CC --with-cxxlib-autodetect=0 --with-batch --with-debugging=0 --with-fc=0 --with-hdf5=0 --with-make-np --with-memalign=64 --with-shared-libraries COPTFLAGS=-O3 CXXOPTFLAGS=-O3 FOPTFLAGS=-O3 PETSC_ARCH=arch-theta-knl-opt

💡Tips

  • Do not change any option before --with-batch
  • For debug version, use --with-debugging and remove the OPTFLAGS
  • PETSC_ARCH is just a folder name of your choice, used to discriminate different builds

✨ Installation of torchdiffeq

git clone https://github.com/rtqichen/torchdiffeq.git
cd torchdiffeq
pip install --user -e .